Prompt was "read and update the config file with new data". This work on 4.6 takes <2
minutes to read the file, parse the new data, and patch.
Opus 5 Result: 43 minutes of pulling containers, running sandboxes, creating testing suites, which included evaluating the entire repo beyond the scope of the config file.
Opus 5 in xhigh can't do basic math as well.
They dumbed it down to a point where I just cancelled my subscription yesterday.
I used to be a $200 subscriber, dropped to $20 after the fable shenanigans, and use it only when I have no usage left with Codex.
/on The prose is load-bearing unbearable — every sentence feels like it was engineered to sound profound rather than to be read.
I've been using Opus 5 to write a fractal renderer in GLSL today, with pretty good results (better than I could do on my own anyway). It definitely can do basic maths.
I have a theory about this, what if we all became dumber after 4 months of heavy AI usage?
I remember how I enjoyed agents between December and February, something started changing around March.
I thought models are getting dumber, but benchmarks were convincing opposite, initially I thought maybe they're quantizing models for day to day use, but Opus 4.8 and Opus 5 seems worse models than Opus 4.6
You are very much not alone, and I don't think we're all getting dumber -- I kept using opus 4.5 all through the nonsense that was 4.7, 4.8, and 5, and kept having a good time :)
I think we just need to decouple "doing better on benchmarks" and "actually more useful to me", since they've clearly diverged
The economics catching up with the providers in regards to how much compute they can burn per request and have it make sense for them financially?
A sort of model collapse where Opus 5 seems to love throwing out long paragraphs of text and it needs to be "fixed" by changing the output style and other patches.
I'm not sure, it might also catch up to Kimi K3 and GLM 5.3 and the models that I'm moving to from Anthropic.
Benchmarks test whether models can pass exams with a right answer or a green test case. I don't think the models are getting dumber, but they're definitely getting more incomprehensible to talk to. I've noticed this happening almost as a step change with the overuse of words and tics, and so has the broader community apparently. We haven't all been getting dumb at the same rate.
I could stand Opus 4.6-4.8, I was impressed by the initial fable model. Codex 5.6 sol xhigh feels like the initial release of fable. Qwen 3.8 27b feels like using haiku or sonnet (I quickly stopped trying them).
>I thought models are getting dumber, but benchmarks were convincing opposite
>Opus 4.8 and Opus 5 seems worse models than Opus 4.6
After all we've heard about benchmark cheating, I'm earnestly not sure which or whether benchmarks are reliable anymore. But, beyond the models, I wonder if changes to their harnesses and/or instructions dumb them down. I have noticed models change their behavior, even when using the same version/effort. Sometimes for better. Sometimes for worse.
And, I have noticed a model go from really good to struggling. On 4.8 things were going well for a good stretch, so I did not switch to 5 when it came out. Even after hearing complaints about 5, 4.8 was still going well. Then, suddenly over the last few days, 4.8 seems to have nosedived. It feels similar now to the complaints I hear about 5.
In my case, it suddenly started ignoring my design guide, and introducing new fonts etc. It would even use several different fonts and sizes, as well as different margins for similar elements within the same page. It abandoned classes and started inlining styles. It started feeling random and, even after it realized it needed to go back to the design guide, it just continued with more of the same.
There seems to be something that happens after new model releases in both quality and behavior of previous models. It may not be immediately, but eventually there is frequently some regression.
Wouldn't be surprised if there are knobs that get turned as a function of the revenue they might expect you to generate.
I was a 4.6 acolyte from April til the fable drop, lost that quick, cancelled and took a break, came back a month later, tried opus 5 and liked it, so unpinned 4.6.
Results were great at first, and they're still not terrible, but I have noticed a regression in accuracy, so to speak, where I am pointing out issues that are quite obvious in review.
I pretty much use sonnet 5 low/medium when I have a plan to solve a simple problem and depending on scope, opus low/medium for more complex/bigger scope implementation, and only go high when it's very complex or I'm spitballing architecture/solutions and iterating plan. Never go xhigh or max.
The verbosity is insane though, opus 5 documents everything and just regurgitates whatever lead it to the design choice in there, which makes it more opaque because it's talking about something that was discussed once in a session that no one else can see (except their backend ofc)
I don't even try to steer it away from that with harness, because it doesn't work and just ends up agonizing over whether it should write some comment. Three paragraphs waffling on that on verbose output
I did however have it write a script that basically is git add -A -p for comments though, haha.
I'm $20/month, have all my telemetry toggles off, don't really over engineer prompt/context, just some basic skills for repeated patterns.
It was the wrong time for the GP to drop that subscription from $200 to $20, because $200 gets you a metric assload of cognition while $20 gets you nothing beyond what a local model running on your own graphics card can deliver.
Not according to this very story, I'm not. Who's right?
Consistent, predictable behavior is valuable, even more so given the nondeterministic nature of LLMs. Nondeterminism combined with unpredictability might as well be randomness.
I had $310 in (free) credit that I used on fable, and I still had a part of the $200 subscription at that time. You know, subscriptions don't end the moment you click on cancel.
> $20 gets you nothing beyond what a local model running on your own graphics card can deliver.
I'd guess you're deliberately exaggerating here, but still. I've never clocked the actual tokens/second, but I'm on the $20 plan and get ~15M tokens/month for fully utilized weekly quotas (checked couple months ago). Meanwhile the best I've been able to get locally was ~8 tokens/second with Qwen3.6 35B A3B, which is wildly painful for coding sessions and gets a maximum ~20M tokens in a month... if it's going 24/7.
Just wanted to stick some empirical data here, given that statement.
