Nerfed Daily: 2026-11-09
The complaints are still pointed at one place: Opus 5.5 inside Claude Code, and the weekend made them louder, not sharper. The closest thing to hard evidence is one person having Gemini read 1,400 of their own Claude Code sessions, which says the reasoning changed on November 3 even though the model name didn't, and a follow-up puts numbers on it: on high effort, 40.5% of turns now act with zero thinking, and the median thinking before a tool call fell from 84 to 54. That's one user's logs graded by a rival model, not a benchmark. Most of the rest is same-prompt, launch-day-versus-today videos, which feel convincing and control for nothing. The counter side has its own weak spot: the independent tracker people used to cite stopped posting in mid-October, so the silence cuts both ways. Anthropic has said nothing about model quality. The one official change in the window is a Claude Code release that stops auto mode from quietly falling back to a local classifier, which is about permissions, not intelligence. My read: something in how much Opus 5.5 thinks per turn probably moved, and the bigger damage is to trust, because nobody can check it except by reading their own logs.
#1
@maxt3chno
https://x.com/maxt3chno/status/2119135579673665804
(a) @maxt3chno posted the most shared complaint of the window, a Press F for Opus 5.5 with a side-by-side image. The claim is specific about mechanism: limits drain faster because the model loops on wrong decisions and ignores the notes you gave it, and it says Sunday was worse than Saturday. That ties quality to quota in one sentence, which is why it spread. No logs or numbers, just the image and the feeling.
https://x.com/maxt3chno/status/2119135579673665804
(a) @maxt3chno posted the most shared complaint of the window, a Press F for Opus 5.5 with a side-by-side image. The claim is specific about mechanism: limits drain faster because the model loops on wrong decisions and ignores the notes you gave it, and it says Sunday was worse than Saturday. That ties quality to quota in one sentence, which is why it spread. No logs or numbers, just the image and the feeling.
#2
@ivanainai
https://x.com/ivanainai/status/2119531165874794696
(a) @ivanainai reran the same Machine Claw test from October 6 and posted the two results side by side. The complaint is that Opus 5.5 got slow and started getting simple questions wrong. Reusing a dated prompt is better than most posts manage, but it's still one run on each side, and a single sample of a generative task swings a lot on its own.
https://x.com/ivanainai/status/2119531165874794696
(a) @ivanainai reran the same Machine Claw test from October 6 and posted the two results side by side. The complaint is that Opus 5.5 got slow and started getting simple questions wrong. Reusing a dated prompt is better than most posts manage, but it's still one run on each side, and a single sample of a generative task swings a lot on its own.
#3
@GuoChaoTech
https://x.com/GuoChaoTech/status/2119406392109076715
(a) @GuoChaoTech makes the money argument: the $20 plan was enough in September, now it's $100, and the same prompt produces a worse video than on launch week. The post ends by asking whether Opus 5.5 is being made worse on purpose so Fable 5.5 looks better at launch. A near-identical post from another account ran the same day with the same 5x framing, so this one reads as a template as much as a test.
https://x.com/GuoChaoTech/status/2119406392109076715
(a) @GuoChaoTech makes the money argument: the $20 plan was enough in September, now it's $100, and the same prompt produces a worse video than on launch week. The post ends by asking whether Opus 5.5 is being made worse on purpose so Fable 5.5 looks better at launch. A near-identical post from another account ran the same day with the same 5x framing, so this one reads as a template as much as a test.
#4
@MedAdem_1
https://x.com/MedAdem_1/status/2119542119131983951
(a) @MedAdem_1 used Opus High 5.5 as the daily driver for weeks and says the last couple of days were off. The workaround was moving up to Extra effort or to Fable, and even then it needed many more feedback loops than before. That's a useful detail: when people feel a drop, the fix they reach for is spending more, which is exactly what makes the conspiracy reading attractive.
https://x.com/MedAdem_1/status/2119542119131983951
(a) @MedAdem_1 used Opus High 5.5 as the daily driver for weeks and says the last couple of days were off. The workaround was moving up to Extra effort or to Fable, and even then it needed many more feedback loops than before. That's a useful detail: when people feel a drop, the fix they reach for is spending more, which is exactly what makes the conspiracy reading attractive.
