November 17, 2026nerfed

Nerfed Daily: 2026-11-17

The nerf complaints this window are loud but thin. Opus 5.5 inside Claude Code takes most of the hits again: the same mistake repeated after every correction, style files suddenly ignored, a whole week where it feels dumber. The new twist is that Fable 5.2 is no longer the escape hatch. Several people now say Fable feels lazier or misses details that Opus just got, which flips last month's story. There is still no hard evidence of a model swap. The best number of the window cuts the other way: one user counted 9 status incidents between Nov 8 and Nov 16, and only 3 touched the models, the longest lasting 73 minutes. Most of what broke was sign-in, calls, billing and connectors. Anthropic did respond, but on usage, not quality: it admitted bugs had been eating usage limits faster than intended and handed paid plans one free reset to claim by December 10. Nothing official addresses whether the models got worse.
πŸ’‘#1
@cwww2332
https://x.com/cwww2332/status/2121514387664232669
@cwww2332 posted the window's most-shared nerf joke, written for Chinese users: the wrong way into AI goes from Doubao to Kimi and GLM, then a lot of fiddling with networks and accounts to reach Claude and ChatGPT, then getting nerfed and banned, then finally paying for Gemini Pro (a). The right way, the punchline says, is to buy Gemini Pro on day one. There is no test behind it. It matters as a mood reading: for users who reach Claude through workarounds, being nerfed and being banned now sit in the same sentence as one risk.
πŸ’‘#2
@trikcode
https://x.com/trikcode/status/2122325674069541339
@trikcode pairs two facts into one jab: Anthropic says over 90% of the code merged into its own codebase is written by Claude, so who ships the patches in a week when Opus 5.5 gets dumber (a)? It assumes the dip is real without showing one. The joke works because it points at a real dependency. If the lab's own engineering runs on the same model users complain about, a quality dip would hurt Anthropic first, which is a decent argument that nobody is degrading it on purpose.
πŸ’‘#3
@i_am_vickyd
https://x.com/i_am_vickyd/status/2122409793159660002
@i_am_vickyd asks outright whether Opus 5.5 has been nerfed (a). The complaint is specific in shape: the task and the goal were spelled out, and every time it is told to go again it makes the same mistakes, gets corrected, and repeats them. That pattern, corrections that do not stick inside one session, is the most common way people describe a nerf in Claude Code. It is also what a bloated or compacted context looks like, so a fresh session is the first thing to try.
πŸ’‘#4
@scottstts
https://x.com/scottstts/status/2122413652124762129
@scottstts replies that even with the fixes, Fable 5.2 is still much lazier than Opus 5.5 (a). Short, but it is one of several posts this window that reverse last month's ranking, when Fable was the model people fled to when Opus felt off. Lazier here usually means shorter answers and more stopping to ask. That can be a tuning choice rather than a capability drop, but users experience both the same way.
πŸ’‘#5
@grbellar
https://x.com/grbellar/status/2122345086600020203
@grbellar asks whether it is just them: Fable 5.2 feels dumber than Opus 5.5 and keeps missing small details that Opus just got (a). Missing small details is the vaguest kind of nerf report and the hardest to test. Paired with the lazier reports, though, it points the same way: this week the complaint is not that one model dropped, it is that the newer, pricier model is not reliably better than the older one.
πŸ’‘#6
@ugurbuilds
https://x.com/ugurbuilds/status/2122172759652327471
@ugurbuilds says Claude has been nerfed so hard that Fable 5.2 on extra effort does not match what the first Opus did on medium (a). This is the nostalgia version of the complaint: the comparison is against memory, not a saved run. Without the same prompt run then and now it cannot be checked, which is exactly why the tracker posts further down matter more than this one.
πŸ’‘#7
@jamesj_cottrell
https://x.com/jamesj_cottrell/status/2122429881761828885
@jamesj_cottrell makes the broadest claim of the window: Fable 5.2, Opus 5.5, GPT and Grok were all beaten on solution quality, not just cost, by Qwen 3.8 27b on a real task (a). The task was debugging an expensive data model, and a whole day went by without a line of code because every frontier model kept pushing indexes it did not need. The author calls it stickiness: models trained to ace closed benchmarks hold onto their own plan in open problems. It is one task, not a test suite, but it is a concrete one.
