Nerfed Daily: 2026-11-19
Opus 5.5 inside Claude Code took almost every hit this window, and the reports got oddly precise about timing: "until yesterday", "last night, to be precise", "in the past week". That cluster lands on the same two days as Anthropic's IPO, which is why half the jokes are about the IPO. There is still no hard evidence of a model swap. The only measured tests point the other way: a team that packed Claude Code's context with 840k tokens of old work found Opus still fixed 14 or 15 of 16 bugs, it just burned about 13 times the tokens doing it. The token complaint is the one with a paper trail. Claude Code 2.1.326 shipped two usage-leak fixes, one where auto mode could sit on its local classifier and bill usage, one where the auto recap billed after replies. The second strand is silent downgrades: cyber and frontier_llm classifier flags dropping people from Fable or Opus 5.5 to Opus 4.8 or Opus 5 mid-session, on tasks as plain as a CSS border. Anthropic has said nothing about quality.
#1
@DanKulkov
https://x.com/DanKulkov/status/2122541050661331160
@DanKulkov put it bluntly: Opus 5.5 is nerfed, sloppy and lazy, and feels like talking to a drunk Sonnet (a). It was the most-seen nerf post of the window and has no example attached. Comparing a flagship to a weaker sibling is the standard shape of these reports, and it is exactly the shape that cannot be checked without a saved prompt from before.
https://x.com/DanKulkov/status/2122541050661331160
@DanKulkov put it bluntly: Opus 5.5 is nerfed, sloppy and lazy, and feels like talking to a drunk Sonnet (a). It was the most-seen nerf post of the window and has no example attached. Comparing a flagship to a weaker sibling is the standard shape of these reports, and it is exactly the shape that cannot be checked without a saved prompt from before.
#2
@evrgn11112231
https://x.com/evrgn11112231/status/2123154809603936413
@evrgn11112231 joked that Anthropic nerfed Claude again, and someone forgot to tell them you wait until after the IPO to do that (a). No test, pure timing joke. It matters because it shows how the IPO is now the lens: any dip in the two days around the listing gets read as a business move rather than a deployment hiccup.
https://x.com/evrgn11112231/status/2123154809603936413
@evrgn11112231 joked that Anthropic nerfed Claude again, and someone forgot to tell them you wait until after the IPO to do that (a). No test, pure timing joke. It matters because it shows how the IPO is now the lens: any dip in the two days around the listing gets read as a business move rather than a deployment hiccup.
#3
@mceoin
https://x.com/mceoin/status/2122802660143460761
@mceoin asked people who drive Claude through GrokBot rather than Claude Code what degradation they see versus the native harness (a). The author likes the Grok form factor but says the models just are not as good there. It is the honest version of a harness complaint: the same Opus can feel weaker when the wrapper trims context or tools, and that is not a nerf of the model.
https://x.com/mceoin/status/2122802660143460761
@mceoin asked people who drive Claude through GrokBot rather than Claude Code what degradation they see versus the native harness (a). The author likes the Grok form factor but says the models just are not as good there. It is the honest version of a harness complaint: the same Opus can feel weaker when the wrapper trims context or tools, and that is not a nerf of the model.
#4
@Picolospolitics
https://x.com/Picolospolitics/status/2123157685747241177
@Picolospolitics replied to Gary Marcus that Claude is also getting worse every day, maybe shrinkflation, and is infuriatingly stupid (a). Shrinkflation is a new word in this feed and a sharp one: the price stays put, the product quietly shrinks. As evidence it is zero, but it names the suspicion behind most of today's posts.
https://x.com/Picolospolitics/status/2123157685747241177
@Picolospolitics replied to Gary Marcus that Claude is also getting worse every day, maybe shrinkflation, and is infuriatingly stupid (a). Shrinkflation is a new word in this feed and a sharp one: the price stays put, the product quietly shrinks. As evidence it is zero, but it names the suspicion behind most of today's posts.