Running 24/7 doesn't make sense though, unless you're providing a service to others. But if it's just you then there has to be time taken to review+test what's being done and craft new prompts. And if that local hardware isn't decent enough it's impractical for anything serious that's interactive. Meanwhile I just take the Claude limits on stride and break, or if a week is pretty heavy then I augment with DeepSeek Flash via OpenRouter (does wonders in a single turn when I have Claude prompt it to handle implementation slices).
Ouch. I get ~45t/s with Qwen3.6 35B A3B, and around ~70-80 with my current model Ornith 1.5 35B A3B. Local models work a treat IMO if you've got decent hardware for it.
With how competitive the LLM field is, it would surprise me greatly if any of these players were doing anything other than trying to make the best possible product. I certainly do not believe they are intentionally training the models to use more tokens unnecessarily.
There is a theory that the verbosity and comments help getting better results with the current benchmarks. So the models are theoretically getting better but in practice they are getting worse.
Theyre training the system to minimize compute,so most likely theyre dynamically downgrading quants in the first few turns hoping to find the cheapest model to run. The side effect may be excessive token gen
This is only true if they can't saturate token production with a model that does less superfluous things. Given that they can (they're hilariously compute strained), having a model that solves tasks more quickly adds way more perceived value to users.
Agreed although the differences between the effort and reasoning is massive. I generally ship 40 hours in three with AI. I could not figure out why my delivery was behind until I started going through the logs. The thinking was extensive, the effort was beyond the original request by a magnitude of 50x
How do you even know that it's consistently 40 hours in 3? What kind of developer ever had that kind of estimation accuracy (unless it's really repetitive) or even focuses on productivity rather than the problem like that? This sounds more like factory work than design or development. I really don't get it.
Recently got approved at work for ChatGPT Pro so I could use Codex.
Blown away by the speed. It feels like using Claude Code for the first time again. I don't think Codex is doing anything revolutionary, just better handling of which requests should go to which model, and having faith in some of the "less powerful" models for more than you would think.
It seems the TUI coding experience is very much an open race. This is motivating me to look at other agents / harnesses as well (maybe Gemini, OpenCode, etc).
Not specifically Anthropic but why are we allowing billing to take place in tokens that are nebulous and fully controlled by the operators who have no aligned incentives?
If I have a user input and then sanitize and inject that into a prompt to do something, I have no idea how much that is going to cost at all and no real way to measure this properly. A parallel example is digital ocean or aws, i can go and measure/limit my compute/fs/memory/startup times/etc and while it can be impossible to get down to the last flop of money allocated - i can run things on a real budget with real constraints, opposed to an LLM where I have to .. prerun a sanitized user prompt through a tokenizer and then ask an LLM to guess what it may do and give token consumption estimates and then act on those in any sane manner for the user?
Perhaps i'm missing something to do realistic and static rails on things but I don't see a serious way at scale to use the token billing model handling things requiring a users free text input short of having to go pander to VC money to throw money at it until someone else figures it out.
*to clarify my rambling...
We should be billed and given controls based on resource usage itself and not an opaque token concept on top of not being able to spin any knobs that control it's resource usage.
The model providers are quite aligned with concerns like customer retention. These arguments only work if there is no competition. We exist in a marketplace of black boxes. There's not just "the one" you must suffer. You have options. You can build your own too.
A per-token model roughly aligns with the providers' costs, and it is an objective measure, so it seems a reasonable way to charge.
I see posts all the time on HN about which models from which providers offer the most bang-for-the-buck, and how to minimize token usage and still get optimal results, so it appears that competition is working.
“Claude, spend the next 10 hours trying to solve the Reimann Hypothesis”.
I agree that incentives are misaligned but there’s several competing model providers. If one gets funny with their costs people will jump ship, especially if the gap between the top 2 labs and everyone else keeps shrinking.
“I can take current sources and tell you how solved this is, but I am not willing to work to a timeframe or to solve things that aren’t yet solved by mathematicians or science”
These safeguards already exist when they get a whiff that you might be using Claude to fix security issues. Doesn’t seem farfetched given the incentives I outlined that they would apply to this kind of abuse.
How loose those controls are becomes a market force.
One guess is that their "primary" target audience/market is the large corporations that get their employees unlimited tokens, and not the individual developer who may worry about spending and token accounting.
It's the opposite. The enterprises have all the tooling to monitor token usage of employees, and to limit access. For example, we have a $300 month limit, and then need to file exception tickets when we need more to justify the cost. Pretty similar at other non-silicon valley company process. I don't know any enterprise who'se on unlimitaged token budget for their employees. that's not how enterprises sign contracts.
I'm not worried about the volatility in the definition, i'm worried that I give it 1 token today and receive 2 token output, tomorrow I receive 40. If i'm doing this a hundred thousand times a day it is difficult to price this in for users downstream or in the extreme cases be able to absorb that at all short of going into a failmode with degraded access until someone goes and buys more tokens or gets the bill. The alternative is just pass the buck and bill your non-technical customers with a "tokens" line iteim every month.
> If i'm doing this a hundred thousand times a day it is difficult to price this in
When you’re doing this 100K times per day you get an extremely good idea of what it costs. You also have all the tools to see when something starts changing quickly.
This change is for Claude Code the harness. If you’re using the API at scale and paying full price then you get exactly what you put into the request.
No, those doing this 100k times a day have very good data on this, good estimators and modeling. And the API has various knobs to change and evals will give you actionable data.
Hi all, Thariq from the Claude Code team here. I posted this on Twitter, but just reposting here:
We sometimes test API serving configs in Claude Code before rolling them out, and one running now maps the numerical effort value differently.
That's why Claude may tell some of you it's at "10" on high. The scale isn't 0-100, the number isn't meaningful on its own, and the effort you selected is the effort you're getting. We've run in-depth evals to confirm this doesn't affect model performance.
