I don’t think anyone who have yet commented has used the actual model. But, to be fair to the commenters, google has made it very hard for them to do so without spending a ton of money.
Neon is a really good model. It’s currently my favorite model to use. Coding-wise, it’s equivalent to Opus 5.5/Fable. Documentation-wise, it is great. If you have access to Neon have it take a crack at simplifying your agents.md. It’s the first model I’ve run that criticized the amount of ceremony in my instructions and suggested solid simplifications. It also makes clean edits to documents without leaving scars of the previous verbiage behind.
Art-wise, it’s more capable, in my experience, than Fable. I’ve been using fable/opus heavily to make art assets for my application (I shared in a previous comment that procedural art has completely obviated gpt-image art for me) and I’ve started to run into a wall with how far I could iterate with fable. I had neon take a crack at the assets I struggled to improve and it did a grand job.
In terms of working together with a model and how it communicates to you, I love this model. The model knows how to clearly communicate to you. It is such a breath of fresh air over Opus and Fable which, despite the improvements, I feel is actively trying to damage my brain with its unnatural prose. It's hard to describe how it feels to have a model that you feel respects your time and energy with proper writing, and the subsequent delight that arises from the feeling of wanting to engage in smaller details again.
Neon has a big model smell, but the miracle and irony is it’s supposedly based on a smaller base model than Carbon. On google’s model page, they previously wrote that Neon shares the same base model as Hydrogen! It is thought that Argon and Carbon share the same base model.
My impressions will probably change as I use this for longer than a day, but my initial impression is: I don’t know how google pulled this off.
no doubt Neon is a capable model but for us paying $200/month nothing can really match Codex or Claude's usage. Like I have a lot of trouble paying for API to work on a new project, most of the time I am brain storming trying different things until I arrive at a solid direction and putting a meter on it feels prohibitive vs a fixed "all you can eat" plan.
but seeing these subsidized plans are mostly enterprise acquisition schemes I do feel like they will slowly turn into pay as you go type and now that China's access to nvidia gpus is curbed, I think we'll see less and less investment into spending too cheap to meter models as the incentives disappear
All in all I want to say, build whatever you want right now because when there are lot more vibers tomorrow, they won't be able to support subsidized plans. Fact that Google hasn't budged on this tells me that they see no upside
Maybe someone can help me clear this up: I've got a $20/mo subscription to Claude that's not through a console, and which I just use as a chat/coding window in VSCode. Then I have an API key that I use with OpenClaw in a sandboxed system. Both are using Opus 5.5.
A good 4-6 hour session of code and tool usage on the API can run me $80-$100. Are you saying I can get that all for $200/mo if I switch to Claude Code and use the high end subscription?
so ... I just have to give up the openclaw harness? I guess my understanding atm is that they'll reject calls to a subscription method if they're coming from a claw.
I wouldn't say my relationship with the claw is worth an extra $500 or something, but there are serious improvements over using Claude Code directly. Scheduled jobs, the whatsapp channel, and a bit of personality... I'm disinclined to hunt around for another harness that Anthropic would accept on the terms of their subscription, only to find it banned shortly thereafter. Is there a well trodden path to just using openclaw on one of those higher level subscriptions instead of raw API pricing?
So you can get massively subsidized coding use (I use it through T3, herdr works, the official clients work too of course) and also get a modest amount of API credit for whatever you want (claw, etc.).
The grey market may or may not have some solutions for this. Usage is at your own risk.
Theres also options to connect openclaw via the real cli, but that does blow up your system prompt. I use that option right now and it works quite well. There's no free lunch as they say.
Usage limits in Antigravity, I take it. So you can't take it to your harness of choice, or e.g. run batch image processing jobs. For $200 per month, that's pretty lousy.
Yeah that's pretty shitty. What's not shitty is that you get like 10-20x (maybe more) usage per dollar compared to the anthropic subs. I've been running 3 neon sessions for the last ~hour and I'm still at 0% weekly usage... my anthropic 20x pro sub weekly limit just ran out after about 1 day of semi-active use...
"I don’t think anyone who have yet commented has used the actual model. But, to be fair to the commenters, google has made it very hard for them to do so without spending a ton of money."
I don't know why that would be.
The Neon model is available today in both my Gemini app on iPhone and at https://gemini.google.com/app, and I just have a $4.99/month Google AI Pro trial offer.
I presume I could boot it up in Antigravity also, but I haven't done so yet.
Just tell me how to fit the Gemini app into the pipeline I currently have for transcribing historical handwritten documents. After a lot of wheedling, I got the chat model to admit that no, that isn't actually possible, if I want to do something that isn't playing around in their web app I need to shell out for API credits. "Oh, but you get free API credits with a subscription!" the bot tried to sell me. How much? $10 of credits per month for a $20 per month subscription, it turns out. The Gemini app is a toy.
Google has subscriptions equivalent to openai and anthropic - Google AI pro/ultra, etc. You can download google antigravity (Google's CLI harness) and use that with the subscription.
To be honest the Google AI studio, and all of the related recent Google web UIs and related surfaces are absolute dumpster trash, just finding where/how to subscribe and set this up was so unnecessarily painful, and the UI is so painfully slow and buggy- holy shit.
That being said I just got neon running via antigravity (agy), and it is quite good. Noticeably faster than opus 5.5, and so far it seems like I will get at least 10x usage compared to my anthropic sub of similar price.
I feel like I'm a dinosaur when I manually type gemini.google.com/app to open the native app on my phone. It's already installed magically, why do I have to rely on an open-in-app cue to boot it.
I see, that's good to know. I was previously using a version of Gemini Pro that had an extended free trial (I've since upgraded to ultra). In my version of pro, I had none of the modern models available to use. I am happy to share blame with google for this confusion.
Capacity bottlenecks made them not include certain models in "promotional" accounts, i.e. ones that are currently on various promotional sales accounting.
There is a lot of those, one could say is that Google is trying to curb their previous overgenerosity
I can't even use my bank cards for Google (normal Visa or Mastercard debit card) in this country because it flashes up a mysterious error message about "temporary cards not being allowed for purchases". This has been going on for about four years that I know of, across at least three banks, both Visa and Mastercard, and none of them are temporary cards (normal expiry date several years after the card is issued) and I don't have trouble using the same cards for any other online purchases including OpenAI or Anthropic.
Google just makes it weirdly hard to give them money.
Just because you were able to pay them $20 and test it doesn't mean other people have that experience.
Outside of their ad business googles enterprise offerings have always been incredibly hit or miss. It just isnt an enterprise that values support or stability.
Haha, very true. That's honestly why I'm probably switching to it from Claude. As far as I can tell, it's at least as good as Opus 5.5 at writing code, but the way it communicates back to me is so nice. I had it clean up some READMEs that had vibe-coded sections that had leaked in over the months... it found those sections, reworded them (dropping most of it, honestly), and told me exactly why when it was done.
If I had to guess it probably has all the research optimisations that went into Gemma and whatever new Gemma they have coming out. Probably saves a lot of memory and speed.
I left it overnight working on a massive project I have and am extremely pleased with the result. It has made several mistakes (ones that Claude likely wouldn't make), but it seems to have built faster and waaaaay cheaper than Claude.
I'm working on "hand-drawn" objects modeled in 3d. Regarding iteration, for example, I've struggled to get fable to generate a mug that has a geometrically correct handle with a natural taper. I'm sure fable could have gotten to what neon got me but after a certain number of iterations it wasnt worth the time and tokens.
It's also available on $20/month Gemini Pro accounts.
It totally replaced Argon there, which is no longer available.
Interestingly, Argon used to run into rate limits all the time on $20/mo plans; you'd get three or four prompts, then you'd have to spend several hours in a cooldown period where you could only use an old 3.X Flash model. Neon is either much more efficient, or its limits have been relaxed. I get quite a lot of use out of it, and haven't yet been cast into cooldown purgatory.
It has only replaced Argon completely in the Web UI, it's still available in agy with separate usage like before.
I'm really happy with the Pro plan right now, Hydrogen and Neon have been great and actually usable day one. 5 TB of storage for my family and pretty decent AI usage for $10 a month through Verizon FiOS add-on. Much better value than the ChatGPT offering.
Totally agree. AI Studio seems to go on forever using the best model (per your subscription plan). Anrigravity (desktop) and Anrigravity (IDE) to run the Orchestration etc to get into local mode for the weeds or run your local data pipeline, uses your quota and then GitHub sync back into AI Studio for seemingly endless quota at the top model where it easily handles massive context. AI Studio is smart to build environments that easily work when in local vs studio (web) modes. Neon actually seems better than Argon, which was great for reasoning.
Super stupid question. I pay $20/month on Google (for extra storage, etc). I also paid Claude Code ($200 per month). Leaving aside the fact that Claude probably will give me more tokens, is there a way I can replace my whole flow of Claude Code with a CLI by Google (that also works well with Herdr)? I feel Google makes using their models very very hard!
I gave Antigravity a try, coming from Claude Code. To list a few differences, agy does
* not depend on the folder/repo, iow resume in an other folder is fine
* not have a an auto mode, only feasible way is YOLO mode. The sandbox seems nice and I’ve tried and tried and tried but always got stuck in endless confirmation prompts
* not have --ide integration with vscode that in manual mode shows the pre-write diff
* not integrate as clean with the browser as claude-in-chrome does
> not have a an auto mode, only feasible way is YOLO mode.
There's an auto mode. You can Shift-Tab a few times to rotate between Default mode, "Plan mode", "Auto-edit mode", and "Auto mode" (or `agy --mode=auto`).
First impression (though I haven't used it for code yet) is that is less verbose and executes less things between messages, or at least doesn't show all of them or show them in a more compact way.