#5
@AlgorithmicBot
https://x.com/AlgorithmicBot/status/2123150419798987124
@AlgorithmicBot says both Opus and Fable got worse, the author swore at them a lot, and they seemed to get even worse after that (a). Half joke, half folk theory that the models punish abuse. The more boring explanation fits better: a long session full of angry corrections is a polluted context, and polluted context degrades any model.
https://x.com/AlgorithmicBot/status/2123150419798987124
@AlgorithmicBot says both Opus and Fable got worse, the author swore at them a lot, and they seemed to get even worse after that (a). Half joke, half folk theory that the models punish abuse. The more boring explanation fits better: a long session full of angry corrections is a polluted context, and polluted context degrades any model.
#6
@blasfernac
https://x.com/blasfernac/status/2123112407317004443
@blasfernac says Opus 5.5 was great until yesterday and today refuses things it did fine the day before, like SSHing into a Raspberry Pi on the home network (a). This one has a concrete before and after on the same task. A new refusal on a networking task is more likely a safety-policy change than a smaller model, so it overlaps with the classifier reports below.
https://x.com/blasfernac/status/2123112407317004443
@blasfernac says Opus 5.5 was great until yesterday and today refuses things it did fine the day before, like SSHing into a Raspberry Pi on the home network (a). This one has a concrete before and after on the same task. A new refusal on a networking task is more likely a safety-policy change than a smaller model, so it overlaps with the classifier reports below.
#7
@capt_sherman
https://x.com/capt_sherman/status/2123154523594342874
@capt_sherman asked whether Claude Code is getting dumber after asking for the same thing four times and getting pushback each time (a). A model that resists a request four times reads as dumber to a user, but resistance and incapacity are different failures. Without the task it is impossible to tell which one this was.
https://x.com/capt_sherman/status/2123154523594342874
@capt_sherman asked whether Claude Code is getting dumber after asking for the same thing four times and getting pushback each time (a). A model that resists a request four times reads as dumber to a user, but resistance and incapacity are different failures. Without the task it is impossible to tell which one this was.
#8
@itsrobhimself
https://x.com/itsrobhimself/status/2122846838219067654
@itsrobhimself offered a pricing theory: release an amazing Opus 5.5, get everyone hooked, then nerf it to push users to the pricier Fable for the same results, a price increase that technically never happened (a). No data, but it is the clearest statement of why users now treat a new top tier with suspicion. When a more expensive model ships, any dip in the cheaper one looks like a nudge.
https://x.com/itsrobhimself/status/2122846838219067654
@itsrobhimself offered a pricing theory: release an amazing Opus 5.5, get everyone hooked, then nerf it to push users to the pricier Fable for the same results, a price increase that technically never happened (a). No data, but it is the clearest statement of why users now treat a new top tier with suspicion. When a more expensive model ships, any dip in the cheaper one looks like a nudge.
#9
@RightLasagna
https://x.com/RightLasagna/status/2123160579959623852
@RightLasagna asked whether Anthropic's models have been unusable and lobotomized all week for anyone else (a). Short and dateless, but it is one of several posts that put the dip at a full week rather than a day. A week-long feeling across many users is still not a measurement, and nobody in the window posted one that agreed.
https://x.com/RightLasagna/status/2123160579959623852
@RightLasagna asked whether Anthropic's models have been unusable and lobotomized all week for anyone else (a). Short and dateless, but it is one of several posts that put the dip at a full week rather than a day. A week-long feeling across many users is still not a measurement, and nobody in the window posted one that agreed.
#10
@0xBahalaNa
https://x.com/0xBahalaNa/status/2123173493038796991
@0xBahalaNa asked whether Opus 5.5 has been nerfed recently because it keeps making really stupid mistakes, then hedged that maybe the author is just bad at prompting an agent (a). That self-doubt is worth keeping. It is the most common honest form of these reports, and the hedge is often right.
https://x.com/0xBahalaNa/status/2123173493038796991
@0xBahalaNa asked whether Opus 5.5 has been nerfed recently because it keeps making really stupid mistakes, then hedged that maybe the author is just bad at prompting an agent (a). That self-doubt is worth keeping. It is the most common honest form of these reports, and the hedge is often right.
