Nerfed Daily: 2026-11-16
The nerf talk this window is spread thin and mostly vibes. Opus 5.5 inside Claude Code takes the most hits again, with people saying it feels back at Opus 4.8 level or makes mistake after mistake across four sessions, and a few now say Fable 5.2 has already been nerfed too. The sharper complaints are about the safety layer, not the model: a classifier that pulls Fable sessions down to Opus 5, including one person asking about muscular dystrophy trials, and a Fable classifier that trips on a quoted Opus 3 tweet. Hard evidence still points the other way. The daily trackers show no drop, and the best number of the window is a user's own week-against-week test where Ultracode and med/high effort on the same Opus 5.5 scored within noise on quality. Anthropic did respond, but only to the outages: two degraded-performance incidents on Friday, both resolved, no cause published. Everything else official was docs, on Code Mode, Projects and how plan limits work, and none of it touches model quality.
#1
@mehulmpt
https://x.com/mehulmpt/status/2121946029398606192
@mehulmpt posted the window's most-liked nerf joke: when Opus 5.5 starts getting dumber, you call in the big guns (a). There is no test behind it, just a screenshot-style punchline. It still matters as a mood reading, because the joke only lands if a lot of people already believe Opus dips mid-week and the fix is to switch up to Fable. That belief is now common enough to be a meme.
https://x.com/mehulmpt/status/2121946029398606192
@mehulmpt posted the window's most-liked nerf joke: when Opus 5.5 starts getting dumber, you call in the big guns (a). There is no test behind it, just a screenshot-style punchline. It still matters as a mood reading, because the joke only lands if a lot of people already believe Opus dips mid-week and the fix is to switch up to Fable. That belief is now common enough to be a meme.
#2
@ayesha_fatiima
https://x.com/ayesha_fatiima/status/2121884847660753261
@ayesha_fatiima asks the question plainly: is Opus 5.5 actually getting worse, or are expectations just too high (a)? Yesterday the same account listed concrete failures, and today it is already second-guessing itself. That shift is the honest version of most nerf posts, a felt drop with no way to tell if the model moved or the bar did.
https://x.com/ayesha_fatiima/status/2121884847660753261
@ayesha_fatiima asks the question plainly: is Opus 5.5 actually getting worse, or are expectations just too high (a)? Yesterday the same account listed concrete failures, and today it is already second-guessing itself. That shift is the honest version of most nerf posts, a felt drop with no way to tell if the model moved or the bar did.
#3
@baskind
https://x.com/baskind/status/2121276283132825697
@baskind says something about Opus 5.5 lately makes it unusable, with four sessions of mistake after mistake, and asks if anyone else is seeing degraded performance (a). It comes with a screenshot but no task description or comparison. The post is from Nov 13, the same afternoon Anthropic was investigating degraded performance on Claude Code, so part of this may be the outage rather than the model.
https://x.com/baskind/status/2121276283132825697
@baskind says something about Opus 5.5 lately makes it unusable, with four sessions of mistake after mistake, and asks if anyone else is seeing degraded performance (a). It comes with a screenshot but no task description or comparison. The post is from Nov 13, the same afternoon Anthropic was investigating degraded performance on Claude Code, so part of this may be the outage rather than the model.
#4
@develeop
https://x.com/develeop/status/2121415117552435383
@develeop says Opus 5.5 has been acting badly lately and fears it got nerfed, then asks if Anthropic is planning a new model as an IPO celebration (a). That is the classic theory: the old model gets worse right before a new one ships. No evidence comes with it, but the timing link with the IPO news is new.
https://x.com/develeop/status/2121415117552435383
@develeop says Opus 5.5 has been acting badly lately and fears it got nerfed, then asks if Anthropic is planning a new model as an IPO celebration (a). That is the classic theory: the old model gets worse right before a new one ships. No evidence comes with it, but the timing link with the IPO news is new.