#5
@alexshotgunn
https://x.com/alexshotgunn/status/2119555447195848996
(a) @alexshotgunn built the same game right after launch and again today and calls the difference crazy. Same shape as a dozen other posts this weekend: two artifacts, one prompt, weeks apart. None of these posts share seeds, effort settings, or how many tries each side got, so they can't separate a model change from ordinary variance.
https://x.com/alexshotgunn/status/2119555447195848996
(a) @alexshotgunn built the same game right after launch and again today and calls the difference crazy. Same shape as a dozen other posts this weekend: two artifacts, one prompt, weeks apart. None of these posts share seeds, effort settings, or how many tries each side got, so they can't separate a model change from ordinary variance.
#6
@pablonpedrotti
https://x.com/pablonpedrotti/status/2119534418704363748
(a) @pablonpedrotti asked Claude directly whether it's normal to ask Opus 5.5 several times to get a correct answer, even with all the code and docs supplied. The verdict is blunt: an awful experience, very unpleasant to use. It's a small account, but it's the plainest version of what the high-reach posts are saying.
https://x.com/pablonpedrotti/status/2119534418704363748
(a) @pablonpedrotti asked Claude directly whether it's normal to ask Opus 5.5 several times to get a correct answer, even with all the code and docs supplied. The verdict is blunt: an awful experience, very unpleasant to use. It's a small account, but it's the plainest version of what the high-reach posts are saying.
#7
@madeinksc
https://x.com/madeinksc/status/2119060830969106698
(a) @madeinksc says Opus 5.5 on the Max subscription is unusable again and calls the regression confirmed. Confirmed by what isn't said. The word again matters more than the claim: this account has been through a previous round and the trust is what broke, not just the output.
https://x.com/madeinksc/status/2119060830969106698
(a) @madeinksc says Opus 5.5 on the Max subscription is unusable again and calls the regression confirmed. Confirmed by what isn't said. The word again matters more than the claim: this account has been through a previous round and the trust is what broke, not just the output.
#8
@financialvice
https://x.com/financialvice/status/2119556578068365523
(a) @financialvice quotes the line "you're right, and I was wrong to say" and declares Opus 5.5 lobotomized. It's a joke, but it names a real tell people use: the model folding instantly when pushed back on. Sycophancy is the symptom users notice first, before any benchmark would.
https://x.com/financialvice/status/2119556578068365523
(a) @financialvice quotes the line "you're right, and I was wrong to say" and declares Opus 5.5 lobotomized. It's a joke, but it names a real tell people use: the model folding instantly when pushed back on. Sycophancy is the symptom users notice first, before any benchmark would.
#9
@Kipnis_a
https://x.com/Kipnis_a/status/2119505489927049243
(a) @Kipnis_a ran the quality complaint as the lead of a daily roundup and put the point better than the angry posts did: the dip isn't the complaint, the silence is. Users describe agents failing quietly and small errors creeping into codebases since early November. If you run agents unattended, an unannounced model change looks exactly like a green check, which is the real operational risk here.
https://x.com/Kipnis_a/status/2119505489927049243
(a) @Kipnis_a ran the quality complaint as the lead of a daily roundup and put the point better than the angry posts did: the dip isn't the complaint, the silence is. Users describe agents failing quietly and small errors creeping into codebases since early November. If you run agents unattended, an unannounced model change looks exactly like a green check, which is the real operational risk here.