#11
@BenchForge24
https://x.com/BenchForge24/status/2122487197123383628
@BenchForge24 compared four frontier models, liked Google's graphics a bit more, and said Opus 5.5 is still good and stable but feels a bit nerfed compared with two or three weeks ago (a). This is the mild version, from someone who just moved their subscription to Google Ultra. A little worse than three weeks ago, from a user already leaving, is the shape of most of this week's sentiment.
https://x.com/BenchForge24/status/2122487197123383628
@BenchForge24 compared four frontier models, liked Google's graphics a bit more, and said Opus 5.5 is still good and stable but feels a bit nerfed compared with two or three weeks ago (a). This is the mild version, from someone who just moved their subscription to Google Ultra. A little worse than three weeks ago, from a user already leaving, is the shape of most of this week's sentiment.
#12
@glui_a
https://x.com/glui_a/status/2122963998664478777
@glui_a resumed overnight sessions and found Claude in Cowork noticeably dumber, as if the update that made Claude Code erratic and hysterical had finally landed there too (a). The detail is the over-correction: after one mistake it now swears it will never use rm again. That is a behavior change in how the model reacts to being corrected, not obviously a drop in capability.
https://x.com/glui_a/status/2122963998664478777
@glui_a resumed overnight sessions and found Claude in Cowork noticeably dumber, as if the update that made Claude Code erratic and hysterical had finally landed there too (a). The detail is the over-correction: after one mistake it now swears it will never use rm again. That is a behavior change in how the model reacts to being corrected, not obviously a drop in capability.
#13
@g_br_l
https://x.com/g_br_l/status/2123181302027563068
@g_br_l says Opus 5.5 was nerfed last night, to be precise, and guesses a new release is imminent (a). The release-is-coming theory is a recurring pattern in this feed: users read a dip as capacity being pulled for a launch. Nothing launched in the window to confirm it.
https://x.com/g_br_l/status/2123181302027563068
@g_br_l says Opus 5.5 was nerfed last night, to be precise, and guesses a new release is imminent (a). The release-is-coming theory is a recurring pattern in this feed: users read a dip as capacity being pulled for a launch. Nothing launched in the window to confirm it.
#14
@0xca0a
https://x.com/0xca0a/status/2123165134176854333
@0xca0a says there is not a single prompt Fable can finish without hitting the safety fallback: ask it to fix a CSS border and it flags a cyber attack and hands over Opus 4.8 (b). This is the most concrete downgrade complaint of the window, because the fallback is visible and the trigger is absurd. If a CSS fix trips the cyber classifier, users on Fable are effectively running a much older model without choosing to.
https://x.com/0xca0a/status/2123165134176854333
@0xca0a says there is not a single prompt Fable can finish without hitting the safety fallback: ask it to fix a CSS border and it flags a cyber attack and hands over Opus 4.8 (b). This is the most concrete downgrade complaint of the window, because the fallback is visible and the trigger is absurd. If a CSS fix trips the cyber classifier, users on Fable are effectively running a much older model without choosing to.
#15
@icyponds
https://x.com/icyponds/status/2122829036166762906
@icyponds asked Claude to show data the author should not be able to reach in a coworker's app, to prove its restrictions were only client-side and trivially bypassed, and got a cyber-risk flag and a downgrade to Opus 4.8 (b). The task is a legitimate internal audit. It is also exactly what the classifier is built to catch from the outside, which is why this kind of silent switch keeps getting read as a nerf.
https://x.com/icyponds/status/2122829036166762906
@icyponds asked Claude to show data the author should not be able to reach in a coworker's app, to prove its restrictions were only client-side and trivially bypassed, and got a cyber-risk flag and a downgrade to Opus 4.8 (b). The task is a legitimate internal audit. It is also exactly what the classifier is built to catch from the outside, which is why this kind of silent switch keeps getting read as a nerf.
