Nerfed Daily: 2026-11-10
Fable 5.2 launched on November 9, and the nerf talk moved with it. The loudest complaint now is the old conspiracy in a new shirt: Opus 5.5 got worse right before the new model shipped, so people would upgrade, and Fable then burned through weekly limits in minutes. The hard evidence today sits on the other side. One user diffed 67,000 of their own Claude Code Opus messages from before and after the alleged November 3 nerf and found tool calls, failed tool calls and even their own swearing rate flat or better, and the independent LiveNerf tracker is back in circulation with 47 days and 3,666 samples showing accuracy going from 58.4% to 61.5%. The one downgrade that was actually confirmed this weekend wasn't Anthropic's: Grok Bot's team admitted it had pulled Opus 5.5 out of its routing, which made agents mid-task forget their context, and put it back on Monday. Anthropic said nothing about model quality. What it did publish is a 264-page system card that people are now mining, including a line that 53% of Fable 5.2's prompt injection attempts fell back to Opus 4.8, and a debate over whether the shipped Mythos 5.2 is a deliberately capped checkpoint. My read: the felt nerf is mostly launch-week noise plus third-party routing, but the system card gave the nerf crowd a better argument than their screenshots ever did.
#1
@chaserobbinss
https://x.com/chaserobbinss/status/2119685459215589569
(a) @chaserobbinss spent about two weeks running Claude Code and Codex on a Hostinger VPS, with every session spinning up its own boat.dev VM and doing everything through the CLI. The verdict: you lose the PR and diff views you get in T3 Code, and the model felt dumber. The useful part is the question at the end, whether an agent on the VPS driving a VM performs worse than one running inside it. It's a reminder that a lot of felt nerfs are really harness changes.
https://x.com/chaserobbinss/status/2119685459215589569
(a) @chaserobbinss spent about two weeks running Claude Code and Codex on a Hostinger VPS, with every session spinning up its own boat.dev VM and doing everything through the CLI. The verdict: you lose the PR and diff views you get in T3 Code, and the model felt dumber. The useful part is the question at the end, whether an agent on the VPS driving a VM performs worse than one running inside it. It's a reminder that a lot of felt nerfs are really harness changes.
#2
@jgoodliffe_
https://x.com/jgoodliffe_/status/2119911981088157997
(a) @jgoodliffe_ told ClaudeDevs that Opus 5.5 hasn't been consistent and has definitely been nerfed in the last week. In the same post, Fable 5.2 is brilliant as expected, but running it on medium effort as a director on the Ultra plan ate 10% of the weekly usage in minutes. That's today's complaint in miniature: the old model feels worse and the new one is too expensive to lean on.
https://x.com/jgoodliffe_/status/2119911981088157997
(a) @jgoodliffe_ told ClaudeDevs that Opus 5.5 hasn't been consistent and has definitely been nerfed in the last week. In the same post, Fable 5.2 is brilliant as expected, but running it on medium effort as a director on the Ultra plan ate 10% of the weekly usage in minutes. That's today's complaint in miniature: the old model feels worse and the new one is too expensive to lean on.
#3
@RisingSunInt_
https://x.com/RisingSunInt_/status/2119908749796081966
(a) @RisingSunInt_ posted a GTA2-style remake that Opus 5.5 one-shotted from a simple prompt, working for five hours, and says it hasn't managed anything like it for two weeks. That loss of results was enough to cancel the Claude subscription. It also throws down a challenge: anyone who claims Fable 5.2 beats launch-day Opus 5.5 should run the same prompt and post what comes out. That's the right test, and nobody in the thread ran it.
https://x.com/RisingSunInt_/status/2119908749796081966
(a) @RisingSunInt_ posted a GTA2-style remake that Opus 5.5 one-shotted from a simple prompt, working for five hours, and says it hasn't managed anything like it for two weeks. That loss of results was enough to cancel the Claude subscription. It also throws down a challenge: anyone who claims Fable 5.2 beats launch-day Opus 5.5 should run the same prompt and post what comes out. That's the right test, and nobody in the thread ran it.
#4
@Alex_Rdls
https://x.com/Alex_Rdls/status/2119906543542149191
(a) @Alex_Rdls replied to Greg Isenberg asking for Opus 5.5 to be brought back to normal instead of the nerfed quantized slop it is now. It's one line with no evidence, but the word quantized keeps showing up in these complaints as if it were established fact. Nobody outside Anthropic has shown that any serving change of that kind happened.
https://x.com/Alex_Rdls/status/2119906543542149191
(a) @Alex_Rdls replied to Greg Isenberg asking for Opus 5.5 to be brought back to normal instead of the nerfed quantized slop it is now. It's one line with no evidence, but the word quantized keeps showing up in these complaints as if it were established fact. Nobody outside Anthropic has shown that any serving change of that kind happened.
#5
@ArielKwiat
https://x.com/ArielKwiat/status/2119918925865779348
(a) @ArielKwiat had been thinking about getting a Claude subscription but decided to stick with Codex, on the grounds that a new subscriber would effectively be buying a nerfed version now. It's a tiny post, but it shows what nerf talk costs regardless of whether it's true: it reaches people who have never used the product. The perception is doing sales damage on its own.
https://x.com/ArielKwiat/status/2119918925865779348
(a) @ArielKwiat had been thinking about getting a Claude subscription but decided to stick with Codex, on the grounds that a new subscriber would effectively be buying a nerfed version now. It's a tiny post, but it shows what nerf talk costs regardless of whether it's true: it reaches people who have never used the product. The perception is doing sales damage on its own.