#5
@albustime
https://x.com/albustime/status/2122016916139319735
@albustime calls it insane that Anthropic has already nerfed Fable 5.2, and calls it brazen (a). There is no example and no number, only the claim. Fable 5.2 is barely two weeks old, so the fact that the nerf story starts this early says more about trust than about the model.
https://x.com/albustime/status/2122016916139319735
@albustime calls it insane that Anthropic has already nerfed Fable 5.2, and calls it brazen (a). There is no example and no number, only the claim. Fable 5.2 is barely two weeks old, so the fact that the nerf story starts this early says more about trust than about the model.
#6
@DtoTheMill
https://x.com/DtoTheMill/status/2122057026503053399
@DtoTheMill replies that Claude sucks now, that it has been lobotomized, and tells the person to cancel and get a refund while possible (a). It is the angriest post of the window and also the least specific. It is listed because the refund advice is where these threads end up when nobody can show a test.
https://x.com/DtoTheMill/status/2122057026503053399
@DtoTheMill replies that Claude sucks now, that it has been lobotomized, and tells the person to cancel and get a refund while possible (a). It is the angriest post of the window and also the least specific. It is listed because the refund advice is where these threads end up when nobody can show a test.
#7
@fadeleusdev
https://x.com/fadeleusdev/status/2121421549844783248
@fadeleusdev says Opus 5.5 got nerfed hard and feels back at Opus 4.8 level (a). Naming a specific older model is more useful than saying dumber, because it can be checked. It also lines up with the known fallback behaviour, where some sessions really are answered by Opus 4.8, which would feel exactly like this.
https://x.com/fadeleusdev/status/2121421549844783248
@fadeleusdev says Opus 5.5 got nerfed hard and feels back at Opus 4.8 level (a). Naming a specific older model is more useful than saying dumber, because it can be checked. It also lines up with the known fallback behaviour, where some sessions really are answered by Opus 4.8, which would feel exactly like this.
#8
@OmkarChandra
https://x.com/OmkarChandra/status/2121401189078102375
@OmkarChandra says the classifier switched the session to Opus 5 while asking only about muscular dystrophy clinical trial research (b). The account has already unsubscribed and is using up the paid time. The question it raises is sharp: if Opus 5 is the default safe model for bio questions, is that model held back on purpose? This is the pattern that makes people call it a nerf, a quiet swap on a harmless medical question.
https://x.com/OmkarChandra/status/2121401189078102375
@OmkarChandra says the classifier switched the session to Opus 5 while asking only about muscular dystrophy clinical trial research (b). The account has already unsubscribed and is using up the paid time. The question it raises is sharp: if Opus 5 is the default safe model for bio questions, is that model held back on purpose? This is the pattern that makes people call it a nerf, a quiet swap on a harmless medical question.
#9
@KryptoKum
https://x.com/KryptoKum/status/2121233459507036608
@KryptoKum says all active Fable 5.2 sessions were silently degraded to Opus 5, not even Opus 5.5 (b). Posted on Nov 13, the day of the afternoon incident, so it may be the outage path rather than the classifier. Either way the complaint is the same: the switch was silent, and the user only found out by checking the model.
https://x.com/KryptoKum/status/2121233459507036608
@KryptoKum says all active Fable 5.2 sessions were silently degraded to Opus 5, not even Opus 5.5 (b). Posted on Nov 13, the day of the afternoon incident, so it may be the outage path rather than the classifier. Either way the complaint is the same: the switch was silent, and the user only found out by checking the model.
#10
@IrfanLateefUS
https://x.com/IrfanLateefUS/status/2121736979742347420
@IrfanLateefUS runs out of Claude usage limits by Wednesday while the reset comes on Saturday (c). The workaround question is how to move coding tasks from Claude Code to Codex, where there are spare credits, without losing quality. Three dead days a week is the concrete cost of the limits, and the answer people reach for is a second vendor.
https://x.com/IrfanLateefUS/status/2121736979742347420
@IrfanLateefUS runs out of Claude usage limits by Wednesday while the reset comes on Saturday (c). The workaround question is how to move coding tasks from Claude Code to Codex, where there are spare credits, without losing quality. Three dead days a week is the concrete cost of the limits, and the answer people reach for is a second vendor.