With Claude Code I feel every answer triggers a lot of actions and agents, while with Antigravity seems there are less of that and it's more straightforward.
I just ran into this Argon rate limiting last weekend - just as you describe. Is there any documentation about how much I can use Neon in the app or browser? I have this $20 plan and it became useless recently because of the rate limits.
I love that we are years into LLMs and Google still doesn't roll out models in a timely fashion.
I was interested to see what Neon was like so downloaded antigravity. Signed in with my personal account, but it's AI Plus and not AI Pro (I just wanted it for the storage, but I assumed I got _some_ AI stuff included given the name).
No problem, sign in with my workspace account, nope despite being on a higher tier I just get 3.8 Flash.
I understand they may be very compute constrained, but surely it's worth giving users even a handful of prompts.
> Google still doesn't roll out models in a timely fashion.
This is the same way Google has been doing roll out across products for a decade. And it is unlikely to change.
At the scale of Google, deployments like this would reach billions users, spanning over the world, and take hours, if not days, for a proper roll out. An instant deployment would carry a lot of risks.
To be clear I'm not talking about the main Gemini UI. Which ironically I have access to hydrogen on my accounts.
I'm talking about antigravity - their coding agent, which has orders of magnitude less usage than Gemini or AI mode.
But tbh Gemini has the same issue, this is a rare exception.
I'd just have thought that given how little dev mind share Google has you might want to give paying (!) customers timely access for a few prompts to evaluate it.
Considering how much OAI and ANT have captured the market I find it hard to believe Google has that many people who will really give their models that much work on paid plans
Note that OpenAI's corresponding tier, ChatGPT Go, doesn't even have Terra (the Hydrogen equivalent), let alone larger models like Sol (the Neon equivalent).
> If you have access to Neon have it take a crack at simplifying your agents.md. It’s the first model I’ve run that criticized the amount of ceremony in my instructions and suggested solid simplifications.
> Neon is a really good model. It’s currently my favorite model to use. Coding-wise, it’s equivalent to Opus 5.5/Fable.
When I read this type of thing I do really start to see why Anthropic et al are putting so much effort into the harness. I have a lot of hooks and things setup to make Opus/Fable tolerable but it is tolerable. The same but change the harness is not really a great selling point. I have definitely talked with fellow devs who prefer OpenAIs models because the word barf out of Anthropic models can be exhausting even if they don't code as well or whatever.
Admittedly since OpenAI/Anthropic models are serving my needs and my past experience getting access to Google's models was a pain in the ass I am less inclined to mess around with Googles shitty consoles and API permissions.
Gemini and Fable both generally perform well when provided a generic prompt like "draw a picture of an abstract landscape using javascript".
Can't share my work but my process involved a lot of planning which included discussing specific feelings i wanted my art to evoke, providing a collection of images (a moodboard), and having the model generate multiple mockups to establish direction and procedure. That took a lot of different iterations but once I finally established the style, the model had no problem transferring it forward.
google has done something very wrong on the marketing side, i know a lot of people in software engineering and no one i know uses their models or is excited about anything they do
I use gemini almost exclusively. Our CEO uses claude and it's like 90% wrong. Maybe he prompts it wrong?
I can pretty much one or two shot everything in gemini. Not sure if that means i'm working on simpler stuff or I’m just better at describing what i want. But i'm also using like orders of magnitude less tokens than everyone else.
People talk constantly about defecting between plans. The grass is always greener, there's always something worth bitching about, and things fly fast in the thunderdome. If any entity comes out with something worth using, it'll get traction.
I have a "fringe" theoretical physics idea[1] that I've been grinding through various frontier models to see how far they can push this kind of thing. It's a great test because it is sufficiently "out there" that they can't memorise the answer, and it has absolutely brutal mathematics.
A simple cross-check is to get two models to advance the paper, and then have the other model validate it.
All of the Gemini 4 models -- including Neon -- consistently make simple, "doofus" errors like neglecting minus signs or doubling a scalar factor due to accidentally multiplying by it twice. Similarly, they often use two different sets of assumptions for a long mathematical argument involving multiple proofs, so the whole argument ends up invalid.
For comparison, these basic errors no longer occur with any of the recent Anthropic or OpenAI models, including every GPT 6 Sol or Astra model that I've tested.
Simply put, if you let Astra do 5-10 iterations in a row (~1 hour of work) and feed it to Gemini 4 Neon, it won't spot any errors. Do the reverse of that and even GPT 6.2 Sol will tear Neon's work to pieces.
PS: I suspect that Gemini models have a higher temperature setting, which makes their English sound more natural, but introduces errors that degrade their hard science capability.
[1] But not crank physics, which I would expect to "trip up" models trained on correct, valid physics and mathematics.
A variant of the notion that the standard model of physics is "simply" the property of some high-dimensional algebra. For example, Cohl Furey proposed a few different variations of the Octonions such as the dual-Octonions and I believe most recently the Dixon algebra ℂ ⊗ ℍ ⊗ O. There are other groups investigating Lie algebras related to E8 and the like.
The idea is that the Standard Model is soooo close to a nice and neat algebra that it may just be possible that it simply "pops out" of one naturally. All of particle physics might just be a very simple mathematical rule, instead of a whole pile of things on top of other things that looks suspiciously like epicycles.
There's a "search" by theoreticians for the equivalent algebraic change in perspective that Galileo did when he moved astronomy from overly complex 2D mathematics to a far simpler 3D algebra.
I have several notions that now -- thanks to AI -- I can throw at the proverbial wall to see if they stick.
The one that I've been playing around with the longest is a real geometric algebra with a total of nine dimensions: three temporal, three spatial, and three nilpotent ("matter"). I also sprinkled on some of my preferred postulates, such as making the theory MWI-flavoured, fully relational (no absolute clocks or rulers), etc.
I knew it probably wouldn't work, but this algebra has an incredibly beautiful symmetry of three-by-three and it is "just" big enough to encompass the SM particle zoo including the three particle generations!
The idea was that the three temporal dimensions might explain the "many worlds" aspects of quantum mechanics.
It almoooost worked!
I got it to the point that all three frontier AIs agreed that it's publication-worthy and one particular branch was "ahead" of Cohl Furey's favored algebra, but... meh... it's (provably) a dead end. Not to mention that everyone and their dog will be doing the same kind of thing with their own toy theory, so nobody will be interested in yet more AI slop output unless it makes a dramatic discovery.
Despite the dead end it was worth a go, especially given that trying is effectively free: I usually burn tokens on this right before using a "reset" that's about to expire.
It's actually a great test to see how much a particular model likes to "butter your bread" versus giving a negative opinion.
Anthropic models: Somewhat brutal actually. Fable told me to "stop wasting my time".
OpenAI models: Incredibly diplomatic and neutral, refuses to state a firm opinion either way even if pressed.
Google Gemini models: "The honesty of this paper refreshing, this should be the benchmark for how all theoretical research is published!" (etc)
I learned not to trust anything a Google model says and to use the combination of an Anthropic and an OpenAI model if I want to know if something really is a bad idea.
Sadly I only get -5 + 10 points for "I'm not good at math, but my theory is conceptually right, so all I need is for someone to express it in terms of equations".
I'm literally using AI to paper over my sub-standard mathematics skills!
Trust me, there's hundreds -- if not thousands -- of people doing this right now, and many of them likely work in serious physics research programmes.
Many people have ideas, not many people can dedicate their lives to them.
Now, it becomes entirely possible for someone with a "notion" to make the next "big" discovery, but not actually understand what they have discovered.
> But, to be fair to the commenters, google has made it very hard for them to do so without spending a ton of money.
Google has an introductory $5 subscription that defaults to Gemini 4 Neon (High). I was curious, so I grabbed that and got things going within 5 minutes.
Neon is a contender. Neon "high" appears faster and cheaper than Opus 5.5 "medium" with similar results on long-running language porting work in a large monorepo.
The speed is the biggest win IMO. When Carbon came out, I tried to give it a shot, but it was so slow that I went back to Opus just for productivity, nevermind quality. Today I started Neon on one worktree and Opus on another, and Neon did so well that I have now switched for both.
Code quality looks great. Comment-wise it has its own Claudese(exclamations like "this function has a concurrency bug and it's a sharp one!"), but overall I like reading it more than Opus. Might just be because I am not yet tired of its quirks though.
Oh, after a productive day on this particular work I had only spent $30 of Neon vs. $70 of Opus yesterday. YMMV
Which Qwen? For Flash-Next there are several tunes that reduce the thinking tokens, like ThinkingCap or Swift. I've been running a Frankenstein version of the latter on my DGX Spark.
I've been using Argon steady since it was released and a few times it's gotten in this really excited about everything mood like that. Most of the time it's totally even keel. Cute is a good word for it, not sure what's going on under the hood.
Faster and cheaper, and apparently better than Opus 5.5 medium? Bro I just got my mind blown by Opus 5.5, now Neon gets to once again shift my mental benchmark and compel me to finally download and try Antigravity? Right after I migrated from the Claude Code TUI to the desktop app and never looked back for all of the 2 weeks it's been?
I'll do it just for the improved prose. I've had it up to here with Opus 5.5's surprisingly poor writing skills, even though it's leagues better than the Willem-Dafoe-on-cocaine dumpster fire that is Opus 5.
If you watch Youtube much at all, Youtube Premium Lite (no ads) is well worth the $9/month. This is included with their $20/month AI plan. Spending that other $11/month wasn't worth it before Neon, but it might be now.
There are 3 different antigravity quotas:
* gemini (previously just Gemini 3.8 Flash)
* claude and gpt (very small quota)
* argon (very small quota)
The Gemini quota now includes Neon and Hydrogen. Hydrogen (medium) competes with DeepSeek (high). Its within 2x on cost and more intelligent.