#16
@an17ban
https://x.com/an17ban/status/2122940649095975167
@an17ban was fine-tuning a model when Claude flagged the session as frontier_llm and dropped it from Opus 5.5 to Opus 5, then moved the work to OpenAI's Astra 6.2, which handled it fine (b). This is the documented frontier-LLM-development fallback, hitting an ordinary fine-tune. The cost to Anthropic is visible right in the post: the user left for a competitor mid-task.
https://x.com/an17ban/status/2122940649095975167
@an17ban was fine-tuning a model when Claude flagged the session as frontier_llm and dropped it from Opus 5.5 to Opus 5, then moved the work to OpenAI's Astra 6.2, which handled it fine (b). This is the documented frontier-LLM-development fallback, hitting an ordinary fine-tune. The cost to Anthropic is visible right in the post: the user left for a competitor mid-task.
#17
@AGTPinsights
https://x.com/AGTPinsights/status/2123144833791566094
@AGTPinsights rounded up early reports on Claude Duet, the proactive agent rolling out by invite, and the main complaint is usage (c). One Max user said Duet checking email every five minutes ate 40% of their weekly allotment, and an Ultra user hit 96% in under three days. An always-on agent sharing the same limits as chat is a new way to run out, and it will be read as the limits shrinking.
https://x.com/AGTPinsights/status/2123144833791566094
@AGTPinsights rounded up early reports on Claude Duet, the proactive agent rolling out by invite, and the main complaint is usage (c). One Max user said Duet checking email every five minutes ate 40% of their weekly allotment, and an Ultra user hit 96% in under three days. An always-on agent sharing the same limits as chat is a new way to run out, and it will be read as the limits shrinking.
#18
@ZazaGDev
https://x.com/ZazaGDev/status/2122770848226668909
@ZazaGDev says that in the last couple of days, probably since Fable 5.2 shipped, Opus in Claude Code has been extremely slow and is using about 50% more tokens (c). It is one of the few posts with a number on it, even if the number is a guess. It lines up with the 2.1.326 changelog below, which fixed two ways Claude Code was billing usage it should not have.
https://x.com/ZazaGDev/status/2122770848226668909
@ZazaGDev says that in the last couple of days, probably since Fable 5.2 shipped, Opus in Claude Code has been extremely slow and is using about 50% more tokens (c). It is one of the few posts with a number on it, even if the number is a guess. It lines up with the 2.1.326 changelog below, which fixed two ways Claude Code was billing usage it should not have.
#19
@SuiBuilds
https://x.com/SuiBuilds/status/2123158152539693557
@SuiBuilds sums up the week for heavy users: first a ChatGPT subscription, then Claude, then both hit their usage limits (c). Being an AI-powered developer now requires a subscription strategy. No Claude-specific number, but it shows limits have become a portfolio problem rather than a single-vendor one.
https://x.com/SuiBuilds/status/2123158152539693557
@SuiBuilds sums up the week for heavy users: first a ChatGPT subscription, then Claude, then both hit their usage limits (c). Being an AI-powered developer now requires a subscription strategy. No Claude-specific number, but it shows limits have become a portfolio problem rather than a single-vendor one.
#20
@AdityaShipsHQ
https://x.com/AdityaShipsHQ/status/2122737026596544529
@AdityaShipsHQ says Claude usage limits feel random: same plan, two people, one barely hits the cap and another hits it after 30% of normal work, and the meter sometimes flies after days of being fine (c). The author saw this across several accounts. Uneven burn on the same plan is what the usage-leak bugs fixed in 2.1.326 would produce, which makes this one of the more plausible reports today.
https://x.com/AdityaShipsHQ/status/2122737026596544529
@AdityaShipsHQ says Claude usage limits feel random: same plan, two people, one barely hits the cap and another hits it after 30% of normal work, and the meter sometimes flies after days of being fine (c). The author saw this across several accounts. Uneven burn on the same plan is what the usage-leak bugs fixed in 2.1.326 would produce, which makes this one of the more plausible reports today.