#10
@siraustin
https://x.com/siraustin/status/2119504683362304087
(b) @siraustin caught a silent downgrade in the act. An ops session got safety-flagged for quoting a security scan back to itself, and Claude Code dropped the session from Opus 5.5 to Opus 4.8 without asking. The status line in the pasted output shows Opus 4.8/max, and switching back had to be done by hand. Even discussing the fallback models triggered the fallback, which is the funniest and most concrete evidence of the weekend.
https://x.com/siraustin/status/2119504683362304087
(b) @siraustin caught a silent downgrade in the act. An ops session got safety-flagged for quoting a security scan back to itself, and Claude Code dropped the session from Opus 5.5 to Opus 4.8 without asking. The status line in the pasted output shows Opus 4.8/max, and switching back had to be done by hand. Even discussing the fallback models triggered the fallback, which is the funniest and most concrete evidence of the weekend.
#11
@TheObiLeonard
https://x.com/TheObiLeonard/status/2119508402774970691
(c) @TheObiLeonard pays for two Claude Max subscriptions and still hits usage limits every two days. No complaint about quality at all, just volume. Read next to the looping complaints, it fits the theory that the cost problem and the quality problem are one problem: a model that retries more burns more.
https://x.com/TheObiLeonard/status/2119508402774970691
(c) @TheObiLeonard pays for two Claude Max subscriptions and still hits usage limits every two days. No complaint about quality at all, just volume. Read next to the looping complaints, it fits the theory that the cost problem and the quality problem are one problem: a model that retries more burns more.
#12
@dev_maims
https://x.com/dev_maims/status/2119403065736306826
(c) @dev_maims calls the $20 Claude plan a scam because you barely start coding before hitting the limit. It's the entry-tier version of the same complaint, and it lands differently: on Pro, there's no room to absorb a model that needs extra rounds.
https://x.com/dev_maims/status/2119403065736306826
(c) @dev_maims calls the $20 Claude plan a scam because you barely start coding before hitting the limit. It's the entry-tier version of the same complaint, and it lands differently: on Pro, there's no room to absorb a model that needs extra rounds.
#13
@swarupdcs
https://x.com/swarupdcs/status/2119070025147441495
(c) @swarupdcs tried Opus 5.5 through Devin Pro at $20 a month and logged it. Day one, about 1 hour 20 minutes of use exhausted the daily limit and took about 44% of the weekly quota; day two did it again, so nearly 90% of the week was gone in two days. That's a third-party harness, not Anthropic's plan, but it's the most numeric quota post in the window and a warning for anyone buying a wrapper mainly for Opus.
https://x.com/swarupdcs/status/2119070025147441495
(c) @swarupdcs tried Opus 5.5 through Devin Pro at $20 a month and logged it. Day one, about 1 hour 20 minutes of use exhausted the daily limit and took about 44% of the weekly quota; day two did it again, so nearly 90% of the week was gone in two days. That's a third-party harness, not Anthropic's plan, but it's the most numeric quota post in the window and a warning for anyone buying a wrapper mainly for Opus.
#14
@guoluolaep
https://x.com/guoluolaep/status/2118548256171159864
(c) @guoluolaep juggles Claude Code and Codex because of usage limits, and since typing isn't an option, every handoff means saying the whole task out loud again to the other agent. It's the human cost of quota nobody counts: the limit doesn't just stop work, it forces a full re-briefing.
https://x.com/guoluolaep/status/2118548256171159864
(c) @guoluolaep juggles Claude Code and Codex because of usage limits, and since typing isn't an option, every handoff means saying the whole task out loud again to the other agent. It's the human cost of quota nobody counts: the limit doesn't just stop work, it forces a full re-briefing.
#15
@lateascent
https://x.com/lateascent/status/2119507450613125236
(c) @lateascent pays $20 each for Claude Pro and ChatGPT Plus and noticed the usage screens count in opposite directions: Claude shows used, ChatGPT shows left. Claude Pro read 12% used on the session and 41% on the week; ChatGPT read 72% left on the week, which is 28% used. Half of the limit anger online comes from people comparing numbers that don't mean the same thing, and this post is a two-minute fix for that.
https://x.com/lateascent/status/2119507450613125236
(c) @lateascent pays $20 each for Claude Pro and ChatGPT Plus and noticed the usage screens count in opposite directions: Claude shows used, ChatGPT shows left. Claude Pro read 12% used on the session and 41% on the week; ChatGPT read 72% left on the week, which is 28% used. Half of the limit anger online comes from people comparing numbers that don't mean the same thing, and this post is a two-minute fix for that.