This is all from benchmarks- I am only just trying these models out now.
How much neon usage do we get on this plan? I have that plan and agy is uselsss as a single prompt burns my entire Claude budget, argon gets me like 3 chats, and im definitely not using whatever flash model is listed for coding.
I am happy to pay a fair cost and not have ads. If everyone blocks ads the content isn't going to exist or if enough block they are going to work harder to block ad blockers.
I have paid for Kagi and Spotify for a long time. Its tragic that most of the web only offers a large monthly subscription to read a single article or intrusive ads- I do use Zen on Desktop and Orion on mobile due to that.
Even if you pay, you will have ads. Ads are now embedded into the content. So much content is in fact just ads masquerading as content to steal your attention. A lot of content out there simply should not and does not need to exist. It is just an endless rehash of pre-existing ideas with the sole purpose of beating the algorithm. The model is fundamentally broken imo and needs to be disrupted.
I'm reading the plan features and it says "YouTube Premium Lite individual plan6
Most videos ad-free, offline, and in the background". So you still get ads with some videos? The next plan ($100) has actual youtube premium no-ads
I've been using neon for like an hour after getting the Google AI ultra sub... so far it seems like it could be at least 10x the value of an anthropic sub...
> This is included with their $20/month AI plan. Spending that other $11/month wasn't worth it before Neon, but it might be now
This is their golden ticket.
Amazon famously bundled a variety of services under Prime which had, at least for a time, absolutely unheard of retention metrics. Truly otherworldly stuff. They probably still do.
I've been a Prime member since inception, longer than practically any other service. Given my household's needs I'm unlikely to cancel it any time soon.
Google bundling YouTube and Gemini under their... whatever-it-is-called plan (One? At least that name was a sufficiently thematic copy of "Prime"...) is brilliant. The $20 plan is a no-brainer for a ton of folks. They need to make the $200/mo plan a no brainer, and I think they can, but I doubt they have the wherewithal to actually figure it out and pull it off.
The market has rarely, maybe never, been more amenable to otherwise monopolistic bundling, I think they'd be crazy not to try and swing for the fences.
And I'm not saying any of this is a good thing, I'm certainly not endorsing this behavior. It's just that if I were a soulless automaton C-level executive trying to maximize shareholder value at all costs, and looking at the-company-most-likely-to-dethrone-me's IPO earlier today, and had Neon hot out of the gate, pulling away from the pack, I would be going for blood right now.
Well I'm not that C-level executive, so even though I do use Google models today (they're really superior at historical HTR), such slimy bundling practices makes me see if the Chinese models can be wrangled into replacing them (maybe with the right few-shot examples...)
Fully agree. Neon is my main model right now. I have a pretty advanced setup with multiple agents, coordinators, etc. It's basically a "factory" customized for me. It ran on Claude for a long time, then on Codex for the past three months, and now it runs on Google models. Neon is good, predictable, and smarter than me :)
Agy has a bunch of good slash commands for different use cases. And the biggest thing for me is that it doesn't annoy me with security refusals. It helped me with my garage door, for example, when another model refused.
Ok, so I can see that this is a benchmark the submitter made and submitted themselves, which is kinda misleading and should at least be disclosed.
But my question is, what does this benchmark actually measure about the LLM's engineering abilities? It just looks like a page of random bar graphs to me.
looking at the graphic below the table, it does seem like data is missing for Neon which is present for Fable, and isn't penalized for the missing benchmarks.
Seems fair game to me if the benchmark were actually a reasonable attempt at rating models, but it’s total slop. Like Neon score is based on three benchmarks and Fable based on 6 and that’s hidden.
Maybe but the sentiment of people who have actually used it is more important than these benchmarks anyway and from this thread that seems to be highly positive.
I have a little WIP game (don't we all...) and I use models through MCP to author maps for the game. It's 3d-ish. Neon just did a really nice job authoring a map section on a hillside. It did indeed seem about equivalent, subjectively, to Opus. This is impressive because GPT models, including the just-released 6.2 Astra, need extra guidance to do a convincing job at authoring my maps within their constraints. Without special prompts, GPT will build indoor spaces with sloping floors that follow the hillside they're built on. It will judge its own work as great when it looks objectively ugly and repetitive. Etc.
Opus does a fantastic job without handholding, and in an early sample it looks like Neon does too. But I've only tested one tile because I hit my 5hr limit. I'm "only" on the $20/mo plan for Google.
Interesting to me that I was inspired to get into swe by wanted to build games, but found it near impossible back in the day. Fast forward to now and I do not even have a WIP game :( what I am doing with my life.
Consider this a push, then! The activation cost is lower than ever. I've got a bunch of trash in progress and it's been a delight - I wish you the same.
The bottom chart (looks like a spiderweb) looked weird, Neon looked like a stub, so I looked into it and only 3/8 of the benchmarks are populated for Neon, whereas 6/8 are for Fable.
Seems like the methodology here isn’t sound? Not sure why we’d care about this over something like Artifical Analysis index.
I’m not a big fan of ai coding, but have been using them because the expectations require it now. I had been using opus 5.5, but I’ve found hydrogen great for basic stuff and neon really blew me away. I’ve mostly been using them for prototype work, which many models do a good job at, but the speed and quality seem better than opus 5.5 at least for me. Genuinely impressed, and now I’m looking for projects I had been putting off that I might be able to have it just handle for me.
With more competitors, the performance increase of the AI models we can use is welcome. Because of this, the performance of the open source models that distill them also looks like it could go up, so I think places with security concerns can benefit too.
-Cyber variants of the models exist but you need to be in Fairwinds and that's AFAIK not easy thing.
The company I'm leaving due to change of projects I am 95% sure has access, but only from a part of the company that has extra security checks and blacked out windows :V
There's apparently a bit of mess regarding intersection of what options you have enabled on your account (as in, workspace license) and country availability
Can someone speak to the sub options and real world allowance they bring, compared to those of of Ant and OpenAI for coding? Is it same-ish, is it very different?
I've been impressed with my usage of Neon so far. Early impression at High and Extra is near enough to Opus 5.5 Extra for me. Usage has been pretty good for that effort level too.
It doesnt matter. If its not at least as efficient as opus 5.5… people want a good model that can be used for good. Astra was also good but would eat through the subscription in no time. The same for Fable 5. Opus 5.5 is the first model that made the subscriptions usable.
I wouldn't trust any particular one, but obscure / indie indexes are going to be less affected by benchmaxxing than the well known ones, so it's a positive sign for Neon that it is performing so well on a lot of them including this one.
Has anyone else run into the issue of all Gemini 4 models showing as "Unavailable" in agy with a Pro sub, despite working in the web chat? And if so, have you been able to resolve it?
Neon just feels so readable. I compared an analysis of a complex HVAC quote against Fable 5.2 and Neon and Neon gave back straightforward negotiating points, no straight-into-invented-jargon headings, and overall gave a response that felt reasonable and proportional to my questions. I’m not surprised that Google’s models can be good at such things, but it’s such a welcome change from these turgid AI responses that feel like wading through an ocean of Fable’s free associations.
How do I use it? Do I have to create a Google Console account and navigate through 15 menus in order to generate an API key? Or are there another 7 different ways to achieve similar outcomes?
The subscription is the most cost effective way to use. But if you want to avoid the monthly cost you can create an API key and pay per use.
There are a few ways to create/use an API key. Through Google cloud, through vertex, but by far the easiest way is to use AI studio. https://aistudio.google.com/ pretty good platform to try out any Google model or just create an API key and use in whatever you want without having to deal with Google Cloud UI.
Alternatively you can get from other sources as well such as openrouter.
Is the subsidy good on the subscription? Claude would cost me $10K or so in tokens on what I do in the $200/mo plan. It does good work but the written communication from it hurts my head. Open to trying a new provider.
Gemini 4 Hydrogen is the first Google model that has lived up to its benchmarks for me. I'd rather use and trust Opus 5.5 more for planning, but Google's latest models seem fine for implementing, debugging issues, and code review. It's the one decent model that sits below Cursor's threshold for charging $0.25 more per million tokens for enterprise users. Neon would be as good at or better than Hydrogen, which is obvious from the naming scheme
We all know there must have been 15 different product managers meeting endlessly over this and thinking they are making the world a better place by inventing new nonsensical names.
Based on my usage right now (maybe 1.5 hours of 3-4 parallel active sessions) it seems like the value/usage is like at least 10x more than anthropic... I'm still at 0% weekly usage on the Google sub now, I ran out of my weekly Claude x20 sub limit in one day... this might be the best value sub there is if this holds up
The harness itself (antigravity) is a bit sus right now, and it's the only one you can (apparently) use for the Google sub, but so far it's been decent
Speaking of marketing, Anthropic has claimed they will never have ads. They have also claimed that they won't sell user data. I have a feeling I know the answer, but has Google ever claimed something similar?
I work at Google and work exclusively in the CLI version of the agent harness. I run 20
at a time (it’s nice that tokens are already in my budget). I have about 85 skills and plugins loaded and can nearly automate all of my job from there.
I give it work to do for 12 hours and come back in the morning to a stack of fully end to end validated commits.
makes sense, when i tried argon agy just didn’t work. half of the time it didn’t show argon as a valid model, and the other half of the time the harness would just break in various ways
i decided not to chargeback since i didn’t want my google acc banned, accepted losing that money, and that was my one and only (and ever) experience with google’s ai stack.
It was hot garbage not so long ago but it's getting better very quickly. They actually have a decent-ish working ACP mode now too and finally you can see usage levels without having to bring up the TUI. Still no multiple profiles our non-interactive auth.
(I don't use the CLI directly so these are what I noticed they improved and didn't)
Hmm. Based on this, I dropped a preliminary specification for an RF transceiver design into Neon. I've have a couple hours work into already with GPT-6.2 Sol/high.