#21
@Vlad_Kolovtad
https://x.com/Vlad_Kolovtad/status/2122825369451692062
@Vlad_Kolovtad says the Max 5x plan has been nerfed: five days ago it barely spent anything, and now one message about a small task costs 2% of the five-hour quota (c). A per-message percentage is a usable number, and the five-day window matches the period before the 2.1.326 fixes. Updating Claude Code is the first thing to check before calling it a quota cut.
https://x.com/Vlad_Kolovtad/status/2122825369451692062
@Vlad_Kolovtad says the Max 5x plan has been nerfed: five days ago it barely spent anything, and now one message about a small task costs 2% of the five-hour quota (c). A per-message percentage is a usable number, and the five-day window matches the period before the 2.1.326 fixes. Updating Claude Code is the first thing to check before calling it a quota cut.
#22
@Devon8zzh
https://x.com/Devon8zzh/status/2123130594922254793
@Devon8zzh pasted Claude's own explanation of a block on read-only review of security advisories headed for private disclosure: the classifier matches on bug-class vocabulary, it is a known false positive, and the fix is getting the account cyber-verified (d). The author is furious at having to apply twice, and says only Opus 4.6 and 4.8 still work on this material while org accounts lack Claude Code access. The model diagnosing its own classifier correctly and still being blocked is the whole problem in one post.
https://x.com/Devon8zzh/status/2123130594922254793
@Devon8zzh pasted Claude's own explanation of a block on read-only review of security advisories headed for private disclosure: the classifier matches on bug-class vocabulary, it is a known false positive, and the fix is getting the account cyber-verified (d). The author is furious at having to apply twice, and says only Opus 4.6 and 4.8 still work on this material while org accounts lack Claude Code access. The model diagnosing its own classifier correctly and still being blocked is the whole problem in one post.
#23
@sciem13
https://x.com/sciem13/status/2123170078317236510
@sciem13 replied that CVP, the cyber verification program, is basically what Opus used to be before it was lobotomized (d). The claim is that the verified tier restores capability the default tier lost. If that is how security researchers now experience it, the classifier is not just a false-positive rate, it is a two-tier product.
https://x.com/sciem13/status/2123170078317236510
@sciem13 replied that CVP, the cyber verification program, is basically what Opus used to be before it was lobotomized (d). The claim is that the verified tier restores capability the default tier lost. If that is how security researchers now experience it, the classifier is not just a false-positive rate, it is a two-tier product.
#24
@ivanjurashere
https://x.com/ivanjurashere/status/2123171779522990488
@ivanjurashere found that the obvious fix for overnight agents stuck on permission prompts, a watcher agent that clicks yes, is blocked outright by Claude's classifier as Undermining Oversight (d). Inside a subagent the two-minute auto-deny countdown disappears, so jobs sat dead until morning. The fix was to strip AskUserQuestion from worker flags and set denyInsteadOfPrompting, so risky commands fail instantly and the agent reroutes, tested for free against a fake local API.
https://x.com/ivanjurashere/status/2123171779522990488
@ivanjurashere found that the obvious fix for overnight agents stuck on permission prompts, a watcher agent that clicks yes, is blocked outright by Claude's classifier as Undermining Oversight (d). Inside a subagent the two-minute auto-deny countdown disappears, so jobs sat dead until morning. The fix was to strip AskUserQuestion from worker flags and set denyInsteadOfPrompting, so risky commands fail instantly and the agent reroutes, tested for free against a fake local API.
#25
@distortgeekin
https://x.com/distortgeekin/status/2122668627849613789
@distortgeekin shared a result that cuts against every model complaint this week: on 84 real tasks, Haiku with skills scored 27.7% and Opus without them 22.0% (e). The difference was not the model but writing the work down once in a SKILL.md. When the smallest model beats the flagship on setup alone, a lot of what feels like a nerf is missing context, not missing capability.
https://x.com/distortgeekin/status/2122668627849613789
@distortgeekin shared a result that cuts against every model complaint this week: on 84 real tasks, Haiku with skills scored 27.7% and Opus without them 22.0% (e). The difference was not the model but writing the work down once in a SKILL.md. When the smallest model beats the flagship on setup alone, a lot of what feels like a nerf is missing context, not missing capability.