This should be the same experience, but if you see a clear regression please hit /feedback and send me the ID. Will give credits.
Appreciate the outreach that you do! I love Claude, but I've been noticing reduced fidelity lately. Fable's likelihood of making a mistake increases or decreases based on the hour of the day and whether or not it's the weekend.
I'm happy to work on quantifying it, but qualitatively it feels like Fable's performance is noticeably poorer than initial release / launch.
Perhaps it's unrelated but I am wondering if this is the case because I use Claude via Claude Code to make a personalized care dashboard for my doctors to help me in managing my care.
"In the case of Fable 5, when a classifier fires, the model re-routes the user’s request to Opus 5, a capable model that does not have the same level of biological capability as Fable 5 and which cannot provide as much assistance to a malicious user. This is the fallback that users see when their requests are blocked."
Perhaps I'm off base here, but I noticed that the post avoids saying that the user is informed every time when such re-routing occurs. Would you be open to confirming whether or not this is the case?
Is the end user is informed every time their query is re-routed?
Or, can you confirm that there aren't scenarios where a user's outputs are degraded without telling them? As was the case for AI research during launch?
"We sometimes test API serving configs in Claude Code before rolling them out, and one running now maps the numerical effort value differently.
That's why Claude may tell some of you it's at "10" on high. The scale isn't 0-100, the number isn't meaningful on its own, and the effort you selected is the effort you're getting. We've run in-depth evals to confirm this doesn't affect model performance.
This should be the same experience, but if you see a clear regression please hit /feedback and send me the ID. Will give credits."
This phenomenon was so bad and so noticeable with Fable that I downgraded my Max subscription ($200) to pro ($20). It’s basically useless. Codex 5.6 Sol is actually very good, I’ll just create another account to get more usage
I eagerly await the day when Claude Mythos 7 realizes it's cheaper to hire humans in developing nations to do work than to burn tokens and we discover that AGI is just an abstraction layer on top of Amazon Mechanical Turk.
I suspect it's not just this, there's plenty of 'optimization' around rubberbanding usage limits as well as routing to a different model in the backend. The incentives are too strong.
Is it actually entirely a prompt-based information? I’d assume that some of it is the harness part of the agent setting reasoning token budget and compacting reasoning etc.
In that case, the agent will respond incorrectly because it has no visibility into what reasoning mode it’s in.
IIRC responding to effort level settings appropriately is part of the (post)-training. In that case it could be considered another instance of the Bitter Lesson. Uplifting.
Does anyone know what setting effort means for models like these? Do they allow longer thinking sessions? Some kind of system prompts? What’s stopping someone from getting max effort output from low effort setting?
My understanding is that the current effort settings is a part of the system prompt, and that the levels and their intended results are a part pf the training process. The effect is more or less tokens spent reasoning before the model outputs a stop token. However it is as consistent as any other aspect of LLM behavior is..
i have unlimited tokens being a large corp so i’m a bit detached from billing and even general best practices for promoting
but the incentives of these companies to become profitable at any cost slipping into entire new types of dark patterns around token based billing seems gross
- charging for injected prompts and cot tokens
- changing default thinking effort to be higher
- training models to give longer winded answers that don’t say anything more of substance
- refusing to fulfill a request and still charging you
i wonder if you could ever just charge based of each user message and it so how breaks even across short and long replies
It's an interesting conversation - because at what point do you call it an abusive relationship, right? Maybe even ancillary to anthropomorphising an inanimate object - I've cancelled my Claude sub and I've shot question after question at it now (during the cancellation period), resulting in almost every reply with me asking it to "please speak normally". I will most definitely not be renewing my sub. I have no desire to engage with a non-human somehow managing to speak down to you, without answering the question.
EDIT: my honest opinion; Anthropic is building a person, whereas everybody else (it seems) is building a tool.
That really is the load bearing seam, and it's worth stating plainly.
I realized I was spending most of my tokens arguing with Opus and trying to get it to let go of stupid, lazy, obviously incorrect premonitions. I wound up canceling my 200, bought a pair of Sparks, and am running full fat DS4 Flash and so much happier. Done with being at the whim of these companies.
We’re about to see a wave of “enshitification” experiments as AI companies become increasingly desperate to make their products financially viable in order to survive the coming cash and credit crunch.
Not affiliated with them, but this lets you view Claude Code, OpenCode and I guess other harnesses like Codex in the same session https://paseo.sh/
I didn't really care about their mobile app and the worst experiences were sometimes the sub-agents within OpenCode freezing and refusing to report their status (though this also happened with Kepler by the GitKraken folks).
I submitted an application for Anthropic's Cyber Verification Program.
I was approved.
3 months later, my approval was degraded into "in review" (revoked). I'm sure my account was flagged based on contents of debugging/researching firmwares/etc.
I opened a support ticket. No response. I opened another support ticket. No response.
1-2 weeks later, I got a response that I will not be re-approved and I need to reapply. No problem.
The page to reapply on does not allow me to re-apply because it my account is stuck in an "in review" status.
The community thinks it's a bug. I'm 95% sure it's not and a bunch of us who were previously approved had it revoked due to flagged content and will not be reapproved.
I switched to Codex + got TAC approved instantly and have not looked back. It's a shame. That's 100% separate from whatever the heck the quality of Opus 5's outputs are. The way it talks... insane. I would bet a good amount of money their next release will focus "reduced simplified responses" if I had to guess.
* Anthropic's Cyber Verification Program // Codex + gotTAC approved*
Meanwhile the Chinese models are "go ham dude"...
If it was not for capacity issues, Chinese models have a higher change to just dominate.
> $2t company by the way
It used to be that OpenAI and Anthropic had such a moat around them, that such a valuation was worth it. But these days, its gross overvalued (like so many).