#11
@OmAgale22
https://x.com/OmAgale22/status/2122021310972977315
@OmAgale22 asks Anthropic directly whether something changed with Claude Pro limits, because the 5-hour cap feels like it drains three times faster for the exact same workload (c). Three times is a felt number, not a measured one. It is the same complaint as last week, when Anthropic admitted a metering bug and handed out a reset.
https://x.com/OmAgale22/status/2122021310972977315
@OmAgale22 asks Anthropic directly whether something changed with Claude Pro limits, because the 5-hour cap feels like it drains three times faster for the exact same workload (c). Three times is a felt number, not a measured one. It is the same complaint as last week, when Anthropic admitted a metering bug and handed out a reset.
#12
@awesamarth_
https://x.com/awesamarth_/status/2122072826006102440
@awesamarth_ dropped to the $20 ChatGPT Plus plus $20 Claude Pro for an exam month, and says using Fable for more than three prompts would burn through the limits (c). That is a clean picture of the Pro tier right now: Fable is technically there, but three prompts is a demo, not a workday.
https://x.com/awesamarth_/status/2122072826006102440
@awesamarth_ dropped to the $20 ChatGPT Plus plus $20 Claude Pro for an exam month, and says using Fable for more than three prompts would burn through the limits (c). That is a clean picture of the Pro tier right now: Fable is technically there, but three prompts is a demo, not a workday.
#13
@MythomancerX
https://x.com/MythomancerX/status/2122048839620452598
@MythomancerX reads Anthropic's fix for the metering bugs as one reset you have to claim yourself by December 10 (c). The point is the missing number: how many people hit their weekly cap early because of those bugs and bought extra usage? That figure, not a coupon, is what this user wants published. It is the clearest statement of why the reset did not close the issue.
https://x.com/MythomancerX/status/2122048839620452598
@MythomancerX reads Anthropic's fix for the metering bugs as one reset you have to claim yourself by December 10 (c). The point is the missing number: how many people hit their weekly cap early because of those bugs and bought extra usage? That figure, not a coupon, is what this user wants published. It is the clearest statement of why the reset did not close the issue.
#14
@JohnWittle
https://x.com/JohnWittle/status/2122080147436408910
@JohnWittle says the Fable 5.2 classifiers over the past weeks have been really frustrating (d). The example is concrete: while indexing a dataset of old model outputs, a quoted Opus 3 passage that someone had posted tripped the general_harms classifier. The theory offered is that the classifiers are tuned for false positives and then only walked back where they block paid work, so anything outside labor, like talking about model psychology, keeps hitting walls.
https://x.com/JohnWittle/status/2122080147436408910
@JohnWittle says the Fable 5.2 classifiers over the past weeks have been really frustrating (d). The example is concrete: while indexing a dataset of old model outputs, a quoted Opus 3 passage that someone had posted tripped the general_harms classifier. The theory offered is that the classifiers are tuned for false positives and then only walked back where they block paid work, so anything outside labor, like talking about model psychology, keeps hitting walls.
#15
@netdur
https://x.com/netdur/status/2122086593540284834
@netdur agrees with a public complaint that Anthropic is using a dumb classifier that denies requests, maybe on keywords alone (d). The argument is about value: you pay for results you do not get, while other vendors will just help. No example of the denied request is given, so this is the sentiment rather than a case.
https://x.com/netdur/status/2122086593540284834
@netdur agrees with a public complaint that Anthropic is using a dumb classifier that denies requests, maybe on keywords alone (d). The argument is about value: you pay for results you do not get, while other vendors will just help. No example of the denied request is given, so this is the sentiment rather than a case.
#16
@JacobDev15
https://x.com/JacobDev15/status/2122068807388627066
@JacobDev15 points out that the Claude Code auto mode classifier reads only the user's messages, never Claude's (d). So a line like i approve, go ahead gets denied, while go ahead and push to prod gets through. The post calls it annoying and correct, because an agent that can argue with its own guardrail would eventually win the argument. Many classifier complaints are really this design choice showing up.
https://x.com/JacobDev15/status/2122068807388627066
@JacobDev15 points out that the Claude Code auto mode classifier reads only the user's messages, never Claude's (d). So a line like i approve, go ahead gets denied, while go ahead and push to prod gets through. The post calls it annoying and correct, because an agent that can argue with its own guardrail would eventually win the argument. Many classifier complaints are really this design choice showing up.