This is really good. Worlds better, in fact. I'm impressed and delighted. Far better insights into power management and clocking. Component identification is awesome; far better than what I've been getting from Codex. It's like a friendly 30 year EEE just decided I'm worthy of as much help as I need.
I start with detailed analysis before I go near EDA tools. The hardest part of design is identifying components that distributors actually stock and sell. There is no point to making schematics for fictional components. Once that part is nailed down, I use kicad-happy to analyze my schematics and PCBs. I haven't felt the need to let the LLM actually operate the EDA itself.
I already used up the full quota for Neon. It is fine and close to Opus 5.5 on high to xhigh. At least the new quotas are much larger with the new models and you can get something done. I optimized multiple deep learning models and integrated them into an application pipeline and everything worked pretty much like Opus (almost).
Is it just me or all Google models are terribly bad?
They all seems benchmaxed to hell and producing total crap compared to Claude or openai models from my own experience (mostly coding)
After hearing some good things from my tokenmaxxing buddies (or whatever you want to call them, AI enthusiasts), I got a $20/mo sub to try out Neon and it's just breezed through some hardware graphics bring-up problems I've been struggling with on Sol/Astra and Fable. I'm gonna drop my $200 OpenAI sub this coming month in favor of Google's $200 plan just for more Neon tokens.
Interesting. I've been considering switching my OpenAI sub for Google also. I think a lot of the variance in results has to do with what programming languages you're using and what type of projects you're doing, but for me, on Swift/SwiftUI work, I've done exact head-to-head comparisons of OpenAI models and Opus 5.5, and Opus is so much better, so consistently, that I'm really wondering why I bother paying OpenAI $100/month. I use ChatGPT for some ancillary projects and personal stuff, but I don't even come close to using 100% of my plan. I might give Google a try.
> I've done exact head-to-head comparisons of OpenAI models and Opus 5.5, and Opus is so much better
I am just curious, have you always noticed this trend or is it somewhat more recent-ish? I swear, a year ago or so, I was having the complete opposite experience, but something changed in the 3-6 months. I do know nor understand Swift UI, so I have no opinions on the matter, but my Swift experiences have been the nearly identical. I felt like GPT used to be so much better with Swift (and C#) than Opus.
I can't speak to the historical trend because I've only been doing substantial SwiftUI work with these models for about three months. But in my head-to-head comparisons, it's not strictly code where Opus has been better. Codex has made some very questionable choices in app architecture, etc., which I don't see from Opus.
It's just you. Since Argon, the Gemini models have been producing good results for me side-by-side with Opus/Fable models on the same prompts. Not always winning, but increasingly so with Carbon/Neon.
I think it's just you. Carbon is excellent but extremely slow. If Neon is better than or equal to Carbon while also being a lot faster, that'd be a great model. I haven't tried Neon, yet, but they seem to have understood the assignment with the latest round of models.
this week I ran out of codex astra quota and switched to antigravity, boy is Gemini Carbon amazing model! It actually helped achieve a breakthrough I was stuck on for weeks with Astra.
And tokens/task is very high. They don't seem to have figured out making the models usable in a typical interactive workflow, ie., they're in a Opus-4.7 era.
i already went through the whole quota for Neon, it is on high to xhigh almost as good as Opus 5.5, almost equal. Most people who dont do deep learning model rewrites and other math stuff and statistics like me wont probably even notice a difference.
Yes. Every Google model so far has been disappointment. And their biggest moat, the integration into other google products has been laughable. Today I've asked gemini in gmail to download all attachments from one e-mail address and it turned out it's unable to do it. It's unbelievable how incompetent they are, Google has such a massive edge over OAI and Anthropic and it's still losing.
I don’t think anyone who have yet commented has used the actual model. But, to be fair to the commenters, google has made it very hard for them to do so without spending a ton of money.
Neon is a really good model. It’s currently my favorite model to use. Coding-wise, it’s equivalent to Opus 5.5/Fable. Documentation-wise, it is great. If you have access to Neon have it take a crack at simplifying your agents.md. It’s the first model I’ve run that criticized the amount of ceremony in my instructions and suggested solid simplifications. It also makes clean edits to documents without leaving scars of the previous verbiage behind.
Art-wise, it’s more capable, in my experience, than Fable. I’ve been using fable/opus heavily to make art assets for my application (I shared in a previous comment that procedural art has completely obviated gpt-image art for me) and I’ve started to run into a wall with how far I could iterate with fable. I had neon take a crack at the assets I struggled to improve and it did a grand job.
In terms of working together with a model and how it communicates to you, I love this model. The model knows how to clearly communicate to you. It is such a breath of fresh air over Opus and Fable which, despite the improvements, I feel is actively trying to damage my brain with its unnatural prose. It's hard to describe how it feels to have a model that you feel respects your time and energy with proper writing, and the subsequent delight that arises from the feeling of wanting to engage in smaller details again.
Neon has a big model smell, but the miracle and irony is it’s supposedly based on a smaller base model than Carbon. On google’s model page, they previously wrote that Neon shares the same base model as Hydrogen! It is thought that Argon and Carbon share the same base model.
My impressions will probably change as I use this for longer than a day, but my initial impression is: I don’t know how google pulled this off.
I used it and was shocked at how much better the test tasks I ran it against worked.
I made a little stock gamelet and it came out great vs the same prompt being given to 6.2 sol at ultra.
I think the previous google models were inferior but neon is much better.
That said, I use it via antigravity with a plan so I don't know the cost perspective for someone who is paying per token.
no doubt Neon is a capable model but for us paying $200/month nothing can really match Codex or Claude's usage. Like I have a lot of trouble paying for API to work on a new project, most of the time I am brain storming trying different things until I arrive at a solid direction and putting a meter on it feels prohibitive vs a fixed "all you can eat" plan.
but seeing these subsidized plans are mostly enterprise acquisition schemes I do feel like they will slowly turn into pay as you go type and now that China's access to nvidia gpus is curbed, I think we'll see less and less investment into spending too cheap to meter models as the incentives disappear
All in all I want to say, build whatever you want right now because when there are lot more vibers tomorrow, they won't be able to support subsidized plans. Fact that Google hasn't budged on this tells me that they see no upside
Maybe someone can help me clear this up: I've got a $20/mo subscription to Claude that's not through a console, and which I just use as a chat/coding window in VSCode. Then I have an API key that I use with OpenClaw in a sandboxed system. Both are using Opus 5.5.
A good 4-6 hour session of code and tool usage on the API can run me $80-$100. Are you saying I can get that all for $200/mo if I switch to Claude Code and use the high end subscription?
Yes. My Claude Max 5x sub gives me approx $2,500 worth of tokens per week ($10,000/mo) for $100/mo. I think 20x is roughly double that.
so ... I just have to give up the openclaw harness? I guess my understanding atm is that they'll reject calls to a subscription method if they're coming from a claw.
I wouldn't say my relationship with the claw is worth an extra $500 or something, but there are serious improvements over using Claude Code directly. Scheduled jobs, the whatsapp channel, and a bit of personality... I'm disinclined to hunt around for another harness that Anthropic would accept on the terms of their subscription, only to find it banned shortly thereafter. Is there a well trodden path to just using openclaw on one of those higher level subscriptions instead of raw API pricing?
The $100 and $200 Claude plans give you $100 / $200 of API credit per month.
https://platform.claude.com/docs/en/about-claude/api-credits...
So you can get massively subsidized coding use (I use it through T3, herdr works, the official clients work too of course) and also get a modest amount of API credit for whatever you want (claw, etc.).
The grey market may or may not have some solutions for this. Usage is at your own risk.
Theres also options to connect openclaw via the real cli, but that does blow up your system prompt. I use that option right now and it works quite well. There's no free lunch as they say.
The Max and Ultra personal subscriptions also include an equivalent amount of API credits each month.
What? You can buy a $200/m subscription from Google too?
https://gemini.google/subscriptions/
$99.99/ month: 5x higher usage limits vs. AI Pro
$199.99 / month: 20x higher usage limits vs. AI Pro
Usage limits in Antigravity, I take it. So you can't take it to your harness of choice, or e.g. run batch image processing jobs. For $200 per month, that's pretty lousy.
Yeah that's pretty shitty. What's not shitty is that you get like 10-20x (maybe more) usage per dollar compared to the anthropic subs. I've been running 3 neon sessions for the last ~hour and I'm still at 0% weekly usage... my anthropic 20x pro sub weekly limit just ran out after about 1 day of semi-active use...
Let's see if this lasts though...
Not really. If you want to do anything fancy, you have to pay per token. And that includes using a harness that isn’t google’s
Anthropic does that as well, right? Yeah it's pretty ridiculous.
Claude subscriptions still work fine in 3rd party harnesses. They tried to remove support but ended up not doing so (yet).
See: LiteLLM.
you forgot to mention that when you ask antigravity the wrong question, it will ban your whole Google account.
Which also includes a ton more real-life benefits, not just purely a model (which funny enough, you also get access to other models too).
"I don’t think anyone who have yet commented has used the actual model. But, to be fair to the commenters, google has made it very hard for them to do so without spending a ton of money."
I don't know why that would be.
The Neon model is available today in both my Gemini app on iPhone and at https://gemini.google.com/app, and I just have a $4.99/month Google AI Pro trial offer.
I presume I could boot it up in Antigravity also, but I haven't done so yet.
Just tell me how to fit the Gemini app into the pipeline I currently have for transcribing historical handwritten documents. After a lot of wheedling, I got the chat model to admit that no, that isn't actually possible, if I want to do something that isn't playing around in their web app I need to shell out for API credits. "Oh, but you get free API credits with a subscription!" the bot tried to sell me. How much? $10 of credits per month for a $20 per month subscription, it turns out. The Gemini app is a toy.