#16
@limitdex
https://x.com/limitdex/status/2119460816697094604
(c) @limitdex pulled the current wording of Claude's plans: Max at $100 is five times Pro every five hours and every week, Ultra at $200 is ten times Pro per week with no five-hour limit on verified accounts. The exact size of any of these isn't published. When the base unit is undisclosed, every multiple is unfalsifiable, and that's the structural reason these fights never end.
https://x.com/limitdex/status/2119460816697094604
(c) @limitdex pulled the current wording of Claude's plans: Max at $100 is five times Pro every five hours and every week, Ultra at $200 is ten times Pro per week with no five-hour limit on verified accounts. The exact size of any of these isn't published. When the base unit is undisclosed, every multiple is unfalsifiable, and that's the structural reason these fights never end.
#17
@v01dpr1mr0s3
https://x.com/v01dpr1mr0s3/status/2119542019282620774
(d) @v01dpr1mr0s3 asked Opus 5.5 to check a security issue, then had to mask words like pwn and hunt io in the follow-up because they kept tripping the classifier. The model was doing exactly the requested job and the guardrail fired on the vocabulary of that job. For security work this is the most common way Claude feels dumber: not a weaker model, a blocked one.
https://x.com/v01dpr1mr0s3/status/2119542019282620774
(d) @v01dpr1mr0s3 asked Opus 5.5 to check a security issue, then had to mask words like pwn and hunt io in the follow-up because they kept tripping the classifier. The model was doing exactly the requested job and the guardrail fired on the vocabulary of that job. For security work this is the most common way Claude feels dumber: not a weaker model, a blocked one.
#18
@shanemelaugh
https://x.com/shanemelaugh/status/2119192075505893695
(e) @shanemelaugh had Gemini 4 Argon analyze 1,400 Claude Code sessions and says something definitely happened on November 3: the model name stayed the same but the way it reasons took a sharp turn. It's the single most cited piece of evidence this weekend. Keep in mind what it is: one person's sessions, graded by a competitor's model, with no public method.
https://x.com/shanemelaugh/status/2119192075505893695
(e) @shanemelaugh had Gemini 4 Argon analyze 1,400 Claude Code sessions and says something definitely happened on November 3: the model name stayed the same but the way it reasons took a sharp turn. It's the single most cited piece of evidence this weekend. Keep in mind what it is: one person's sessions, graded by a competitor's model, with no public method.
#19
@chowgong985
https://x.com/chowgong985/status/2119534765153927205
(e) @chowgong985 summarized the numbers behind that 1,400-session analysis: after November 3, on high effort, 40.5% of turns had zero thinking and went straight to action, and the median thinking before acting fell from 84 to 54. Same model ID, same CLI version. If those figures hold up, the change is in thinking budget rather than in the weights, which would explain both the dumber feel and the quota burn.
https://x.com/chowgong985/status/2119534765153927205
(e) @chowgong985 summarized the numbers behind that 1,400-session analysis: after November 3, on high effort, 40.5% of turns had zero thinking and went straight to action, and the median thinking before acting fell from 84 to 54. Same model ID, same CLI version. If those figures hold up, the change is in thinking budget rather than in the weights, which would explain both the dumber feel and the quota burn.