πŸ’‘#8
@MatthieuTalbot
https://x.com/MatthieuTalbot/status/2122263227035767133
@MatthieuTalbot asks ClaudeDevs directly whether Opus 5.5 has been quantized, because it suddenly cannot follow style rules or the lines in a SOUL.md file, almost like Sonnet (a). Instruction files being ignored is a sharper symptom than feeling dumber, because it can be checked: the same file, the same task, before and after. No reply from Anthropic in the window.
πŸ’‘#9
@Super_Nius
https://x.com/Super_Nius/status/2122457181979054430
@Super_Nius asks why every model gets nerfed and slowed down the longer it has been out: brilliant at launch, then output limits creep in (a). Sol 6.1 slowing down did not bother this user, but Opus 5.5, the favorite, getting dumber is unacceptable. The post is pure feeling, no test. It shows the launch-then-decline story is now a default belief, applied to every model on schedule.
πŸ’‘#10
@24thIDRefugee
https://x.com/24thIDRefugee/status/2122455652039880813
@24thIDRefugee says Claude Duet is dumb, probably dumber than Grok Bots at this point, and dates it to a big nerf a few weeks ago that ended a stretch of being super impressed (a). The ask is for two Ultra subscriptions, not more assistant products. Duet is a product surface, not a model, so this is as much about what Anthropic wraps around the model as about the model itself.
πŸ’‘#11
@exitcode0dev
https://x.com/exitcode0dev/status/2122412289974510048
@exitcode0dev says the Fable 5 launch moment never came back: the first three days were the best, then it was nerfed like crazy (a). The same three-days-of-magic story gets told about almost every launch. One plausible reading is that early users are on light servers with generous limits, and the model feels different once load and limits catch up, without any change to the weights.
πŸ’‘#12
@gurby1
https://x.com/gurby1/status/2122163906403508696
@gurby1 says Opus 5.5 has seriously degraded and even decided to bump the account down to Opus 4.8 (b). That second part is the checkable claim: a visible switch to an older model, not a feeling. The post blames it on talking to itself endlessly and says Grok Bot and Muse are better now. No screenshot of the switch, so it cannot be told apart from the classifier fallback or a usage cap.
πŸ’‘#13
@scottcolsen
https://x.com/scottcolsen/status/2121594200676241555
@scottcolsen noticed something odd about limits: Fable was maxed out for the week, but advising in Claude Code kept working (c). The question is whether advisor calls count against usage at all. If they do not, it is a quiet gift; if they do and the meter is wrong, it is the same kind of bug Anthropic admitted to this week. Either way, nobody outside Anthropic can tell from the meter what is being charged.
πŸ’‘#14
@Kasramx
https://x.com/Kasramx/status/2122348571530068391
@Kasramx relays reports that Claude Code burns unusually many tokens on Windows, with some users saying their limits run out much faster than on macOS (c). No numbers attached. It fits the week's official admission that bugs had been draining usage faster than intended, but nothing ties the Windows reports to those specific bugs.
πŸ’‘#15
@biswajitnk7
https://x.com/biswajitnk7/status/2122193063175225347
@biswajitnk7 hit the usage limits on both Claude and Codex by Monday and declared it a no-coding day, switching to writing posts instead (c). It is a small post, but it shows the practical version of the limits complaint: running two subscriptions as a fallback does not help when both meters run out in the same week.
πŸ’‘#16
@iykshani
https://x.com/iykshani/status/2122445043860693088
@iykshani finds Fable 5.2 and Opus 5.5 frustrating to use these days because the classifier keeps rejecting normal chats (d). No example of a blocked prompt is given. It belongs with the steady stream of false-positive reports; Anthropic's own guide, further down, admits these still happen.
πŸ’‘#17
@marcospaulosd
https://x.com/marcospaulosd/status/2122439523657515497
@marcospaulosd has a workaround for what the post calls Anthropic's classifier hell: run claudex -r, pick the Claude Code session, and continue it with the same tools, CLAUDE.md, hooks and skills on a different model (d). The same move is used for a second opinion without re-explaining the task. It is a sign of where users are heading: when a block hits, they move the session rather than rephrase.
