Posted by lairv 22 hours ago
Update: fixed after updating to latest version of the app
Eval context: 30 bug fixes and features evaluated by tests the agent didn't see. Ran on Opus 5.5 at medium effort in Claude Code.
20 of the tasks were last run on Nov 1, twice each. Nov 1 solved 36/40, and today solved 19/20. Thinking tokens per task were 0.95x Nov 1’s and output tokens 0.97x. The one significant change was speed: median time to first token decreased from 3.0s to 1.7s.
Caveats: one run per task today, medium effort only.
Paragraph 1 says there were 30 bug fixes.
Paragraph 2 first says they tested it on 20 of the “tasks”. But why is the denominator 40? Where did that come from? It should be 30.
It goes on to mention another test, this time with 20 tasks. Again, it’s unclear where this is even coming from. Based on the first paragraph it should be 30 not 20.
There are 30 tasks in my eval total:
- 20 of them ran on Nov 1, twice each (so 40)
- I created 10 more tasks and ran them on Nov 5 (also ran them twice each)
- Today I ran all 30 tasks, once each, on Opus 5.5 to see if the performance decreased at all
- I only put the comparison against the original 20 in my post, since I saw in another post that people are claiming Opus was nerfed on November 3
This will all get less confusing when I make my benchmark public (hopefully this week).
Also - LLM text? If GP is confusing it's because it's overly succinct and might have typos, which are not exactly the hallmarks of LLM text.
1. 30 was a typo, it should have been 20.
2. 30 was the full suite, but they only ran 20 of those 30 on Nov 1st and so ran the same corpus again for a fair comparison.
Why paragraph 1 even mentions the full set of 30 tasks if 10 of them are irrelevant to the comparison, I don't know.
I'm not sure if it's LLM text, or somebody who has been reading way too much LLM text and is now corrupted into mimicking it. My guess would be a mix of human and LLM. "...evaluated by tests the agent didn't see" sounds very Claude. The first line (2 sentences) sounds human to me. Everything after sounds AI, though the last line is ambiguous.
It's kind of sad how our meat brains are getting non-consensually trained on this categorization task....
https://news.ycombinator.com/item?id=50422951
So only if you were "selected" for their experiments, your account was nerfed.
I’m working on a task. And the task is never to evaluate the agent performance.
If the nerfing is lower effort (seems to be, presumably why they hide the CoT) then they may be wondering if they can get away permanently with lowering effort on the new release provided it gives a wee improvement.
Sometimes you just have insufficient evidence to be sure of something either way, and nobody gets to win the argument or shut somebody else down.
In this case, some have reported analytical work[0] that goes beyond "gut feel" anecodotal evidence, but even that's not going to conclusive and may not be trustworthy at all.
[0] https://www.reddit.com/r/ClaudeCode/comments/1xr9x5y/the_ner...
Ok, it probably burns tokens faster, but is it really a nerf?
so not everyone will have the same experience. If I was saving on costs that would be the easiest way to do it without deploying, without any nefarious intent. The persons in charge of that likely don't even know about how individual users are impacted
For example a provider quietly underclocks your cpu or puts it is an energy saving plan hoping you won't notice.
Or switch to a faster, cheaper cache mechanism or something that's almost as good, but much cheaper to run, in the name of improving efficiency and keeping up with demand. This switch to cheaper mechanisms is what it feels like to me.
https://www.reddit.com/r/ClaudeCode/s/64DXRERLyU
It would be nice to know how they select. I max out my $200 subscription every week and I was selected. On Saturday, the Opus on my account that happened to be supervising a running live process, became unusable.
I really really really do not trust companies on the verge of IPOing from goosing their numbers. Even if the lawyers are up their ass, the social pressures are more than enough to compel the behavior. For every good person willing to do the right thing even at the cost of a delayed career progression, there are 20 people willing to do anything not even for a dollar but for an "atta boy!" from the boss.
The most plausible approach would be for Anthropic to train some separate classifiers to detect benchmarks. I'm not saying it's completely impossible but IMO this is conspiracy theory levels of effort.
You're absolutely right that it is a game of whack-a-mole and eventually we may have leading private + credible + popular benchmarks that ultimately dissuade the behavior, if it can be dissuaded.
They don't detect a quantized model until the quantization gets down to 4 bits or less.
Humans absolutely do detect quantization well before 4 bits...
Try it yourself if you want. Try a small local model, both quantized and not, and I would be very surprised if it isn't immediately obvious which is which.
But no, evals and benchmarks do not detect quantization, until around 4 bit quantization.
https://developers.redhat.com/articles/2024/10/17/we-ran-ove...
https://pub.towardsai.net/your-model-passed-safety-evals-the...
And the second article seems to reinforce the idea. Yes, broad perplexity tests are not a good proxy for the tests used during training.
Quantizing anything is effectively changing the weights, even if you don’t quantize the weights themselves.
It probably need a whistleblower to uncover what they are doing.