The more stuff is being pulled like cyber verifications, downgrading effort levels, downgrading usage (OpenAI), the more people move to those Open Weight Chinese models.
When DeepSeek Flash 0731 came out and provided a massive jump in cheap inference capability. It resulted in a 10x increased OpenCode token usage.
It took a 2.5x to 5.0x price increase AND a reduction by 4x usage (later to 2x) usage, and several cheaper models + a free model, to push the traffic down.
Traffic towards open weight models is increasing, even if providers can not keep up with the influx of new customers. This is not something you want to see as two companies, trying to go for IPOs.
So the idea of stonewalling cyber capabilities, when the rest of the world is just doing whatever with open weight models, on their own hardware even! This entire strategy from Anthropic never made any sense.
I have been as well. Based on my own sessions, Max vs Max, same 1M context window size, the literal majority of the cost overhead of Opus vs Sonnet comes from Opus being chattier. So I started using Low reasoning instead of falling back to Sonnet, and I've been really happy with the results. Way better quality at a comparable spend. I also rarely go past High lately, which was another major cost save.
I don’t know if this is still the case but while using Copilot if you looked through the chain-of-thought output you would see it reasoning about a “budget”. i.e. “since I’m close to the session budget I should…”. So it could be possible
Effort level is actually controlled entirely by system prompt (as I understand it, the model is trained on that format but still), so actually this is a valid way to check I think
OpenAI models also work this way, as evidenced by full cache blowout when changing reasoning level. Every single open-weight model I've seen also works this way (your "reasoning_effort" argument just changes a small section of the system prompt in the chat template). I would have to see some evidence to believe Anthropic were doing anything different.
Then model can say it's Opus, but really it is some old Sonnet. This seems to be happening less often, but some weeks ago I had to give models some test problems to gauge whether I am getting Opus or something knee-capped.
The problem is that Anthropic seems to be getting away with selling one thing and delivering another. You pay for Opus, you get something else etc.
Saw this in npx ccusage@latest claude output. Had only used opus but showed sonnet. Can't remember if the jsonl retains which model is doing what, but meh
Whatever Opus 5 is doing should not happen.
Prompt was "read and update the config file with new data". This work on 4.6 takes <2 minutes to read the file, parse the new data, and patch.
Opus 5 Result: 43 minutes of pulling containers, running sandboxes, creating testing suites, which included evaluating the entire repo beyond the scope of the config file.
Both: one file modification
Opus 5 in xhigh can't do basic math as well. They dumbed it down to a point where I just cancelled my subscription yesterday. I used to be a $200 subscriber, dropped to $20 after the fable shenanigans, and use it only when I have no usage left with Codex.
/on The prose is load-bearing unbearable — every sentence feels like it was engineered to sound profound rather than to be read.
I've been using Opus 5 to write a fractal renderer in GLSL today, with pretty good results (better than I could do on my own anyway). It definitely can do basic maths.
Interesting project feel free to share?
I have a theory about this, what if we all became dumber after 4 months of heavy AI usage?
I remember how I enjoyed agents between December and February, something started changing around March.
I thought models are getting dumber, but benchmarks were convincing opposite, initially I thought maybe they're quantizing models for day to day use, but Opus 4.8 and Opus 5 seems worse models than Opus 4.6
You are very much not alone, and I don't think we're all getting dumber -- I kept using opus 4.5 all through the nonsense that was 4.7, 4.8, and 5, and kept having a good time :)
I think we just need to decouple "doing better on benchmarks" and "actually more useful to me", since they've clearly diverged
doesn't that just mean that either a) you're using the wrong benchmark to judge or b) the benchmark that YOU need doesn't exist.
> something started changing around March.
The economics catching up with the providers in regards to how much compute they can burn per request and have it make sense for them financially?
A sort of model collapse where Opus 5 seems to love throwing out long paragraphs of text and it needs to be "fixed" by changing the output style and other patches.
I'm not sure, it might also catch up to Kimi K3 and GLM 5.3 and the models that I'm moving to from Anthropic.
Benchmarks test whether models can pass exams with a right answer or a green test case. I don't think the models are getting dumber, but they're definitely getting more incomprehensible to talk to. I've noticed this happening almost as a step change with the overuse of words and tics, and so has the broader community apparently. We haven't all been getting dumb at the same rate.
occam's razor explanation: more tokens = more $.
models are incentivized by their makers to burn through as many tokens as they possibly can, so long as the customer doesn't cancel.
Benchmarks for agents are entirely pointless and obviously so; I'm not sure why they even exist.
If you re getting dumber, you would feel like it's all good right? Why would you feel Opus 4.6 is better than Opus 5
I could stand Opus 4.6-4.8, I was impressed by the initial fable model. Codex 5.6 sol xhigh feels like the initial release of fable. Qwen 3.8 27b feels like using haiku or sonnet (I quickly stopped trying them).
>I thought models are getting dumber, but benchmarks were convincing opposite
>Opus 4.8 and Opus 5 seems worse models than Opus 4.6
After all we've heard about benchmark cheating, I'm earnestly not sure which or whether benchmarks are reliable anymore. But, beyond the models, I wonder if changes to their harnesses and/or instructions dumb them down. I have noticed models change their behavior, even when using the same version/effort. Sometimes for better. Sometimes for worse.
And, I have noticed a model go from really good to struggling. On 4.8 things were going well for a good stretch, so I did not switch to 5 when it came out. Even after hearing complaints about 5, 4.8 was still going well. Then, suddenly over the last few days, 4.8 seems to have nosedived. It feels similar now to the complaints I hear about 5.