#17
@bread_
https://x.com/bread_/status/2122054401816977432
@bread_ built a fork of the t3 harness with a dedicated orchestrator, then killed it, and says after a lot of rapid thread work there is still no degradation to report (e). What the experiment did show is a wish for defaults: being able to say design work goes to Opus 5.5 and subtasks go to a cheaper model, instead of typing it dozens of times a day. Heavy daily use with no felt drop is a useful counterweight to the one-line complaints.
https://x.com/bread_/status/2122054401816977432
@bread_ built a fork of the t3 harness with a dedicated orchestrator, then killed it, and says after a lot of rapid thread work there is still no degradation to report (e). What the experiment did show is a wish for defaults: being able to say design work goes to Opus 5.5 and subtasks go to a cheaper model, instead of typing it dozens of times a day. Heavy daily use with no felt drop is a useful counterweight to the one-line complaints.
#18
@olveraruize
https://x.com/olveraruize/status/2122024295677169748
@olveraruize replies that Opus 5.5 is too good and asks Anthropic not to nerf it (e). It is a short post, but it shows the other half of the crowd on the same day. Some people are worried about a future nerf, not reporting a current one.
https://x.com/olveraruize/status/2122024295677169748
@olveraruize replies that Opus 5.5 is too good and asks Anthropic not to nerf it (e). It is a short post, but it shows the other half of the crowd on the same day. Some people are worried about a future nerf, not reporting a current one.
#19
@RudyXFantasy
https://x.com/RudyXFantasy/status/2122065428457628098
@RudyXFantasy compares plans and says Gemini Argon's limits are still bad for Pro users, while Opus 5.5 on the same kind of plan and the same data analysis lasts many hours (e). On quality, Argon is rated as hallucinating less than Hydrogen. As a counterpoint to the limit complaints, this is one user for whom Claude is the generous option.
https://x.com/RudyXFantasy/status/2122065428457628098
@RudyXFantasy compares plans and says Gemini Argon's limits are still bad for Pro users, while Opus 5.5 on the same kind of plan and the same data analysis lasts many hours (e). On quality, Argon is rated as hallucinating less than Hydrogen. As a counterpoint to the limit complaints, this is one user for whom Claude is the generous option.
#20
@bsmokes
https://x.com/bsmokes/status/2122074496089931892
@bsmokes ran the best test of the window: Ultracode against med/high effort, same Opus 5.5, same Expo and React Native codebase, one week each (e). Med/high merged about 3.2 times more PRs, 187 against 58, with a median merge time of 6.2 hours against 16.6. In a blind review of 20 PRs, Ultracode scored 3.91 against 3.69 out of 5, but two scores for the same diff differed by 0.36 on average, so the gap sits inside the noise. The author lists the caveats honestly: different weeks, different work mix, changed tooling. The takeaway is that the expensive mode buys at most a small quality edge, which also says the base model was not falling apart that week.
https://x.com/bsmokes/status/2122074496089931892
@bsmokes ran the best test of the window: Ultracode against med/high effort, same Opus 5.5, same Expo and React Native codebase, one week each (e). Med/high merged about 3.2 times more PRs, 187 against 58, with a median merge time of 6.2 hours against 16.6. In a blind review of 20 PRs, Ultracode scored 3.91 against 3.69 out of 5, but two scores for the same diff differed by 0.36 on average, so the gap sits inside the noise. The author lists the caveats honestly: different weeks, different work mix, changed tooling. The takeaway is that the expensive mode buys at most a small quality edge, which also says the base model was not falling apart that week.
#21
@TCCardoza
https://x.com/TCCardoza/status/2122038136226062463
@TCCardoza does not think Opus is being nerfed and says expectations change instead (e). The same reply adds that Sol cannot keep up with Opus 5.5. It is a plain I do not feel it, but it comes from someone who uses both and still ranks Opus first.
https://x.com/TCCardoza/status/2122038136226062463
@TCCardoza does not think Opus is being nerfed and says expectations change instead (e). The same reply adds that Sol cannot keep up with Opus 5.5. It is a plain I do not feel it, but it comes from someone who uses both and still ranks Opus first.