I doubt this is the case. I suspect all variation in quality and output can be explained in terms of prompt differences, and changing how much thinking happens at each level, as we see in the post above.
And what about A/B testing? Anthropic doesn't even try to hide that they do this. Limited compute and A/B testing alone, would guarantee a nerfed experience some of the time. That is, you don't even need a nefarious reason.
I consider it gas lighting to deny model nerfing.
They do. I often get the message "This model is currently experiencing high demand, please try again later"
They do drop requests though? At peak times I sometimes have 1-2 requests fail with Opus; no biggie, the harness just retries it. But that, in aggregate, sheds load.
Thats not what Openai and Anthropic do though
If you want to go accuse someone of denying model nerfing, go find someone doing it and argue with them.
Edit: maybe not impossible. Side channels perhaps.
Oh, this Opus X.Y is incredible. Hmm... Now it's dumb. Rinse, repeat.
Doesn't really matter if the cause is quantization, system prompt changes, changes to thinking etc.
It's hard to see how continuously rug pulling users at every release is a good idea. Only reason I can fathom is that it's needed to keep costs inline.
Either that or it's some weird tweaking to the prompts to fix whatever. But given the prevalent sentiment that it's worse that sounds like a long shot
but realistically I think you're right, it's prompt quality and the inherent non-determinism elements of LLMs showing up; when both of those things meet you get brain dead outputs.
Everyone likes to shout "slot machines! non-determinism!" until it actually explains something, heh.
This is getting ridiculous. The number of times I have switched from one model back to the other is crazy. I do not have any proof, so its quite possible I'm just delusional. AI psychosis? Either way, it FEELS like these models work really well for a while and then fall off. Switching to the latest model on the competitor always seems better, until it too seems to taper off. By then, a new model is usually available at the competitor, so I switch.
Its possible there's no degradation at all and just a PERCEPTION of it. But to me the end user, it really doesn't matter.
People seem to underestimate how high normal variance is with these tools.
Max, streamed (128K cap): 128,000 tokens, $2.56, empty file, in September and today.
Max, Batch API (300K): 255,957 tokens for $2.56 in September, 243,707 for $2.44 today. The two drawings: https://playcode.io/marketing/assets/static/claude-opus-5-5-... and https://playcode.io/marketing/assets/static/claude-opus-5-5-...
Xhigh: 86,096 tokens for $1.72 in September, 107,333 for $2.15 today.
The Batch API bills half price. Max ran out of the 128K ceiling again, to the token. Given 300K, it drew the same kind of laptop with 5% fewer tokens. Xhigh used 25% more and drew the same kind of laptop. That's inside the noise: 17 identical calls at high ranged from 17,757 to 59,705 tokens.
The model is still the best of everything we've tried. Most likely the next best would be Fable 5.2. But we don't like Fable because essentially the result is the same, just the price is higher. Its medium and high are close to what Opus 5.5 Max had, but the price is higher.
(I run Playcode, an AI app builder. We benchmark every model before we add it.)
If this turns out to be accurate, I think that would be a pretty unethical thing for Anthropic. Even if it is just a "we had to allocate compute to this new model release".
It should be fraud to swap out the model. We need a term for this, something like a bait-and-switch.
The media coined the term "dual quality" for a European food scandal a few years back, in which conglomerates (including Coca-Cola) adulterated products with cheap, substitute ingredients, if they were destined for distribution in specific markets (mainly, lower-income).
The packaging exteriors were the same. You wouldn't know which version you got unless you were paying close attention to taste. It was essentially what the software industry calls "A/B testing", except with sneaking things into what people eat.
I don't have any good reading links on hand; here's one source,
https://ec.europa.eu/commission/presscorner/detail/en/qanda_... ("Dual Food Quality: Questions and Answers")
I'm a bit disappointed with how this nonsense is spread across HN today.
Are you saying that the many, many people posting and experiencing this are wrong and you are right?
And not just in this instance, nerfing has been complained about in a cyclical fashion for many months.
But, they're all wrong?
You're so confident in your experience that you cannot think that other's experiences are also true?
Like most experiments do not save the company money, there was an experiment to test the new fable, people got to use it at opus-5.5 usage. But this experiment clears the thinking blocks when people go back to the vanilla model; not sure if Anthropic overlooked this, but idk, it does not sound like "nerf on purpose".
There are experiments to adjust the effort level, at the moment the ones that I've seen switch to a higher effort. Again this is not a nerf; might be annoying perhaps, the model does start to behave differently (and consume more tokens). But it's not a nerf.
Then there is apparently one experiment that replaces opus-5.5 system prompt with the one from opus-4.8; the model is the same; this experiment is perhaps a bit more questionable, but still it's not "anthropic nerfed the models".
If people think they are getting nerfed they should check what experiments are enabled (claude.json and session logs).
No, it gets complained about constantly even if there is no model coming out. A completely unrelated thread on Claude will feature comments like "can I just say that [model] has suddenly become useless this week?" Merely asking that question is bound to get people agreeing with you because we are evaluating all of this on vibes and placebos.