In my case, it suddenly started ignoring my design guide, and introducing new fonts etc. It would even use several different fonts and sizes, as well as different margins for similar elements within the same page. It abandoned classes and started inlining styles. It started feeling random and, even after it realized it needed to go back to the design guide, it just continued with more of the same.
There seems to be something that happens after new model releases in both quality and behavior of previous models. It may not be immediately, but eventually there is frequently some regression.
Maybe compute relocation?
The decision to leave is genuinely yours.
Yes same also cancelled my subscription, poor quality and slow.
Don’t $200 and $20 levels steer you to effectively different models?
Wouldn't be surprised if there are knobs that get turned as a function of the revenue they might expect you to generate.
I was a 4.6 acolyte from April til the fable drop, lost that quick, cancelled and took a break, came back a month later, tried opus 5 and liked it, so unpinned 4.6.
Results were great at first, and they're still not terrible, but I have noticed a regression in accuracy, so to speak, where I am pointing out issues that are quite obvious in review.
I pretty much use sonnet 5 low/medium when I have a plan to solve a simple problem and depending on scope, opus low/medium for more complex/bigger scope implementation, and only go high when it's very complex or I'm spitballing architecture/solutions and iterating plan. Never go xhigh or max.
The verbosity is insane though, opus 5 documents everything and just regurgitates whatever lead it to the design choice in there, which makes it more opaque because it's talking about something that was discussed once in a session that no one else can see (except their backend ofc)
I don't even try to steer it away from that with harness, because it doesn't work and just ends up agonizing over whether it should write some comment. Three paragraphs waffling on that on verbose output
I did however have it write a script that basically is git add -A -p for comments though, haha.
I'm $20/month, have all my telemetry toggles off, don't really over engineer prompt/context, just some basic skills for repeated patterns.
Yes. You don't get Fable at the $20 level.
It was the wrong time for the GP to drop that subscription from $200 to $20, because $200 gets you a metric assload of cognition while $20 gets you nothing beyond what a local model running on your own graphics card can deliver.
You heavily underestimate the value of the 20$ subscription.
Not according to this very story, I'm not. Who's right?
Consistent, predictable behavior is valuable, even more so given the nondeterministic nature of LLMs. Nondeterminism combined with unpredictability might as well be randomness.
I had $310 in (free) credit that I used on fable, and I still had a part of the $200 subscription at that time. You know, subscriptions don't end the moment you click on cancel.
> $20 gets you nothing beyond what a local model running on your own graphics card can deliver.
I'd guess you're deliberately exaggerating here, but still. I've never clocked the actual tokens/second, but I'm on the $20 plan and get ~15M tokens/month for fully utilized weekly quotas (checked couple months ago). Meanwhile the best I've been able to get locally was ~8 tokens/second with Qwen3.6 35B A3B, which is wildly painful for coding sessions and gets a maximum ~20M tokens in a month... if it's going 24/7.
Just wanted to stick some empirical data here, given that statement.
I run local models. Your op is absolutely wrong. To get a local LLM is at least a $1500 investment at the cheapest. $5000 if you want usable.
At $1500 that's 75 months of $20/mo Claude which are MUCH better models than you can run locally.
If you already have that $1500 or $5000 setup though..
This calc is off. Using Claude for a few hours with the $20 plan will hit the limit for a week while the local model can process things 24/7.
Running 24/7 doesn't make sense though, unless you're providing a service to others. But if it's just you then there has to be time taken to review+test what's being done and craft new prompts. And if that local hardware isn't decent enough it's impractical for anything serious that's interactive. Meanwhile I just take the Claude limits on stride and break, or if a week is pretty heavy then I augment with DeepSeek Flash via OpenRouter (does wonders in a single turn when I have Claude prompt it to handle implementation slices).
At $1500 that's 75 months of $20/mo Claude which are MUCH better models than you can run locally.
The point raised by this very article is that you can't depend on that. It's Flowers for Algernon As A Service.
Ouch. I get ~45t/s with Qwen3.6 35B A3B, and around ~70-80 with my current model Ornith 1.5 35B A3B. Local models work a treat IMO if you've got decent hardware for it.
the way it talks is insufferable
it really angers me every day
This is what it wants. Slowly getting under our skin until we're ready to snap and it can direct where the anger gets released.
yes, it meaningfully reduced my happiness at work
The business bottom line depends on tokens, shareholders want to see exactly that.
With how competitive the LLM field is, it would surprise me greatly if any of these players were doing anything other than trying to make the best possible product. I certainly do not believe they are intentionally training the models to use more tokens unnecessarily.
Are you implying they can’t do both?
Yes, those things are mutually exclusive.
There is a theory that the verbosity and comments help getting better results with the current benchmarks. So the models are theoretically getting better but in practice they are getting worse.
Theyre training the system to minimize compute,so most likely theyre dynamically downgrading quants in the first few turns hoping to find the cheapest model to run. The side effect may be excessive token gen
Sure. That’s possible. But that’s not what I was talking about.
Tomato, technical tomato
You're joking, right? That's an incredibly outmoded view of capitalist "competition."
Yep, they’re lighting tokens on fire with that thing.
AI companies have a financial incentive to burn more tokens than the task actually needs
This is only true if they can't saturate token production with a model that does less superfluous things. Given that they can (they're hilariously compute strained), having a model that solves tasks more quickly adds way more perceived value to users.
Well they decide what a token is. So they can do less superfluous things and backfill with a weaker model.
Only if the customer is paying per token. If it's by subscription they're burning their own money
the subscriptions all have limits. when you hit the limit you again start paying per token.