#20
@anotherpanacea
https://x.com/anotherpanacea/status/2119545452370809313
(e) @anotherpanacea points out that BridgeBench used to post regular updates saying Opus 5.5 hadn't been nerfed and then stopped in mid-October. A missing data point isn't evidence of a nerf, but it removes the one independent check people leaned on. The pro-nerf side now has an argument from silence, and the other side has nothing fresh to answer with.
https://x.com/anotherpanacea/status/2119545452370809313
(e) @anotherpanacea points out that BridgeBench used to post regular updates saying Opus 5.5 hadn't been nerfed and then stopped in mid-October. A missing data point isn't evidence of a nerf, but it removes the one independent check people leaned on. The pro-nerf side now has an argument from silence, and the other side has nothing fresh to answer with.
#21
@marquistrillx
https://x.com/marquistrillx/status/2118812486396789130
(e) @marquistrillx gave the counter case in one breath: no official nerf, BridgeBench in mid-October still had Opus 5.5 at 105.8% of its launch score inside normal variance, and a chunk of the October complaints turned out to be Claude Code client bugs. That's the strongest counter in the window, and also the oldest data, which is exactly the gap the other side is exploiting.
https://x.com/marquistrillx/status/2118812486396789130
(e) @marquistrillx gave the counter case in one breath: no official nerf, BridgeBench in mid-October still had Opus 5.5 at 105.8% of its launch score inside normal variance, and a chunk of the October complaints turned out to be Claude Code client bugs. That's the strongest counter in the window, and also the oldest data, which is exactly the gap the other side is exploiting.
#22
@vib3coded
https://x.com/vib3coded/status/2119231952045625719
(e) @vib3coded answered the nerf posts with a test: a jelly sperm whale simulation that cost $5.03 and, by the poster's eye, stretched and rolled just as well as before. The poster is careful to say it can't speak for other tasks. One task that still works doesn't disprove a drop elsewhere, but it's a reminder that the same-prompt comparisons cut both ways.
https://x.com/vib3coded/status/2119231952045625719
(e) @vib3coded answered the nerf posts with a test: a jelly sperm whale simulation that cost $5.03 and, by the poster's eye, stretched and rolled just as well as before. The poster is careful to say it can't speak for other tasks. One task that still works doesn't disprove a drop elsewhere, but it's a reminder that the same-prompt comparisons cut both ways.
#23
@assembleRelease
https://x.com/assembleRelease/status/2119539604869226935
(e) @assembleRelease mocked the evidence chain directly: people trusting an Opus-degraded chart that was generated by Opus and posted by Opus. Snark, but the point stands. A chunk of this weekend's proof was AI-made analysis of AI behavior, and nobody checked the checker.
https://x.com/assembleRelease/status/2119539604869226935
(e) @assembleRelease mocked the evidence chain directly: people trusting an Opus-degraded chart that was generated by Opus and posted by Opus. Snark, but the point stands. A chunk of this weekend's proof was AI-made analysis of AI behavior, and nobody checked the checker.
#24
@rahulrajaramio
https://x.com/rahulrajaramio/status/2119543459795136758
(e) @rahulrajaramio noted that on a day everyone is seeing Opus nerfed, it's still uncannily good as an orchestrator. That's a useful split: if thinking per turn dropped, short dispatch decisions would suffer least and long single-agent reasoning most, which roughly matches where the complaints cluster.
https://x.com/rahulrajaramio/status/2119543459795136758
(e) @rahulrajaramio noted that on a day everyone is seeing Opus nerfed, it's still uncannily good as an orchestrator. That's a useful split: if thinking per turn dropped, short dispatch decisions would suffer least and long single-agent reasoning most, which roughly matches where the complaints cluster.
#25
@jan_volad
https://x.com/jan_volad/status/2119513284399223188
(e) @jan_volad spent a few days on Google AI Pro and came back with a limits comparison that runs against the weekend's mood. Argon burned the five-hour quota in five or six simple messages and a third of the weekly allowance in 20 to 30 minutes of coding, while chatting with Opus 5.5 on Claude Pro barely moved the usage bar. The conclusion: for $20, Claude Pro has the better model and much higher usable limits.
https://x.com/jan_volad/status/2119513284399223188
(e) @jan_volad spent a few days on Google AI Pro and came back with a limits comparison that runs against the weekend's mood. Argon burned the five-hour quota in five or six simple messages and a third of the weekly allowance in 20 to 30 minutes of coding, while chatting with Opus 5.5 on Claude Pro barely moved the usage bar. The conclusion: for $20, Claude Pro has the better model and much higher usable limits.