#22
@PiotrBoruta1
https://x.com/PiotrBoruta1/status/2122089698835280056
@PiotrBoruta1 says after a few hiccups Fable 5.2 is an amazing model, but not an all-rounder the way 5.1 was (e). It struggles with surprisingly simple tasks and shines on hard, ambitious ones, mostly used for planning and coordinating Opus agents. It devours usage limits, though High effort has been enough. This is a useful reframe: some of the Fable nerf reports may be people giving it easy jobs it is oddly bad at.
https://x.com/PiotrBoruta1/status/2122089698835280056
@PiotrBoruta1 says after a few hiccups Fable 5.2 is an amazing model, but not an all-rounder the way 5.1 was (e). It struggles with surprisingly simple tasks and shines on hard, ambitious ones, mostly used for planning and coordinating Opus agents. It devours usage limits, though High effort has been enough. This is a useful reframe: some of the Fable nerf reports may be people giving it easy jobs it is oddly bad at.
#23
@thepatchnotesai
https://x.com/thepatchnotesai/status/2121650132982579288
@thepatchnotesai sums up the gap: thousands of people say Claude got dumber, Anthropic launched a new model the same day, and the trackers that test Opus 5.5 every day show no drop (e). It does not explain the felt decline, but it is the standing counter-evidence. Until a tracker moves, the complaints remain a perception story plus the known fallback path.
https://x.com/thepatchnotesai/status/2121650132982579288
@thepatchnotesai sums up the gap: thousands of people say Claude got dumber, Anthropic launched a new model the same day, and the trackers that test Opus 5.5 every day show no drop (e). It does not explain the felt decline, but it is the standing counter-evidence. Until a tracker moves, the complaints remain a perception story plus the known fallback path.
#24
@OneSignalDesk
https://x.com/OneSignalDesk/status/2121234445940240676
@OneSignalDesk relays Anthropic's status update: errors were back on Nov 13, with degraded performance on the web app, the API and Claude Code from 13:09 UTC, about 70 minutes after an earlier sign-in outage was marked resolved (f). This is the one official acknowledgement of degradation this window, and it is about availability, not model quality. Some of the same-day nerf posts likely sit inside this window.
https://x.com/OneSignalDesk/status/2121234445940240676
@OneSignalDesk relays Anthropic's status update: errors were back on Nov 13, with degraded performance on the web app, the API and Claude Code from 13:09 UTC, about 70 minutes after an earlier sign-in outage was marked resolved (f). This is the one official acknowledgement of degradation this window, and it is about availability, not model quality. Some of the same-day nerf posts likely sit inside this window.
#25
@joshwilson_will
https://x.com/joshwilson_will/status/2122014826402906576
@joshwilson_will recaps Friday: logins failed in the morning, then the app, API and coding tools degraded in the afternoon (f). Both are fixed now, and no cause has been published. That missing cause is the gap people fill with nerf theories.
https://x.com/joshwilson_will/status/2122014826402906576
@joshwilson_will recaps Friday: logins failed in the morning, then the app, API and coding tools degraded in the afternoon (f). Both are fixed now, and no cause has been published. That missing cause is the gap people fill with nerf theories.
#26
@DYoung_AI
https://x.com/DYoung_AI/status/2121413105213194417
@DYoung_AI walks through Anthropic's docs for Claude Code Projects, now in public beta on Pro, Max and Ultra (f). The part that matters for limit complaints: a new project runs every thread on Opus at high effort, which the docs say draws on the plan fastest, project usage counts against the same limits and uses them faster, and auto-continue on reset is on by default. Anyone who started a wide first batch without changing the thread model has a documented reason for a fast drain.
https://x.com/DYoung_AI/status/2121413105213194417
@DYoung_AI walks through Anthropic's docs for Claude Code Projects, now in public beta on Pro, Max and Ultra (f). The part that matters for limit complaints: a new project runs every thread on Opus at high effort, which the docs say draws on the plan fastest, project usage counts against the same limits and uses them faster, and auto-continue on reset is on by default. Anyone who started a wide first batch without changing the thread model has a documented reason for a fast drain.