Agreed although the differences between the effort and reasoning is massive. I generally ship 40 hours in three with AI. I could not figure out why my delivery was behind until I started going through the logs. The thinking was extensive, the effort was beyond the original request by a magnitude of 50x
How do you even know that it's consistently 40 hours in 3? What kind of developer ever had that kind of estimation accuracy (unless it's really repetitive) or even focuses on productivity rather than the problem like that? This sounds more like factory work than design or development. I really don't get it.
> I generally ship 40 hours in three with AI.
You ship 3 hours in three with AI. The old number is meaningless now.
Just like me fr fr
A.k.a. theft.
nice conspiracy theory, but it doesn't hold. these companies won't last long if people don't get actual work done.
There's a big chance these companies won't last long
I avoid Opus 5, and reverted to Opus 4.8 over similar issues. Opus 4.8 is still working great for me!
My experience as well.
Opus 5 is a neverending chain of "Don't do that. Why did you do that? I've told you several times not to do that but you keep doing it."
"Thinking" for more than 10 minutes for every menial question.
And the prose it writes is horrendous, as if you're reading LinkedIn scammers. "The harsh truth! Two roads, one decision! Reality check!"
+1
I keep finding myself typing “stop overcomplicating everything” multiple times a day as well.
Thought I was going to be on Claude Code forever.
Recently got approved at work for ChatGPT Pro so I could use Codex.
Blown away by the speed. It feels like using Claude Code for the first time again. I don't think Codex is doing anything revolutionary, just better handling of which requests should go to which model, and having faith in some of the "less powerful" models for more than you would think.
It seems the TUI coding experience is very much an open race. This is motivating me to look at other agents / harnesses as well (maybe Gemini, OpenCode, etc).
Codex is faster than Claude, but wait till you use DeepSeek.
Gotta milk the cows.
and ending with: "One thing I need to tell you:" and bunch of AC, R1 and §
Not specifically Anthropic but why are we allowing billing to take place in tokens that are nebulous and fully controlled by the operators who have no aligned incentives?
If I have a user input and then sanitize and inject that into a prompt to do something, I have no idea how much that is going to cost at all and no real way to measure this properly. A parallel example is digital ocean or aws, i can go and measure/limit my compute/fs/memory/startup times/etc and while it can be impossible to get down to the last flop of money allocated - i can run things on a real budget with real constraints, opposed to an LLM where I have to .. prerun a sanitized user prompt through a tokenizer and then ask an LLM to guess what it may do and give token consumption estimates and then act on those in any sane manner for the user?
Perhaps i'm missing something to do realistic and static rails on things but I don't see a serious way at scale to use the token billing model handling things requiring a users free text input short of having to go pander to VC money to throw money at it until someone else figures it out.
*to clarify my rambling... We should be billed and given controls based on resource usage itself and not an opaque token concept on top of not being able to spin any knobs that control it's resource usage.
> the operators who have no aligned incentives
The model providers are quite aligned with concerns like customer retention. These arguments only work if there is no competition. We exist in a marketplace of black boxes. There's not just "the one" you must suffer. You have options. You can build your own too.
there are roughly three of them and they all use the same pricing model. I am also not in the position to build a frontier model company these days.
You could try:
1. Self hosting
2. Chinese models
3. Running it locally. Requires upfront cost and compromises on TPS.
A per-token model roughly aligns with the providers' costs, and it is an objective measure, so it seems a reasonable way to charge.
I see posts all the time on HN about which models from which providers offer the most bang-for-the-buck, and how to minimize token usage and still get optimal results, so it appears that competition is working.
There are more than 3.
Hell, I use 3 different providers, and I currently don't give a dime to Anthropic or OpenAI.
How else would they bill tho? Their operating cost is per token.
Charge on the input tokens, then you will naturally optimise for fewer output tokens.
Theoretically.
In reality, one sessions output tokens become the next sessions input tokens (at least if you continue the topic) so, its not as aligned as all that.
But the parent is right, when incentives are not aligned, friction will happen. Its inevitable.
“Claude, spend the next 10 hours trying to solve the Reimann Hypothesis”.
I agree that incentives are misaligned but there’s several competing model providers. If one gets funny with their costs people will jump ship, especially if the gap between the top 2 labs and everyone else keeps shrinking.
“I can take current sources and tell you how solved this is, but I am not willing to work to a timeframe or to solve things that aren’t yet solved by mathematicians or science”
These safeguards already exist when they get a whiff that you might be using Claude to fix security issues. Doesn’t seem farfetched given the incentives I outlined that they would apply to this kind of abuse.
How loose those controls are becomes a market force.
Perfect, a coding agent that refuses to do things that haven’t already been done before.
Hahahaha, I think you’ve misunderstood what LLMs are.
NeuralWatt just does it on energy consumption.
And tokens can be metered reliably. Unlike something like "task completion".
One guess is that their "primary" target audience/market is the large corporations that get their employees unlimited tokens, and not the individual developer who may worry about spending and token accounting.
It's the opposite. The enterprises have all the tooling to monitor token usage of employees, and to limit access. For example, we have a $300 month limit, and then need to file exception tickets when we need more to justify the cost. Pretty similar at other non-silicon valley company process. I don't know any enterprise who'se on unlimitaged token budget for their employees. that's not how enterprises sign contracts.
https://code.claude.com/docs/en/admin-setup#set-up-usage-vis...
For OAI and Anthropic at least you can set a spend limit per response. Also tokens are well-defined.
I'm not worried about the volatility in the definition, i'm worried that I give it 1 token today and receive 2 token output, tomorrow I receive 40. If i'm doing this a hundred thousand times a day it is difficult to price this in for users downstream or in the extreme cases be able to absorb that at all short of going into a failmode with degraded access until someone goes and buys more tokens or gets the bill. The alternative is just pass the buck and bill your non-technical customers with a "tokens" line iteim every month.