πŸ’‘#18
@rimtoln
https://x.com/rimtoln/status/2122370200683135362
@rimtoln did the counting nobody else did: from Nov 8 to Nov 16 the Anthropic status page logged 9 incidents, and only 3 were about the models, the longest of those lasting 73 minutes (e). The rest were calls, sign-in, billing and connectors, including Sign in with Apple failing for some new iOS users for 56 hours. Ninety-day uptime reads 99.76% for the API and 99.46% for claude ai. This is the best evidence of the window, and it says the frequent breakage is around the model, not in it.
πŸ’‘#19
@etymologiae33
https://x.com/etymologiae33/status/2122450014165246065
@etymologiae33 reports the opposite of the week's main complaint: great results with Opus 5.5, quite poor ones with Fable 5.2 (e). That is counter-evidence on Opus and an extra complaint on Fable at once. It matches the pattern across the window, that whichever model a person has tuned their workflow around feels like the good one.
πŸ’‘#20
@tomvorel45
https://x.com/tomvorel45/status/2122456757725266172
@tomvorel45 is disappointed in xAI's Ultra plan and says Claude Max gives more usage and a much better model for half the price (e). It is a complaint about a competitor that doubles as a data point for Claude: in a direct comparison of paid plans, Claude's limits came out ahead. That rarely shows up in a week full of limits complaints.
πŸ’‘#21
@devopsbulletin
https://x.com/devopsbulletin/status/2122045746128044510
@devopsbulletin relays the window's one real admission from Anthropic: bugs had been consuming more of users' usage limits than intended, the issues are fixed, and Pro, Max, Ultra and Team users get one free reset across all models and limits (f). It has to be claimed in Settings > Usage by December 10, 2pm PT. This is the first time in weeks the official side has confirmed that part of the limits pain was a defect, not tighter rules.
πŸ’‘#22
@Ishant23April
https://x.com/Ishant23April/status/2122449429886292133
@Ishant23April flags a quiet drain in the new Projects feature: new projects run every thread on Opus at high effort, which Anthropic's own docs say uses a plan up fastest, and threads that hit the limit auto-continue at reset, eating the next window too (f). The fix is in Project settings > General: lower the thread model or effort, and turn off auto-continue. Some limits complaints this week are probably this default, not the model.
πŸ’‘#23
@cryptojezuz
https://x.com/cryptojezuz/status/2122454705741959350
@cryptojezuz summarizes Claude Code 2.1.326, and one line matters here: auto mode no longer gets stuck on its local classifier, which had been counting toward usage (f). The new /status row also shows when a provider setting or a variable like DISABLE_TELEMETRY switches features off. Both chip at the guesswork behind nerf reports, where a feature quietly off can feel like the model getting worse.
πŸ’‘#24
@ST_Automation
https://x.com/ST_Automation/status/2121980025109205426
@ST_Automation counted three fake accounts in one day, spelled CIaude with a capital I, all pushing the same bait: reset your Claude usage limits with one PowerShell line (f). The limits pain is now big enough to be a phishing hook. The real reset this week is claimed in Settings > Usage; anything asking you to paste a command is not it.
πŸ’‘#25
@park_danie36173
https://x.com/park_danie36173/status/2122386379506397567
@park_danie36173 points to Anthropic's Fable 5.1 guide, which says safety-classifier false positives have dropped but still happen (f). A blocked request comes back with stop_reason set to refusal, and the guide names three situations that make blocks likelier. That is the closest thing to an official answer to the classifier complaints above: they are acknowledged, measured as falling, and not zero.
πŸ’‘#26
@divyam_sirswal
https://x.com/divyam_sirswal/status/2121568991001161953
@divyam_sirswal relayed the Nov 14 status page entry: partially degraded service at Anthropic for about four hours (f). It is the kind of official signal that tends to get read as a nerf after the fact. When the status page says degraded for hours, worse answers in that window are an outage, not a new model.
πŸ’‘#27
@p2c2kbf
https://x.com/p2c2kbf/status/2122237317956219247
@p2c2kbf posted the Nov 16 status alert: multiple Anthropic services affected at once, the Claude API, claude ai and Claude Code (f). It landed the same day as many of the Opus 5.5 complaints above. Anyone who felt Claude Code was off on Monday should check that incident window before calling it a nerf.
πŸ’‘#28
Signal Board
Today: 17 complaints vs 3 evidence or counter-evidence posts vs 5 official items from Anthropic. The one official admission is about usage limits being drained by bugs; none addresses model quality.
← Previous
AI for AI Daily: 2026-11-17
← Back to all articles

Comments

Loading...
>_