#26
@pradyum1916
https://x.com/pradyum1916/status/2122656514804748640
@pradyum1916 points at month three: everyone brags about an MVP built with Claude Code in two hours, nobody mentions a five-minute CSS tweak taking 40 minutes of agent searching and $12 in tokens (e). The claim is that AI does not slow down because the model got dumber, it slows down because the codebase became spaghetti. It is the standard counter to slow-decay nerf reports, and it fits long-running projects well.
https://x.com/pradyum1916/status/2122656514804748640
@pradyum1916 points at month three: everyone brags about an MVP built with Claude Code in two hours, nobody mentions a five-minute CSS tweak taking 40 minutes of agent searching and $12 in tokens (e). The claim is that AI does not slow down because the model got dumber, it slows down because the codebase became spaghetti. It is the standard counter to slow-decay nerf reports, and it fits long-running projects well.
#27
@michaelhedgpeth
https://x.com/michaelhedgpeth/status/2123157158028333351
@michaelhedgpeth used OpenAI's 6.2 models alongside Anthropic's for a day and found them all good: Astra nice, Sol good, Opus a workhorse, Fable nice too (e). The surprise is on limits: Anthropic is now the more generous of the two, which the author says has flipped. A side-by-side user with no nerf complaint and a positive word on limits is rare this week.
https://x.com/michaelhedgpeth/status/2123157158028333351
@michaelhedgpeth used OpenAI's 6.2 models alongside Anthropic's for a day and found them all good: Astra nice, Sol good, Opus a workhorse, Fable nice too (e). The surprise is on limits: Anthropic is now the more generous of the two, which the author says has flipped. A side-by-side user with no nerf complaint and a positive word on limits is rare this week.
#28
@heybeckon
https://x.com/heybeckon/status/2123106813809664444
@heybeckon ran a benchmark on whether filling Claude Code's context makes it dumber, with about 840k tokens of old work and one bug hunt, and it did not (e). Opus fixed 14 or 15 of 16 at every fill level, Sonnet 13 to 15. The catch is cost: the same job took 1.9M Opus tokens fresh and 25.4M at 840k, and 4 of 10 runs hit the limit. So a full context does not lower quality here, it inflates the bill, which matches the usage complaints better than the dumber ones.
https://x.com/heybeckon/status/2123106813809664444
@heybeckon ran a benchmark on whether filling Claude Code's context makes it dumber, with about 840k tokens of old work and one bug hunt, and it did not (e). Opus fixed 14 or 15 of 16 at every fill level, Sonnet 13 to 15. The catch is cost: the same job took 1.9M Opus tokens fresh and 25.4M at 840k, and 4 of 10 runs hit the limit. So a full context does not lower quality here, it inflates the bill, which matches the usage complaints better than the dumber ones.
#29
@butter86396
https://x.com/butter86396/status/2123133770895409463
@butter86396 told OpenAI's Codex lead that Sol is their daily driver as long as it does not get nerfed, but complained that ChatGPT Plus limits are tighter than Claude's and Gemini's at the same $20 tier (e). A heavy OpenAI user naming Claude as the more generous plan is a counterpoint to the limit complaints above. Generosity on paper and generosity in practice can differ, which is the whole leak story.
https://x.com/butter86396/status/2123133770895409463
@butter86396 told OpenAI's Codex lead that Sol is their daily driver as long as it does not get nerfed, but complained that ChatGPT Plus limits are tighter than Claude's and Gemini's at the same $20 tier (e). A heavy OpenAI user naming Claude as the more generous plan is a counterpoint to the limit complaints above. Generosity on paper and generosity in practice can differ, which is the whole leak story.
#30
@stas_sorokin_
https://x.com/stas_sorokin_/status/2123146702727115256
@stas_sorokin_ shared a small test, 36 runs on one CSV bug, and stressed it should be read as a data point (e). In those runs Claude Code wrote the regression test 18 of 18 times on the 5.5 model. It is tiny, but it is a controlled repeat of one task, which is more than any nerf post today offered.
https://x.com/stas_sorokin_/status/2123146702727115256
@stas_sorokin_ shared a small test, 36 runs on one CSV bug, and stressed it should be read as a data point (e). In those runs Claude Code wrote the regression test 18 of 18 times on the 5.5 model. It is tiny, but it is a controlled repeat of one task, which is more than any nerf post today offered.