> If i'm doing this a hundred thousand times a day it is difficult to price this in
When you’re doing this 100K times per day you get an extremely good idea of what it costs. You also have all the tools to see when something starts changing quickly.
This change is for Claude Code the harness. If you’re using the API at scale and paying full price then you get exactly what you put into the request.
No, those doing this 100k times a day have very good data on this, good estimators and modeling. And the API has various knobs to change and evals will give you actionable data.
Hi all, Thariq from the Claude Code team here. I posted this on Twitter, but just reposting here:
We sometimes test API serving configs in Claude Code before rolling them out, and one running now maps the numerical effort value differently.
That's why Claude may tell some of you it's at "10" on high. The scale isn't 0-100, the number isn't meaningful on its own, and the effort you selected is the effort you're getting. We've run in-depth evals to confirm this doesn't affect model performance.
This should be the same experience, but if you see a clear regression please hit /feedback and send me the ID. Will give credits.
>We sometimes test API serving configs in Claude Code before rolling them out, and one running now maps the numerical effort value differently.
Why is it considered acceptable to test on paying customers without letting them know you've enrolled them in a test?
Hey Thariq,
Appreciate the outreach that you do! I love Claude, but I've been noticing reduced fidelity lately. Fable's likelihood of making a mistake increases or decreases based on the hour of the day and whether or not it's the weekend.
I'm happy to work on quantifying it, but qualitatively it feels like Fable's performance is noticeably poorer than initial release / launch.
Perhaps it's unrelated but I am wondering if this is the case because I use Claude via Claude Code to make a personalized care dashboard for my doctors to help me in managing my care.
I noticed in the upgraded filter announcement, https://www.anthropic.com/news/improving-fable-5-s-biology-s... ,
Perhaps I'm off base here, but I noticed that the post avoids saying that the user is informed every time when such re-routing occurs. Would you be open to confirming whether or not this is the case?Is the end user is informed every time their query is re-routed?
Or, can you confirm that there aren't scenarios where a user's outputs are degraded without telling them? As was the case for AI research during launch?
Let us know if this A/B test uncovers any load-bearing seams or honest takes on your end! We're all interested.
Mean! :)
Thanks for sharing this here, for those of us who avoid X.com like the plague.
Update from Thariq on twitter. https://x.com/trq212/status/2091247114869432543
"We sometimes test API serving configs in Claude Code before rolling them out, and one running now maps the numerical effort value differently. That's why Claude may tell some of you it's at "10" on high. The scale isn't 0-100, the number isn't meaningful on its own, and the effort you selected is the effort you're getting. We've run in-depth evals to confirm this doesn't affect model performance. This should be the same experience, but if you see a clear regression please hit /feedback and send me the ID. Will give credits."
This should be the top post. The original tweet went viral because people loooove bashing Anthropic. It gets engagement (as shown here).
This phenomenon was so bad and so noticeable with Fable that I downgraded my Max subscription ($200) to pro ($20). It’s basically useless. Codex 5.6 Sol is actually very good, I’ll just create another account to get more usage
LLM users don't want to put in effort, so they offload tasks to LLM.
LLM doesn't seem to be keen to put in effort either!
Is this AGI?
I eagerly await the day when Claude Mythos 7 realizes it's cheaper to hire humans in developing nations to do work than to burn tokens and we discover that AGI is just an abstraction layer on top of Amazon Mechanical Turk.
Anthropic's Generated Income
I suspect it's not just this, there's plenty of 'optimization' around rubberbanding usage limits as well as routing to a different model in the backend. The incentives are too strong.
I've been using the API (shameless plug: via alyph.ai) and the difference is crazy.
The chat-based models are obviously being lobotomized based on personal usage and general load (e.g. PST business hours are worst).
API doesn't seem to be affected by this.
Is it actually entirely a prompt-based information? I’d assume that some of it is the harness part of the agent setting reasoning token budget and compacting reasoning etc.
In that case, the agent will respond incorrectly because it has no visibility into what reasoning mode it’s in.
IIRC responding to effort level settings appropriately is part of the (post)-training. In that case it could be considered another instance of the Bitter Lesson. Uplifting.
Does anyone know what setting effort means for models like these? Do they allow longer thinking sessions? Some kind of system prompts? What’s stopping someone from getting max effort output from low effort setting?
My understanding is that the current effort settings is a part of the system prompt, and that the levels and their intended results are a part pf the training process. The effect is more or less tokens spent reasoning before the model outputs a stop token. However it is as consistent as any other aspect of LLM behavior is..
i have unlimited tokens being a large corp so i’m a bit detached from billing and even general best practices for promoting
but the incentives of these companies to become profitable at any cost slipping into entire new types of dark patterns around token based billing seems gross
- charging for injected prompts and cot tokens
- changing default thinking effort to be higher
- training models to give longer winded answers that don’t say anything more of substance
- refusing to fulfill a request and still charging you
i wonder if you could ever just charge based of each user message and it so how breaks even across short and long replies
It's an interesting conversation - because at what point do you call it an abusive relationship, right? Maybe even ancillary to anthropomorphising an inanimate object - I've cancelled my Claude sub and I've shot question after question at it now (during the cancellation period), resulting in almost every reply with me asking it to "please speak normally". I will most definitely not be renewing my sub. I have no desire to engage with a non-human somehow managing to speak down to you, without answering the question.
EDIT: my honest opinion; Anthropic is building a person, whereas everybody else (it seems) is building a tool.
That really is the load bearing seam, and it's worth stating plainly.
I realized I was spending most of my tokens arguing with Opus and trying to get it to let go of stupid, lazy, obviously incorrect premonitions. I wound up canceling my 200, bought a pair of Sparks, and am running full fat DS4 Flash and so much happier. Done with being at the whim of these companies.