Google has subscriptions equivalent to openai and anthropic - Google AI pro/ultra, etc. You can download google antigravity (Google's CLI harness) and use that with the subscription.
To be honest the Google AI studio, and all of the related recent Google web UIs and related surfaces are absolute dumpster trash, just finding where/how to subscribe and set this up was so unnecessarily painful, and the UI is so painfully slow and buggy- holy shit.
That being said I just got neon running via antigravity (agy), and it is quite good. Noticeably faster than opus 5.5, and so far it seems like I will get at least 10x usage compared to my anthropic sub of similar price.
Ahh is also absolute garbage compare to Claude code. It’s stunning the difference in ux between the two.
My biggest gripe - I have thoroughly tried to set up permissions correctly, but it still asks me permission for every tiny thing it tries to do.
I feel like I'm a dinosaur when I manually type gemini.google.com/app to open the native app on my phone. It's already installed magically, why do I have to rely on an open-in-app cue to boot it.
> google has made it very hard for them to do so without spending a ton of money.
I dunno... I could afford to drop 20 bucks to try it out, and did so. Not sure how it's different than Anthropic or OpenAI that way.
I see, that's good to know. I was previously using a version of Gemini Pro that had an extended free trial (I've since upgraded to ultra). In my version of pro, I had none of the modern models available to use. I am happy to share blame with google for this confusion.
Non-paying subscribers of Gemini Pro don't get access to the recent models.
How can you be a non-paying subscriber to pro? Google's pricing and packaging architecture makes no sense as usual.
Capacity bottlenecks made them not include certain models in "promotional" accounts, i.e. ones that are currently on various promotional sales accounting.
There is a lot of those, one could say is that Google is trying to curb their previous overgenerosity
I have AI Pro as well and have Hydrogen and Neon access. Just no Argon or Carbon
Weird. I see Neon, Argon, and Hydrogen.
(Argon I think was just retired so that may be why)
I still see it on OpenCode Go
I can't even use my bank cards for Google (normal Visa or Mastercard debit card) in this country because it flashes up a mysterious error message about "temporary cards not being allowed for purchases". This has been going on for about four years that I know of, across at least three banks, both Visa and Mastercard, and none of them are temporary cards (normal expiry date several years after the card is issued) and I don't have trouble using the same cards for any other online purchases including OpenAI or Anthropic.
Google just makes it weirdly hard to give them money.
Just because you were able to pay them $20 and test it doesn't mean other people have that experience.
Outside of their ad business googles enterprise offerings have always been incredibly hit or miss. It just isnt an enterprise that values support or stability.
> Not sure how it's different than Anthropic or OpenAI that way.
It writes text that is not some complete gibberish.
Haha, very true. That's honestly why I'm probably switching to it from Claude. As far as I can tell, it's at least as good as Opus 5.5 at writing code, but the way it communicates back to me is so nice. I had it clean up some READMEs that had vibe-coded sections that had leaked in over the months... it found those sections, reworded them (dropping most of it, honestly), and told me exactly why when it was done.
If I had to guess it probably has all the research optimisations that went into Gemma and whatever new Gemma they have coming out. Probably saves a lot of memory and speed.
It's an unbelievably good model. Seriously.
I really like the communication, anthropic is noisy, grok basically won't talk to you, but neon is concise and clear.
I left it overnight working on a massive project I have and am extremely pleased with the result. It has made several mistakes (ones that Claude likely wouldn't make), but it seems to have built faster and waaaaay cheaper than Claude.
Fantastic.
What kind of assets are you working with and how are to improving them?
I'm working on "hand-drawn" objects modeled in 3d. Regarding iteration, for example, I've struggled to get fable to generate a mug that has a geometrically correct handle with a natural taper. I'm sure fable could have gotten to what neon got me but after a certain number of iterations it wasnt worth the time and tokens.
It has been available if you pay for API since day1, right?
It's also available on $20/month Gemini Pro accounts.
It totally replaced Argon there, which is no longer available.
Interestingly, Argon used to run into rate limits all the time on $20/mo plans; you'd get three or four prompts, then you'd have to spend several hours in a cooldown period where you could only use an old 3.X Flash model. Neon is either much more efficient, or its limits have been relaxed. I get quite a lot of use out of it, and haven't yet been cast into cooldown purgatory.
It has only replaced Argon completely in the Web UI, it's still available in agy with separate usage like before.
I'm really happy with the Pro plan right now, Hydrogen and Neon have been great and actually usable day one. 5 TB of storage for my family and pretty decent AI usage for $10 a month through Verizon FiOS add-on. Much better value than the ChatGPT offering.
Totally agree. AI Studio seems to go on forever using the best model (per your subscription plan). Anrigravity (desktop) and Anrigravity (IDE) to run the Orchestration etc to get into local mode for the weeds or run your local data pipeline, uses your quota and then GitHub sync back into AI Studio for seemingly endless quota at the top model where it easily handles massive context. AI Studio is smart to build environments that easily work when in local vs studio (web) modes. Neon actually seems better than Argon, which was great for reasoning.
Super stupid question. I pay $20/month on Google (for extra storage, etc). I also paid Claude Code ($200 per month). Leaving aside the fact that Claude probably will give me more tokens, is there a way I can replace my whole flow of Claude Code with a CLI by Google (that also works well with Herdr)? I feel Google makes using their models very very hard!
The official replacement would be Antigravity (https://antigravity.google/) which works relatively well with and is explicitly supported by Herdr (https://herdr.dev/tui/docs/agents/).
A word of warning though, it will need a bit of getting used to if you've used Claude Code extensively.
I gave Antigravity a try, coming from Claude Code. To list a few differences, agy does
* not depend on the folder/repo, iow resume in an other folder is fine
* not have a an auto mode, only feasible way is YOLO mode. The sandbox seems nice and I’ve tried and tried and tried but always got stuck in endless confirmation prompts
* not have --ide integration with vscode that in manual mode shows the pre-write diff
* not integrate as clean with the browser as claude-in-chrome does
> not have a an auto mode, only feasible way is YOLO mode.
There's an auto mode. You can Shift-Tab a few times to rotate between Default mode, "Plan mode", "Auto-edit mode", and "Auto mode" (or `agy --mode=auto`).
You can start it with --mode auto, but it's useless. It will still grind to a halt and ask permission for tons of innocuous things.
I basically have to run it with --dangerously-skip-permissions, which is batshit nuts. Google, fix your harness.
Care to eexplain why it will need a bit of getting used to it?
Edit: Already installed it and using it through Herdr! I realize what you meant ha!
can you share with the rest of the class?
First impression (though I haven't used it for code yet) is that is less verbose and executes less things between messages, or at least doesn't show all of them or show them in a more compact way.
With Claude Code I feel every answer triggers a lot of actions and agents, while with Antigravity seems there are less of that and it's more straightforward.
I just ran into this Argon rate limiting last weekend - just as you describe. Is there any documentation about how much I can use Neon in the app or browser? I have this $20 plan and it became useless recently because of the rate limits.
but is it as cost effective as Opus 5.5?
I have paid $200/year for Claude Pro sub and almost never runs out of either session or weekly usage limits. Not sure if Neon can last that long?
>and almost never runs out of either session or weekly usage limits.
You are using it gently. Try using it with heavier workflows and subagent or run getshitdone and you will spend the weekly limits in an hour.
I love that we are years into LLMs and Google still doesn't roll out models in a timely fashion.
I was interested to see what Neon was like so downloaded antigravity. Signed in with my personal account, but it's AI Plus and not AI Pro (I just wanted it for the storage, but I assumed I got _some_ AI stuff included given the name).
No problem, sign in with my workspace account, nope despite being on a higher tier I just get 3.8 Flash.
I understand they may be very compute constrained, but surely it's worth giving users even a handful of prompts.
Opinions are my own.
> Google still doesn't roll out models in a timely fashion.
This is the same way Google has been doing roll out across products for a decade. And it is unlikely to change.
At the scale of Google, deployments like this would reach billions users, spanning over the world, and take hours, if not days, for a proper roll out. An instant deployment would carry a lot of risks.
Personally, I see it as engineering done right.
To be clear I'm not talking about the main Gemini UI. Which ironically I have access to hydrogen on my accounts.
I'm talking about antigravity - their coding agent, which has orders of magnitude less usage than Gemini or AI mode.
But tbh Gemini has the same issue, this is a rare exception.
I'd just have thought that given how little dev mind share Google has you might want to give paying (!) customers timely access for a few prompts to evaluate it.
Considering how much OAI and ANT have captured the market I find it hard to believe Google has that many people who will really give their models that much work on paid plans
>And it is unlikely to change.
and this is the reason google will lose the race. other companies have figure out the way already.
also with the "#1" models they have why cant they solve this problem too?
And yet somehow Anthropic and OpenAI manage to do it
Note that OpenAI's corresponding tier, ChatGPT Go, doesn't even have Terra (the Hydrogen equivalent), let alone larger models like Sol (the Neon equivalent).
Chatgpt go is clearly not meant for normal development work.
At most it'll be useful for a kid leaning programming, or someone playing around with less coding-focused jobs.
For someone who's limited to ≤$10/month, nothing compares to opencode go.
Not true, I'm talking about antigravity not the main UI. ChatGPT Go does include (a small amount) of these models.
> If you have access to Neon have it take a crack at simplifying your agents.md. It’s the first model I’ve run that criticized the amount of ceremony in my instructions and suggested solid simplifications.
that is a very compelling endorsement...
> Neon is a really good model. It’s currently my favorite model to use. Coding-wise, it’s equivalent to Opus 5.5/Fable.