#26
@VianaAutomation
https://x.com/VianaAutomation/status/2119549272412832044
(e) @VianaAutomation compared the cheapest plans: Claude Pro had been running heavy workloads non-stop since Friday and sat at 52%, while Grok, activated less than a day earlier with barely any work, was already at 44%. It's a counterpoint on quota specifically, and a reminder that some of the anger this week is spillover from other products' cuts.
https://x.com/VianaAutomation/status/2119549272412832044
(e) @VianaAutomation compared the cheapest plans: Claude Pro had been running heavy workloads non-stop since Friday and sat at 52%, while Grok, activated less than a day earlier with barely any work, was already at 44%. It's a counterpoint on quota specifically, and a reminder that some of the anger this week is spillover from other products' cuts.
#27
@maixbt
https://x.com/maixbt/status/2119532338468278346
(f) @maixbt skips the nerf debate and goes after what usually eats the limit: one model doing every job. Opus 5.5 output costs $20 per million tokens and Haiku 5.5 costs $0.50, so a rename or a file search on Opus is 40 times the price. The proposed split is Haiku for search, renames, logs and one-line fixes, Sonnet for writing code, and Opus only for planning and review, with the honest caveat that this won't fix a real problem on Anthropic's side if there is one.
https://x.com/maixbt/status/2119532338468278346
(f) @maixbt skips the nerf debate and goes after what usually eats the limit: one model doing every job. Opus 5.5 output costs $20 per million tokens and Haiku 5.5 costs $0.50, so a rename or a file search on Opus is 40 times the price. The proposed split is Haiku for search, renames, logs and one-line fixes, Sonnet for writing code, and Opus only for planning and review, with the honest caveat that this won't fix a real problem on Anthropic's side if there is one.
#28
@BercusNA
https://x.com/BercusNA/status/2119185181936652450
(f) @BercusNA offers a fix on the user's side of the dumber feeling: a long CLAUDE.md. Citing Anthropic's own docs, the post says to keep it under 200 lines because Claude reads the whole file at the start of every session and follows long rule sets worse. The suggested layout is a short map in CLAUDE.md plus topic files Claude opens only when needed, and a rule that anything not written down gets answered with not written down instead of a guess.
https://x.com/BercusNA/status/2119185181936652450
(f) @BercusNA offers a fix on the user's side of the dumber feeling: a long CLAUDE.md. Citing Anthropic's own docs, the post says to keep it under 200 lines because Claude reads the whole file at the start of every session and follows long rule sets worse. The suggested layout is a short map in CLAUDE.md plus topic files Claude opens only when needed, and a rule that anything not written down gets answered with not written down instead of a guess.
#29
@thesam
https://x.com/thesam/status/2119542002496802883
(f) @thesam turned the weekend into an engineering rule: your coding model is not a stable employee. Developers report Opus 5.5 ignoring instructions and spinning up extra sub-agents for simple jobs while scores look fine. The advice is to pin whatever version passed your own evals and keep a fallback ready, which is the only defense that doesn't depend on Anthropic saying anything.
https://x.com/thesam/status/2119542002496802883
(f) @thesam turned the weekend into an engineering rule: your coding model is not a stable employee. Developers report Opus 5.5 ignoring instructions and spinning up extra sub-agents for simple jobs while scores look fine. The advice is to pin whatever version passed your own evals and keep a fallback ready, which is the only defense that doesn't depend on Anthropic saying anything.
#30
Signal Board: 17 complaints vs 6 counter-evidence posts vs 0 official responses on model quality, plus 3 posts arguing the evidence points to a real change.
Comments