I too feel that way sometimes, I don't think it's ridiculous that Anthropic might silently downgrade you as demand warrants. But asking others if they "feel" something I feel is less than useless.
"Could all these people seriously be wrong?" Yes, yes they could.
if they just straight up dont sell you something you bought then its fraud.
I think there is a ripe opportunity for expanding the definition of fraud to include changing the "ingredients" of a named product on the market. It is not limited here, we see it with food, and even in the medical field
(I will die on the hill that medical fraud was committed when mRNA therapy was used as a treatment for covid and it was 'sold' to the public as a "vaccine" despite not being the same thing as any vaccine before it. The definition of vaccine was expanded to include it without sifficient public disclosure to warrant informed consent imo. This is not a statement about conspiracy or the efficacy of vaccines or a moral judgement on whether it was for the greater good or not.. it was simply an act of fraud that robbed the public of proper informed consent by making people believe it was X when it was really Y.)
So yes, I think there is a large need for the public to at the very least put a strong handle on the ability to change what a product name means. It is a trivial problem to solve considering it is only a problem due fraud or fraud adjacent activities.
heres the solution: You have your product name, and then any changes to it REQUIRE a new version name. Like how you have Coke and Diet Coke, only formalized such that Coke cannot legally replace the ingredients of regular Coke with the ingredients of Diet coke and still call it Coke. With Software they could at least use version numbers and release modifiers. most of the time its built in already, but it needs to include the full stack. if they change anything that has an interaction with the user it should change the label seen by the user for the product.
thank you for coming to my ted talk
A Opus 5.5 is not a Opus 5.5. Its Opus 5.5 of some checkpoint.
mRNA btw. was well known how it worked, it wasn't hidden and some people waited for the non mRNA version because of it. I never seen or read the argument you make and I do not get it.
A vaccine is something which protects you from an infection. What definition do you think changed in humans? That we make the protein? How is that so different from viral vector vaccines?
The consent is a certain level of trust from the society to a group of researchers limited by being humans and having to act in a certain time frame.
Nontheless, I do agree I would like to have a fingerprint and proper changelogs etc. for products I buy including digital and analog. But this has a downside as usual: The smaller you are, the more it costs you to (probably) register, maintain and change.
You know the good thing about Claude etc. right now? It doesn't matter. My invest is not 4 digits and there are that many changes, i'm quite happy to have been upgraded for 'free' from Opus 4.8 to Opus 5 (what do you call this upgrade, which was significant better?)
Opus 5.5 was so good that if it wasn't nerfed - at least just looking at the benchmarks - the difference in cost here does not seem to justify using this for anything besides design and review.
That's like maybe 1% of total token usage for SWE BUT a better design can have enormous benefits, so it's almost always worth using the best model for that.
Fable 5.1 Max $11.46
Fable 5.2 Max $44.70
Anthropic does this time and time again, unserious company
System prompt would be the easiest one.
But there were flash or inbetween models, different checkpoints etc. THey might also try different hardware which might not be binary compatible.
Right?
Fuck Anthropic.
https://x.com/follistatindev/status/2119129611510383014
With early handouts of Qwen4 starting to be reported 5 days ago ... so if they're good, they would have had some hours to evaluate Qwen4 against Claude 5.5 and, then or right after was when Anthropic decided to put out 5.2 ...
So yeah ... benchmaxxing and all that. BUT nevertheless, the published results seem to say that Qwen4-Max appears to match Opus 5.5 on some benchmarks, and Qwen4-27B (probably by going even harder than Qwen3.8-27B on thinking/inference time compute) can beat Sonnet 5.5 on some things.
I guess they're not taking any chances. Keep in mind that Anthropic has declared they have non-public models (it's in their IPO filing). These are referred to by the shorthand names Model 2, Model 3, and Model 4. Essentially 4 is better than 3 is better than 2.
Fable 5.2 is "2.5" so to speak. So Anthropic was holding back, but not anymore. And now they can dial the performance up once more with what they have ready, and probably once more after that.
It would be funny though if you are right and that model is Fable 5.5 in a couple weeks.
I suspect we will see people being adamant that this model simply 'feels' better due to long horizon goal alignment improvements ( like this kind that would cause this jump)
I will stick with Sonnet 5.5 - 10 x the amount of work done, at 80% of the quality.
I originally subscribed to $20/mo Pro plan less than a year ago, and upgraded to $100/mo Max plan ~2-3 months ago.
So, the difference between them is just the harness?
I am curious as to why they launched this one and why now.
0: https://x.com/tanayj/status/2118047122939658341
Cloudflare doesn't have token leaderboards or limits, but uses alerts for token spend management:
- $300 in a day: that employee gets an email
- $500: manager gets an email
- $1,000: VP gets an email
(via CTO @dok2001 on No Priors)
As a preliminary verdict after using 25% of the weekly limit: it's a miss.
I'm not sure "we couldn't top Opus 5.5" is the right message for the upcoming IPO. I was expecting a significant release.