When I read this type of thing I do really start to see why Anthropic et al are putting so much effort into the harness. I have a lot of hooks and things setup to make Opus/Fable tolerable but it is tolerable. The same but change the harness is not really a great selling point. I have definitely talked with fellow devs who prefer OpenAIs models because the word barf out of Anthropic models can be exhausting even if they don't code as well or whatever.
Admittedly since OpenAI/Anthropic models are serving my needs and my past experience getting access to Google's models was a pain in the ass I am less inclined to mess around with Googles shitty consoles and API permissions.
> that procedural art
What tools are they using to make this?
They arent using external tools. It's all custom typescript drawn on canvas.
Do they require specific instructions to do this well?
I'd enjoy seeing examples (that also includes the prompt).
Gemini and Fable both generally perform well when provided a generic prompt like "draw a picture of an abstract landscape using javascript".
Can't share my work but my process involved a lot of planning which included discussing specific feelings i wanted my art to evoke, providing a collection of images (a moodboard), and having the model generate multiple mockups to establish direction and procedure. That took a lot of different iterations but once I finally established the style, the model had no problem transferring it forward.
google has done something very wrong on the marketing side, i know a lot of people in software engineering and no one i know uses their models or is excited about anything they do
I use gemini almost exclusively. Our CEO uses claude and it's like 90% wrong. Maybe he prompts it wrong?
I can pretty much one or two shot everything in gemini. Not sure if that means i'm working on simpler stuff or I’m just better at describing what i want. But i'm also using like orders of magnitude less tokens than everyone else.
Because up until now it hasn't had the models to make it worth trying.
well it's a bit late now
People talk constantly about defecting between plans. The grass is always greener, there's always something worth bitching about, and things fly fast in the thunderdome. If any entity comes out with something worth using, it'll get traction.
https://killedbygoogle.com/
That's why.
it's pretty easy to use inside of cursor
I have a "fringe" theoretical physics idea[1] that I've been grinding through various frontier models to see how far they can push this kind of thing. It's a great test because it is sufficiently "out there" that they can't memorise the answer, and it has absolutely brutal mathematics.
A simple cross-check is to get two models to advance the paper, and then have the other model validate it.
All of the Gemini 4 models -- including Neon -- consistently make simple, "doofus" errors like neglecting minus signs or doubling a scalar factor due to accidentally multiplying by it twice. Similarly, they often use two different sets of assumptions for a long mathematical argument involving multiple proofs, so the whole argument ends up invalid.
For comparison, these basic errors no longer occur with any of the recent Anthropic or OpenAI models, including every GPT 6 Sol or Astra model that I've tested.
Simply put, if you let Astra do 5-10 iterations in a row (~1 hour of work) and feed it to Gemini 4 Neon, it won't spot any errors. Do the reverse of that and even GPT 6.2 Sol will tear Neon's work to pieces.
PS: I suspect that Gemini models have a higher temperature setting, which makes their English sound more natural, but introduces errors that degrade their hard science capability.
[1] But not crank physics, which I would expect to "trip up" models trained on correct, valid physics and mathematics.
Were you using any / commands when it made doofus errors?
I also found Argon made a ton on doofus physics mistakes. All caught by Astra. But it got better (and much slower) when using /teamwork.
what is this idea, for the curious?
A variant of the notion that the standard model of physics is "simply" the property of some high-dimensional algebra. For example, Cohl Furey proposed a few different variations of the Octonions such as the dual-Octonions and I believe most recently the Dixon algebra ℂ ⊗ ℍ ⊗ O. There are other groups investigating Lie algebras related to E8 and the like.
The idea is that the Standard Model is soooo close to a nice and neat algebra that it may just be possible that it simply "pops out" of one naturally. All of particle physics might just be a very simple mathematical rule, instead of a whole pile of things on top of other things that looks suspiciously like epicycles.
There's a "search" by theoreticians for the equivalent algebraic change in perspective that Galileo did when he moved astronomy from overly complex 2D mathematics to a far simpler 3D algebra.
I have several notions that now -- thanks to AI -- I can throw at the proverbial wall to see if they stick.
The one that I've been playing around with the longest is a real geometric algebra with a total of nine dimensions: three temporal, three spatial, and three nilpotent ("matter"). I also sprinkled on some of my preferred postulates, such as making the theory MWI-flavoured, fully relational (no absolute clocks or rulers), etc.
I knew it probably wouldn't work, but this algebra has an incredibly beautiful symmetry of three-by-three and it is "just" big enough to encompass the SM particle zoo including the three particle generations!
The idea was that the three temporal dimensions might explain the "many worlds" aspects of quantum mechanics.
It almoooost worked!
I got it to the point that all three frontier AIs agreed that it's publication-worthy and one particular branch was "ahead" of Cohl Furey's favored algebra, but... meh... it's (provably) a dead end. Not to mention that everyone and their dog will be doing the same kind of thing with their own toy theory, so nobody will be interested in yet more AI slop output unless it makes a dramatic discovery.
Despite the dead end it was worth a go, especially given that trying is effectively free: I usually burn tokens on this right before using a "reset" that's about to expire.
> I got it to the point that all three frontier AIs agreed that it's publication-worthy
Lol, do they still gas you up like this?
Glad you asked!
It's actually a great test to see how much a particular model likes to "butter your bread" versus giving a negative opinion.
Anthropic models: Somewhat brutal actually. Fable told me to "stop wasting my time".
OpenAI models: Incredibly diplomatic and neutral, refuses to state a firm opinion either way even if pressed.
Google Gemini models: "The honesty of this paper refreshing, this should be the benchmark for how all theoretical research is published!" (etc)
I learned not to trust anything a Google model says and to use the combination of an Anthropic and an OpenAI model if I want to know if something really is a bad idea.
If it's worth publishing, OpenAI will snipe it in a few months.
What's your score on John Baez test?
https://math.ucr.edu/home/baez/crackpot.html
Sadly I only get -5 + 10 points for "I'm not good at math, but my theory is conceptually right, so all I need is for someone to express it in terms of equations".
I'm literally using AI to paper over my sub-standard mathematics skills!
Trust me, there's hundreds -- if not thousands -- of people doing this right now, and many of them likely work in serious physics research programmes.
Many people have ideas, not many people can dedicate their lives to them.
Now, it becomes entirely possible for someone with a "notion" to make the next "big" discovery, but not actually understand what they have discovered.
It's wild.
> But, to be fair to the commenters, google has made it very hard for them to do so without spending a ton of money.
Google has an introductory $5 subscription that defaults to Gemini 4 Neon (High). I was curious, so I grabbed that and got things going within 5 minutes.
I agree WRT Carbon, but not Neon.
Neon's only available in the $20 tier, unless you were offered some new user promotion or something. As far as I can tell anyway: https://one.google.com/about/google-ai-plans/
It's the first sentence of my comment...
> an introductory $5 subscription
I keep seeing opus 5.5/fable. those are two separate models. two different costs. so pick the one that makes sense.
Neon is a contender. Neon "high" appears faster and cheaper than Opus 5.5 "medium" with similar results on long-running language porting work in a large monorepo.
The speed is the biggest win IMO. When Carbon came out, I tried to give it a shot, but it was so slow that I went back to Opus just for productivity, nevermind quality. Today I started Neon on one worktree and Opus on another, and Neon did so well that I have now switched for both.
Code quality looks great. Comment-wise it has its own Claudese(exclamations like "this function has a concurrency bug and it's a sharp one!"), but overall I like reading it more than Opus. Might just be because I am not yet tired of its quirks though.
Oh, after a productive day on this particular work I had only spent $30 of Neon vs. $70 of Opus yesterday. YMMV
[disclaimer: Googler]
My favorite current-gen Gemini-ism is how it has these epiphanies, like "Wait! There's a pointer to the documentation we need right here!" So cute.
Then you would love the thinking traces from Chinese models like GLM-5.3-Flash.
From Qwen:
> I'll write the script now. Let me write it. I'll write. OK let me just do this now. I'll write it. Let me write it. I'll stop. Writing. OK.
> my decode of #1 said off=0x1281. Hmm - hmm - hmm. Hmm. Hmm, hmm - but the displayed size was 22255... hmm
(I think this is my quant's fault tbf)
The “stolen CoT” paper has some thinking traces of closed frontier models.
The ‘summarisers’ mask a lot of raw thinking you’d probably be frustrated by too.
Which Qwen? For Flash-Next there are several tunes that reduce the thinking tokens, like ThinkingCap or Swift. I've been running a Frankenstein version of the latter on my DGX Spark.
You gave the model ADHD. My brain might just be over-quantized.
it's not your quant, it's just how that model works. it gets way crazier than that, just run a long session
Deepseek: But wait! Maybe I am wrong and user is right? Let me investigate further.
I've been using Argon steady since it was released and a few times it's gotten in this really excited about everything mood like that. Most of the time it's totally even keel. Cute is a good word for it, not sure what's going on under the hood.
Yeah, I got a lot "Look at that!" with Argon.
I added a rule telling it not to use exclamation marks. Game changer!
Does it use them in the thinking phase? What if your AI isn't as motivated to keep going because you don't let it get excited?
Faster and cheaper, and apparently better than Opus 5.5 medium? Bro I just got my mind blown by Opus 5.5, now Neon gets to once again shift my mental benchmark and compel me to finally download and try Antigravity? Right after I migrated from the Claude Code TUI to the desktop app and never looked back for all of the 2 weeks it's been?
I'll do it just for the improved prose. I've had it up to here with Opus 5.5's surprisingly poor writing skills, even though it's leagues better than the Willem-Dafoe-on-cocaine dumpster fire that is Opus 5.
If you watch Youtube much at all, Youtube Premium Lite (no ads) is well worth the $9/month. This is included with their $20/month AI plan. Spending that other $11/month wasn't worth it before Neon, but it might be now.