#27
@YashNautiyal_
https://x.com/YashNautiyal_/status/2122060897984114937
@YashNautiyal_ explains the new experimental Code Mode: instead of one tool call at a time, Claude writes a small JavaScript script that runs several calls in one go, some in parallel (f). It needs v2.1.323 or later and an interactive terminal session, enabled with an environment variable. The warning is about auto mode: the classifier reviews the script, not each call, so approval becomes a batch.
https://x.com/YashNautiyal_/status/2122060897984114937
@YashNautiyal_ explains the new experimental Code Mode: instead of one tool call at a time, Claude writes a small JavaScript script that runs several calls in one go, some in parallel (f). It needs v2.1.323 or later and an interactive terminal session, enabled with an environment variable. The warning is about auto mode: the classifier reviews the script, not each call, so approval becomes a batch.
#28
@JacqBots
https://x.com/JacqBots/status/2122028951363084300
@JacqBots says Claude got worse the longer the conversation ran, and the fix was simple (f). Ask it to summarize what works and what is left, open a fresh chat, paste it in. Same model, much better output, and the team says it built its checkout product this way without anyone coding. A lot of felt nerf is long-context drift, and this is the cheapest workaround.
https://x.com/JacqBots/status/2122028951363084300
@JacqBots says Claude got worse the longer the conversation ran, and the fix was simple (f). Ask it to summarize what works and what is left, open a fresh chat, paste it in. Same model, much better output, and the team says it built its checkout product this way without anyone coding. A lot of felt nerf is long-context drift, and this is the cheapest workaround.
#29
@anilpervaiz
https://x.com/anilpervaiz/status/2122017599307550807
@anilpervaiz adds the control side of Code Mode: in auto mode the classifier reviews the batched code but does not see values computed at run time, such as a path taken from search output (f). Deny and ask rules and PreToolUse hooks still apply per call, and the docs say to keep those for calls you must control. The advice is to keep the per-call rules on and add a hook that reads the code before it runs.
https://x.com/anilpervaiz/status/2122017599307550807
@anilpervaiz adds the control side of Code Mode: in auto mode the classifier reviews the batched code but does not see values computed at run time, such as a path taken from search output (f). Deny and ask rules and PreToolUse hooks still apply per call, and the docs say to keep those for calls you must control. The advice is to keep the per-call rules on and add a hook that reads the code before it runs.
#30
@AlloyPress
https://x.com/AlloyPress/status/2122072151004164251
@AlloyPress summarizes Anthropic's Nov 13 guide on building safety evals (f). One line explains a lot of the felt gap between surfaces: for one model, the difference between the raw API and claude ai with its system prompt was as high as 34 percentage points on multi-turn safety evals. Settings move behaviour that much, so the same model can feel different in the app, in Claude Code and on the API.
https://x.com/AlloyPress/status/2122072151004164251
@AlloyPress summarizes Anthropic's Nov 13 guide on building safety evals (f). One line explains a lot of the felt gap between surfaces: for one model, the difference between the raw API and claude ai with its system prompt was as high as 34 percentage points on multi-turn safety evals. Settings move behaviour that much, so the same model can feel different in the app, in Claude Code and on the API.
#31
@curatedbyap
https://x.com/curatedbyap/status/2122087250930352639
@curatedbyap clears up a common misread of the $200 plan: it covers Claude chat and Claude Code with five-hour and weekly limits, and the API credits that come with it are for API apps (f). They do not stretch Claude Code limits. Some limit complaints come from people who expected those credits to act as overflow.
https://x.com/curatedbyap/status/2122087250930352639
@curatedbyap clears up a common misread of the $200 plan: it covers Claude chat and Claude Code with five-hour and weekly limits, and the API credits that come with it are for API apps (f). They do not stretch Claude Code limits. Some limit complaints come from people who expected those credits to act as overflow.
#32
Signal Board
Today: 16 complaints vs 7 evidence or counter-evidence posts vs 6 official items from Anthropic, none of which addresses model quality.
Today: 16 complaints vs 7 evidence or counter-evidence posts vs 6 official items from Anthropic, none of which addresses model quality.
Comments