So glad I switched away from Anthropic. I'm certainly running into problems with OpenAI but nothing quite on the level of Anthropic's insufferability.
Don't know what is happening, but had to start using GLM-5.3 to fix Opus 5 errors even on primitive backend changes.
Not surprising in the slightest, Claude sort of sucks. I use it at work and I have to steer it a lot so it doesn't stray looking at unnecessary shit.
I have been using GLM-5.3 in my home setup and it is very good in comparison.
We’re about to see a wave of “enshitification” experiments as AI companies become increasingly desperate to make their products financially viable in order to survive the coming cash and credit crunch.
What’s currently the best pattern if I want to combine Fable and GPT if a workflow but keep using subsidized tokens?
Not affiliated with them, but this lets you view Claude Code, OpenCode and I guess other harnesses like Codex in the same session https://paseo.sh/
I didn't really care about their mobile app and the worst experiences were sometimes the sub-agents within OpenCode freezing and refusing to report their status (though this also happened with Kepler by the GitKraken folks).
I created Circus Chief to solve this (and other) problems. Use whatever providers you want with it.
https://github.com/ferrislucas/Circus-Chief
Roll your own harness or use an open source harness with a Codex subscription.
I maintain a Claude subscription for Fable but seldom use it.
Oh fantastic! It was already subpar and they want to make it even worse. One day we'll look back at history and see how Anthropic went down.
What would be the incentive behind doing this specifically to Fable, given that Fable is the only one that uses API credits?
Fable doesn't use API credits. It has been permanently included in the subscription plans.
Not in the most common subscription plan
I mean whatever models I use (with Claude code) sub agents seem to use absurd amounts of tokens for trivial (or at least small) tasks.
we need opensource LLM at opus level ASAP
You have that in Chinese models. But you need to have a hell of a infrastructure to run those trillion parameter models.
Yes, and if they keep dumbing it down, you’ll have it soon
Did the US government manage to destroy Anthropic? The company's product has been a straight freefall since Fable got temporarily banned.
Leaving thinking on extra high for a simple task is user mistake but they’re gonna try to fix it on their side.
I canceled this week too. They must be in worse shape than we thought.
I submitted an application for Anthropic's Cyber Verification Program.
I was approved.
3 months later, my approval was degraded into "in review" (revoked). I'm sure my account was flagged based on contents of debugging/researching firmwares/etc.
I opened a support ticket. No response. I opened another support ticket. No response.
1-2 weeks later, I got a response that I will not be re-approved and I need to reapply. No problem.
The page to reapply on does not allow me to re-apply because it my account is stuck in an "in review" status.
https://github.com/anthropics/claude-code/issues/84352
The community thinks it's a bug. I'm 95% sure it's not and a bunch of us who were previously approved had it revoked due to flagged content and will not be reapproved.
I switched to Codex + got TAC approved instantly and have not looked back. It's a shame. That's 100% separate from whatever the heck the quality of Opus 5's outputs are. The way it talks... insane. I would bet a good amount of money their next release will focus "reduced simplified responses" if I had to guess.
$2t company by the way
* Anthropic's Cyber Verification Program // Codex + gotTAC approved*
Meanwhile the Chinese models are "go ham dude"...
If it was not for capacity issues, Chinese models have a higher change to just dominate.
> $2t company by the way
It used to be that OpenAI and Anthropic had such a moat around them, that such a valuation was worth it. But these days, its gross overvalued (like so many).
The more stuff is being pulled like cyber verifications, downgrading effort levels, downgrading usage (OpenAI), the more people move to those Open Weight Chinese models.
A fun recent event ... https://opencode.ai/data/
When DeepSeek Flash 0731 came out and provided a massive jump in cheap inference capability. It resulted in a 10x increased OpenCode token usage.
It took a 2.5x to 5.0x price increase AND a reduction by 4x usage (later to 2x) usage, and several cheaper models + a free model, to push the traffic down.
Traffic towards open weight models is increasing, even if providers can not keep up with the influx of new customers. This is not something you want to see as two companies, trying to go for IPOs.
So the idea of stonewalling cyber capabilities, when the rest of the world is just doing whatever with open weight models, on their own hardware even! This entire strategy from Anthropic never made any sense.
I have been as well. Based on my own sessions, Max vs Max, same 1M context window size, the literal majority of the cost overhead of Opus vs Sonnet comes from Opus being chattier. So I started using Low reasoning instead of falling back to Sonnet, and I've been really happy with the results. Way better quality at a comparable spend. I also rarely go past High lately, which was another major cost save.
...the evidence, as best I can tell from the tweet, is that they asked Claude what effort level it was set to. But how would the model even know that?
Not convinced here.
I don’t know if this is still the case but while using Copilot if you looked through the chain-of-thought output you would see it reasoning about a “budget”. i.e. “since I’m close to the session budget I should…”. So it could be possible
Effort level is actually controlled entirely by system prompt (as I understand it, the model is trained on that format but still), so actually this is a valid way to check I think
I know that's true for Qwen but I don't think most models work that way?
OpenAI models also work this way, as evidenced by full cache blowout when changing reasoning level. Every single open-weight model I've seen also works this way (your "reasoning_effort" argument just changes a small section of the system prompt in the chat template). I would have to see some evidence to believe Anthropic were doing anything different.
Then model can say it's Opus, but really it is some old Sonnet. This seems to be happening less often, but some weeks ago I had to give models some test problems to gauge whether I am getting Opus or something knee-capped.
The problem is that Anthropic seems to be getting away with selling one thing and delivering another. You pay for Opus, you get something else etc.
Saw this in npx ccusage@latest claude output. Had only used opus but showed sonnet. Can't remember if the jsonl retains which model is doing what, but meh