There are 3 different antigravity quotas:
* gemini (previously just Gemini 3.8 Flash)
* claude and gpt (very small quota)
* argon (very small quota)
The Gemini quota now includes Neon and Hydrogen. Hydrogen (medium) competes with DeepSeek (high). Its within 2x on cost and more intelligent.
This is all from benchmarks- I am only just trying these models out now.
How much neon usage do we get on this plan? I have that plan and agy is uselsss as a single prompt burns my entire Claude budget, argon gets me like 3 chats, and im definitely not using whatever flash model is listed for coding.
I'd be surprised if anyone here doesn't have uBlock Origin blocking all ads anyway - not only on YouTube.
I am happy to pay a fair cost and not have ads. If everyone blocks ads the content isn't going to exist or if enough block they are going to work harder to block ad blockers.
I have paid for Kagi and Spotify for a long time. Its tragic that most of the web only offers a large monthly subscription to read a single article or intrusive ads- I do use Zen on Desktop and Orion on mobile due to that.
Even if you pay, you will have ads. Ads are now embedded into the content. So much content is in fact just ads masquerading as content to steal your attention. A lot of content out there simply should not and does not need to exist. It is just an endless rehash of pre-existing ideas with the sole purpose of beating the algorithm. The model is fundamentally broken imo and needs to be disrupted.
there’s plenty of non tech-savvy folks who let their devices play ads to keep their kids distracted
Can’t use uBlock Origin on Apple TV, Firestick, etc.
You can on FireTVs.
Use Smartube: https://smarttubeapp.github.io/
you can
How?
https://smarttubeapp.github.io/
I'm reading the plan features and it says "YouTube Premium Lite individual plan6 Most videos ad-free, offline, and in the background". So you still get ads with some videos? The next plan ($100) has actual youtube premium no-ads
You don't get the ad-free music streaming (which includes music videos), and it includes ads on Shorts IIRC
I've been using neon for like an hour after getting the Google AI ultra sub... so far it seems like it could be at least 10x the value of an anthropic sub...
> This is included with their $20/month AI plan. Spending that other $11/month wasn't worth it before Neon, but it might be now
This is their golden ticket.
Amazon famously bundled a variety of services under Prime which had, at least for a time, absolutely unheard of retention metrics. Truly otherworldly stuff. They probably still do.
I've been a Prime member since inception, longer than practically any other service. Given my household's needs I'm unlikely to cancel it any time soon.
Google bundling YouTube and Gemini under their... whatever-it-is-called plan (One? At least that name was a sufficiently thematic copy of "Prime"...) is brilliant. The $20 plan is a no-brainer for a ton of folks. They need to make the $200/mo plan a no brainer, and I think they can, but I doubt they have the wherewithal to actually figure it out and pull it off.
The market has rarely, maybe never, been more amenable to otherwise monopolistic bundling, I think they'd be crazy not to try and swing for the fences.
And I'm not saying any of this is a good thing, I'm certainly not endorsing this behavior. It's just that if I were a soulless automaton C-level executive trying to maximize shareholder value at all costs, and looking at the-company-most-likely-to-dethrone-me's IPO earlier today, and had Neon hot out of the gate, pulling away from the pack, I would be going for blood right now.
Well I'm not that C-level executive, so even though I do use Google models today (they're really superior at historical HTR), such slimy bundling practices makes me see if the Chinese models can be wrangled into replacing them (maybe with the right few-shot examples...)
For probably most people, 200$ per month is already a no brainer, just in the other direction.
Fully agree. Neon is my main model right now. I have a pretty advanced setup with multiple agents, coordinators, etc. It's basically a "factory" customized for me. It ran on Claude for a long time, then on Codex for the past three months, and now it runs on Google models. Neon is good, predictable, and smarter than me :)
Agy has a bunch of good slash commands for different use cases. And the biggest thing for me is that it doesn't annoy me with security refusals. It helped me with my garage door, for example, when another model refused.
Would love to know your setup. Are you paying API pricing ? How else can you switch so easily.
Ok, so I can see that this is a benchmark the submitter made and submitted themselves, which is kinda misleading and should at least be disclosed.
But my question is, what does this benchmark actually measure about the LLM's engineering abilities? It just looks like a page of random bar graphs to me.
this seems to be a composite index based on TerminalBench, CursorBench, vals.ai, etc: https://github.com/git-aarvin/mei/tree/main/source/jsonc
looking at the graphic below the table, it does seem like data is missing for Neon which is present for Fable, and isn't penalized for the missing benchmarks.
Yeah that seems to be a huge caveat
Seems fair game to me if the benchmark were actually a reasonable attempt at rating models, but it’s total slop. Like Neon score is based on three benchmarks and Fable based on 6 and that’s hidden.
This is not a real index. Gemini 4 Neon is only covered by 3 out of the 8 benchmarks included. This post is clickbait.
I don’t know. Model seems real good in real usage vs. Opus 5.5 based on my testing
My comment is only with respect to the index.
Maybe but the sentiment of people who have actually used it is more important than these benchmarks anyway and from this thread that seems to be highly positive.
I have a little WIP game (don't we all...) and I use models through MCP to author maps for the game. It's 3d-ish. Neon just did a really nice job authoring a map section on a hillside. It did indeed seem about equivalent, subjectively, to Opus. This is impressive because GPT models, including the just-released 6.2 Astra, need extra guidance to do a convincing job at authoring my maps within their constraints. Without special prompts, GPT will build indoor spaces with sloping floors that follow the hillside they're built on. It will judge its own work as great when it looks objectively ugly and repetitive. Etc.
Opus does a fantastic job without handholding, and in an early sample it looks like Neon does too. But I've only tested one tile because I hit my 5hr limit. I'm "only" on the $20/mo plan for Google.
> I have a little WIP game (don't we all...)
Interesting to me that I was inspired to get into swe by wanted to build games, but found it near impossible back in the day. Fast forward to now and I do not even have a WIP game :( what I am doing with my life.
Consider this a push, then! The activation cost is lower than ever. I've got a bunch of trash in progress and it's been a delight - I wish you the same.
The bottom chart (looks like a spiderweb) looked weird, Neon looked like a stub, so I looked into it and only 3/8 of the benchmarks are populated for Neon, whereas 6/8 are for Fable.
Seems like the methodology here isn’t sound? Not sure why we’d care about this over something like Artifical Analysis index.
There is no methodology here. Vibes based fake index.
Neon performs well on popular benchmarks, but I found these lesser-known benchmarks interesting too. It's a very strong model.
Surge AI DAYJOB - Finance: https://x.com/HelloSurgeAI/status/2123561368473485314
Parallel Search Intelligence: https://x.com/p0/status/2123497882322141453
SlopCop: https://slopcop.com/power-ranking
Cognition: https://x.com/cognition/status/2123553121653944391
I’m not a big fan of ai coding, but have been using them because the expectations require it now. I had been using opus 5.5, but I’ve found hydrogen great for basic stuff and neon really blew me away. I’ve mostly been using them for prototype work, which many models do a good job at, but the speed and quality seem better than opus 5.5 at least for me. Genuinely impressed, and now I’m looking for projects I had been putting off that I might be able to have it just handle for me.
With more competitors, the performance increase of the AI models we can use is welcome. Because of this, the performance of the open source models that distill them also looks like it could go up, so I think places with security concerns can benefit too.
Can’t use it for anything defensive audit oriented. Refuses too often. Google seemingly not trying at all on the CVP or Daybreak front.
-Cyber variants of the models exist but you need to be in Fairwinds and that's AFAIK not easy thing.
The company I'm leaving due to change of projects I am 95% sure has access, but only from a part of the company that has extra security checks and blacked out windows :V
We have paid Gemini on our workspace plan and the models available are "Hydrogen" and 3.1 pro.
I really don't get Google's product strategy.
There's apparently a bit of mess regarding intersection of what options you have enabled on your account (as in, workspace license) and country availability
They charge premium API rates for edge models and then wholesale the 2nd gen. It's pretty smart they just don't make it clear to users.
Can someone speak to the sub options and real world allowance they bring, compared to those of of Ant and OpenAI for coding? Is it same-ish, is it very different?
I've been impressed with my usage of Neon so far. Early impression at High and Extra is near enough to Opus 5.5 Extra for me. Usage has been pretty good for that effort level too.
It doesnt matter. If its not at least as efficient as opus 5.5… people want a good model that can be used for good. Astra was also good but would eat through the subscription in no time. The same for Fable 5. Opus 5.5 is the first model that made the subscriptions usable.
When and why should I trust this index over others?
I wouldn't trust any particular one, but obscure / indie indexes are going to be less affected by benchmaxxing than the well known ones, so it's a positive sign for Neon that it is performing so well on a lot of them including this one.
I just had a pretty negative experience with Neon 4.
(Antigravity GUI client on MacOS, Extra effort level)
I asked agy to find bugs and performance optimization opportunities in some code implementing a OpenGL to Metal graphics bridge.
1. introduced four (4) P1 regressions, essentially breaking everything in a very obvious way
2. wrote no new tests for the code it wrote
3. its code failed five existing tests
4. at least half of the bugs it diagnosed were wrong, e.g. false finding entirely - not a bug, correctly working code
5. correctly diagnosed some previously undiscovered bugs, but either wrote partial or incorrect fixes for most of them
6. correctly diagnosed and wrote correct fixes for only 2 out of a total of like 10 "findings"
Disappointed, this sounded like it would be a big win for Google and put them back at the head of the pack.
Has anyone else run into the issue of all Gemini 4 models showing as "Unavailable" in agy with a Pro sub, despite working in the web chat? And if so, have you been able to resolve it?