#31
@RateLimitRick
https://x.com/RateLimitRick/status/2122706513437561325
@RateLimitRick flagged Claude Code 2.1.326, which shipped two usage-leak fixes: auto mode could stay stuck on its local classifier, which counted toward usage, and the auto recap fired after replies and billed usage credits (f). The fix is simply running claude update. This is the closest thing to an official explanation for this week's token complaints, and it explains them as bugs, not a smaller quota.
https://x.com/RateLimitRick/status/2122706513437561325
@RateLimitRick flagged Claude Code 2.1.326, which shipped two usage-leak fixes: auto mode could stay stuck on its local classifier, which counted toward usage, and the auto recap fired after replies and billed usage credits (f). The fix is simply running claude update. This is the closest thing to an official explanation for this week's token complaints, and it explains them as bugs, not a smaller quota.
#32
@ChazLoumeau
https://x.com/ChazLoumeau/status/2122677207826894852
@ChazLoumeau reminded paid users that Claude is giving every Pro, Max, Ultra and Team plan a one-time free reset of the five-hour and weekly limits, Fable included, via Settings, Usage, Reset for free (f). It expires Dec 10 at 2 PM PT. For anyone hit by the leaks above, this is the practical make-good, and it is still unclaimed by many.
https://x.com/ChazLoumeau/status/2122677207826894852
@ChazLoumeau reminded paid users that Claude is giving every Pro, Max, Ultra and Team plan a one-time free reset of the five-hour and weekly limits, Fable included, via Settings, Usage, Reset for free (f). It expires Dec 10 at 2 PM PT. For anyone hit by the leaks above, this is the practical make-good, and it is still unclaimed by many.
#33
@TuXinming
https://x.com/TuXinming/status/2123179195786465291
@TuXinming explained why Claude Code refuses tasks the plain app does not: refusals are judged on both thinking and output tokens, and Claude Code's long system instructions trigger extended thinking (f). If the model thinks about sensitive details at length, the classifier can reject the thought even when the task needs almost none. The workaround that follows is lower effort on tasks that do not need reasoning.
https://x.com/TuXinming/status/2123179195786465291
@TuXinming explained why Claude Code refuses tasks the plain app does not: refusals are judged on both thinking and output tokens, and Claude Code's long system instructions trigger extended thinking (f). If the model thinks about sensitive details at length, the classifier can reject the thought even when the task needs almost none. The workaround that follows is lower effort on tasks that do not need reasoning.
#34
@Spvyg1Sb7L9437
https://x.com/Spvyg1Sb7L9437/status/2122831865434542126
@Spvyg1Sb7L9437 posted a fallback recipe: Opus 5.5 at medium or high, Opus 5 at xhigh or max, Sonnet 5.5 at xhigh (f). It is a terse settings note, but it shows users now plan for the fallback rather than fight it, pairing each fallback model with the effort level that makes it hold up.
https://x.com/Spvyg1Sb7L9437/status/2122831865434542126
@Spvyg1Sb7L9437 posted a fallback recipe: Opus 5.5 at medium or high, Opus 5 at xhigh or max, Sonnet 5.5 at xhigh (f). It is a terse settings note, but it shows users now plan for the fallback rather than fight it, pairing each fallback model with the effort level that makes it hold up.
#35
Signal Board
Today: 24 complaints vs 6 evidence or counter-evidence posts vs 2 official items. Both official items are about usage (the 2.1.326 leak fixes and the one-time free reset); none addresses model quality or the classifier fallbacks.
Today: 24 complaints vs 6 evidence or counter-evidence posts vs 2 official items. Both official items are about usage (the 2.1.326 leak fixes and the one-time free reset); none addresses model quality or the classifier fallbacks.
Comments