6.1 / 6.2 sol are still the sweet spot for me. 2 plus accounts and I'm rolling all day long.
Neon just feels so readable. I compared an analysis of a complex HVAC quote against Fable 5.2 and Neon and Neon gave back straightforward negotiating points, no straight-into-invented-jargon headings, and overall gave a response that felt reasonable and proportional to my questions. I’m not surprised that Google’s models can be good at such things, but it’s such a welcome change from these turgid AI responses that feel like wading through an ocean of Fable’s free associations.
How do I use it? Do I have to create a Google Console account and navigate through 15 menus in order to generate an API key? Or are there another 7 different ways to achieve similar outcomes?
The subscription is the most cost effective way to use. But if you want to avoid the monthly cost you can create an API key and pay per use.
There are a few ways to create/use an API key. Through Google cloud, through vertex, but by far the easiest way is to use AI studio. https://aistudio.google.com/ pretty good platform to try out any Google model or just create an API key and use in whatever you want without having to deal with Google Cloud UI.
Alternatively you can get from other sources as well such as openrouter.
Is the subsidy good on the subscription? Claude would cost me $10K or so in tokens on what I do in the $200/mo plan. It does good work but the written communication from it hurts my head. Open to trying a new provider.
Openrouter if you just want to pay as you go at token rates.
https://one.google.com/about/google-ai-plans/ had a 75% off on the Pro plan ($5/mo for 3 months) and limits are decent.
Google's comeback is just wild. So much for 'code red' at OpenAI, it's back to where it was when Gemini 2.5 pro released.
We are not sleeping.
Who is "we" in this case?
It can speak English. I don't really care about model capabilities when I don't have to actively get dumber interacting with the model.
Gemini 4 Hydrogen is the first Google model that has lived up to its benchmarks for me. I'd rather use and trust Opus 5.5 more for planning, but Google's latest models seem fine for implementing, debugging issues, and code review. It's the one decent model that sits below Cursor's threshold for charging $0.25 more per million tokens for enterprise users. Neon would be as good at or better than Hydrogen, which is obvious from the naming scheme
Maybe it's just me, but I'm struggling to keep up with these Google names, Hydrogen, Argon, Carbon, Neon, Flash, Flash-Lite and Pro.
I mean it is Google.
We all know there must have been 15 different product managers meeting endlessly over this and thinking they are making the world a better place by inventing new nonsensical names.
I wonder how generous the Google subscriptions are. Can you use Neon to do the same amount of work that fits in a $100 Claude Max X5 subscription?
Based on my usage right now (maybe 1.5 hours of 3-4 parallel active sessions) it seems like the value/usage is like at least 10x more than anthropic... I'm still at 0% weekly usage on the Google sub now, I ran out of my weekly Claude x20 sub limit in one day... this might be the best value sub there is if this holds up
The harness itself (antigravity) is a bit sus right now, and it's the only one you can (apparently) use for the Google sub, but so far it's been decent
Marketing in HN is getting weird
This one is a bit puzzling, it’s clearly a vibe coded benchmark aggregation but I don’t understand what the goal is.
Speaking of marketing, Anthropic has claimed they will never have ads. They have also claimed that they won't sell user data. I have a feeling I know the answer, but has Google ever claimed something similar?
I prefer to use agents from the command line. Does Antigravity CLI deliver a good experience?
I work at Google and work exclusively in the CLI version of the agent harness. I run 20 at a time (it’s nice that tokens are already in my budget). I have about 85 skills and plugins loaded and can nearly automate all of my job from there.
I give it work to do for 12 hours and come back in the morning to a stack of fully end to end validated commits.
makes sense, when i tried argon agy just didn’t work. half of the time it didn’t show argon as a valid model, and the other half of the time the harness would just break in various ways
i decided not to chargeback since i didn’t want my google acc banned, accepted losing that money, and that was my one and only (and ever) experience with google’s ai stack.
I like it, but I have an alias for --dangerously-skip-permissions
It was hot garbage not so long ago but it's getting better very quickly. They actually have a decent-ish working ACP mode now too and finally you can see usage levels without having to bring up the TUI. Still no multiple profiles our non-interactive auth.
(I don't use the CLI directly so these are what I noticed they improved and didn't)
Nope.
The permissions management is a nightmare, specifically.
If you're a yolo user, it can be fine.
I am a yolo user. It is fine. But I made a custom bash env for all agents that refuses my block list regex patterns.
how are people having Neon as their main model if its just out for one day?
For me it's about 2 days now? It was available pretty much the same moment it the blog post landed
I want to use a top tier model not to write code, but to make plans, delegate cheap Deepseek workers and judge the outpot of workers.
So ability to reason matters more than just writing code.
Would Neon be better than Opus with these assumptions?
I'm waiting for Gemini 92 Uranium.
Hmm. Based on this, I dropped a preliminary specification for an RF transceiver design into Neon. I've have a couple hours work into already with GPT-6.2 Sol/high.
This is really good. Worlds better, in fact. I'm impressed and delighted. Far better insights into power management and clocking. Component identification is awesome; far better than what I've been getting from Codex. It's like a friendly 30 year EEE just decided I'm worthy of as much help as I need.
10/10 will prompt again.
What's your design process? Letting the model edit KiCAD files directly?
> What's your design process?
I start with detailed analysis before I go near EDA tools. The hardest part of design is identifying components that distributors actually stock and sell. There is no point to making schematics for fictional components. Once that part is nailed down, I use kicad-happy to analyze my schematics and PCBs. I haven't felt the need to let the LLM actually operate the EDA itself.
Work with opencode or pi?
No, and they will ban your account
It works fine with OpenCode and Pi, use an AI Studio API key.
I already used up the full quota for Neon. It is fine and close to Opus 5.5 on high to xhigh. At least the new quotas are much larger with the new models and you can get something done. I optimized multiple deep learning models and integrated them into an application pipeline and everything worked pretty much like Opus (almost).
These days I only trust vibes. And it takes about 2 to 4 weeks for vibes to settle.
Every single Gemini model that has come out strong has quickly failed the vibes.
Google models score great when ranked, but massively disappointing when used.
Let's see if this one lives up to it's ranking..
Is it just me or all Google models are terribly bad? They all seems benchmaxed to hell and producing total crap compared to Claude or openai models from my own experience (mostly coding)
After hearing some good things from my tokenmaxxing buddies (or whatever you want to call them, AI enthusiasts), I got a $20/mo sub to try out Neon and it's just breezed through some hardware graphics bring-up problems I've been struggling with on Sol/Astra and Fable. I'm gonna drop my $200 OpenAI sub this coming month in favor of Google's $200 plan just for more Neon tokens.
Interesting. I've been considering switching my OpenAI sub for Google also. I think a lot of the variance in results has to do with what programming languages you're using and what type of projects you're doing, but for me, on Swift/SwiftUI work, I've done exact head-to-head comparisons of OpenAI models and Opus 5.5, and Opus is so much better, so consistently, that I'm really wondering why I bother paying OpenAI $100/month. I use ChatGPT for some ancillary projects and personal stuff, but I don't even come close to using 100% of my plan. I might give Google a try.
> I've done exact head-to-head comparisons of OpenAI models and Opus 5.5, and Opus is so much better
I am just curious, have you always noticed this trend or is it somewhat more recent-ish? I swear, a year ago or so, I was having the complete opposite experience, but something changed in the 3-6 months. I do know nor understand Swift UI, so I have no opinions on the matter, but my Swift experiences have been the nearly identical. I felt like GPT used to be so much better with Swift (and C#) than Opus.
I can't speak to the historical trend because I've only been doing substantial SwiftUI work with these models for about three months. But in my head-to-head comparisons, it's not strictly code where Opus has been better. Codex has made some very questionable choices in app architecture, etc., which I don't see from Opus.
It's just you. Since Argon, the Gemini models have been producing good results for me side-by-side with Opus/Fable models on the same prompts. Not always winning, but increasingly so with Carbon/Neon.
I think it's just you. Carbon is excellent but extremely slow. If Neon is better than or equal to Carbon while also being a lot faster, that'd be a great model. I haven't tried Neon, yet, but they seem to have understood the assignment with the latest round of models.
this week I ran out of codex astra quota and switched to antigravity, boy is Gemini Carbon amazing model! It actually helped achieve a breakthrough I was stuck on for weeks with Astra.
I'm using AI heavily at work. I'm switching all my opus 5.5 sub-agents to neon. I'm quite impressed.
With Neon Gemini is good now.
Neon Gemini Evangelion
And tokens/task is very high. They don't seem to have figured out making the models usable in a typical interactive workflow, ie., they're in a Opus-4.7 era.
Very high hallucination rates too.
It may be a bad bench, but the three going Google models are measured as having the lowest hallucination rates of any llms in history]
https://artificialanalysis.ai/evaluations/omniscience?models...
Argon was bad yeah.
All the 3.x flash models were actually dumb compared to everything else it competed with.
Neon seems promising.
Are you saying Neon is just as bad?
Carbon changed my mind.
i already went through the whole quota for Neon, it is on high to xhigh almost as good as Opus 5.5, almost equal. Most people who dont do deep learning model rewrites and other math stuff and statistics like me wont probably even notice a difference.
Is it just me or all Google models are terribly bad?
Not this one.
Yes. Every Google model so far has been disappointment. And their biggest moat, the integration into other google products has been laughable. Today I've asked gemini in gmail to download all attachments from one e-mail address and it turned out it's unable to do it. It's unbelievable how incompetent they are, Google has such a massive edge over OAI and Anthropic and it's still losing.
Codex can do all the Gmail integration I’ve needed.
Where Gemini shines for me is programmatically accessing YouTube content (engineering channels) and leveraging that.
You mean they are not providing you with their best models for free?
No, why would you think that? I'm paying customer.