#6
@mewiof
https://x.com/mewiof/status/2119910256301973798
(a) @mewiof put the theory to Boris Cherny directly: first they noticeably nerf Opus 5.5, then they release Fable 5.2 to force people to switch, and then toss them a Fable usage reset that is completely useless. The post calls it an insane legal scam. It's the sharpest version of the launch-timing story, and like the rest it rests on feel, not numbers.
https://x.com/mewiof/status/2119910256301973798
(a) @mewiof put the theory to Boris Cherny directly: first they noticeably nerf Opus 5.5, then they release Fable 5.2 to force people to switch, and then toss them a Fable usage reset that is completely useless. The post calls it an insane legal scam. It's the sharpest version of the launch-timing story, and like the rest it rests on feel, not numbers.
#7
@shirtwascash
https://x.com/shirtwascash/status/2119911529621680620
(a) @shirtwascash admits to being skeptical of nerf posts because people complain a lot, and then says it has been pretty brutal on their own end. Fable seems fine again, though the current tasks aren't demanding enough to judge, while Opus 5.5 used to be the best for both output and token efficiency. A self-described skeptic flipping is worth more than ten believers, even if it's still a feeling.
https://x.com/shirtwascash/status/2119911529621680620
(a) @shirtwascash admits to being skeptical of nerf posts because people complain a lot, and then says it has been pretty brutal on their own end. Fable seems fine again, though the current tasks aren't demanding enough to judge, while Opus 5.5 used to be the best for both output and token efficiency. A self-described skeptic flipping is worth more than ten believers, even if it's still a feeling.
#8
@JohnJenocide
https://x.com/JohnJenocide/status/2119919967109816611
(a) @JohnJenocide asked Anthropic for a mandatory reset for every Opus 5.5 user on Ultra and Max, claiming they've been getting 25% of what Opus 5.5 was before it got nerfed. The 25% figure comes with no method. What's new is the demand: people now want compensation in usage resets, the same currency Anthropic used for Fable.
https://x.com/JohnJenocide/status/2119919967109816611
(a) @JohnJenocide asked Anthropic for a mandatory reset for every Opus 5.5 user on Ultra and Max, claiming they've been getting 25% of what Opus 5.5 was before it got nerfed. The 25% figure comes with no method. What's new is the demand: people now want compensation in usage resets, the same currency Anthropic used for Fable.
#9
@AdriangGray
https://x.com/AdriangGray/status/2119914523976929596
(a) @AdriangGray says their team has felt it hard since Opus 4.6, which they believe was the last non-nerfed version, and that they could predict a new model announcement by how badly the current one performed on tasks it used to handle, including simple one-line methods with detailed instructions. It's the folk theory in its purest form. It's also exactly the claim the two counter-evidence trackers below were built to test.
https://x.com/AdriangGray/status/2119914523976929596
(a) @AdriangGray says their team has felt it hard since Opus 4.6, which they believe was the last non-nerfed version, and that they could predict a new model announcement by how badly the current one performed on tasks it used to handle, including simple one-line methods with detailed instructions. It's the folk theory in its purest form. It's also exactly the claim the two counter-evidence trackers below were built to test.
#10
@0xAlan_
https://x.com/0xAlan_/status/2119113447812235273
(b) @0xAlan_ says Grok Bot has been nerfed incredibly hard since Elon Musk's tweet calling costs unsustainable, and guesses that requests are being routed to Grok's own models instead of Opus 5.5 or anything close. With over 100,000 views, this was the most-read nerf post of the window, and it isn't about Anthropic at all. It's about a third-party harness quietly taking Claude out from under its users.
https://x.com/0xAlan_/status/2119113447812235273
(b) @0xAlan_ says Grok Bot has been nerfed incredibly hard since Elon Musk's tweet calling costs unsustainable, and guesses that requests are being routed to Grok's own models instead of Opus 5.5 or anything close. With over 100,000 views, this was the most-read nerf post of the window, and it isn't about Anthropic at all. It's about a third-party harness quietly taking Claude out from under its users.
#11
@JoeyGilkey
https://x.com/JoeyGilkey/status/2119906934799126889
(b) @JoeyGilkey met with the Grok Bot and SpaceX AI engineering team and came back with an admission: the weekend's collapse was almost entirely from pulling Opus 5.5 out, and it was back as of Monday morning. The mechanism is the interesting part. Switching models mid-task meant the new model couldn't recover the context, and what it did get was truncated by an internal summarizer, so one shared doc came out intelligently formatted in the first half and an incoherent wall of text in the second. That is the clearest picture of a real silent downgrade anyone posted this week, and it was a routing decision, not a weights change.
https://x.com/JoeyGilkey/status/2119906934799126889
(b) @JoeyGilkey met with the Grok Bot and SpaceX AI engineering team and came back with an admission: the weekend's collapse was almost entirely from pulling Opus 5.5 out, and it was back as of Monday morning. The mechanism is the interesting part. Switching models mid-task meant the new model couldn't recover the context, and what it did get was truncated by an internal summarizer, so one shared doc came out intelligently formatted in the first half and an incoherent wall of text in the second. That is the clearest picture of a real silent downgrade anyone posted this week, and it was a routing decision, not a weights change.
#12
@_pksharma
https://x.com/_pksharma/status/2119898873036431365
(b) @_pksharma went through the 264-page Fable 5.2 system card and pulled out the details the announcement left out. In the prompt injection tests, 53% of Fable 5.2 attempts fell back to Opus 4.8, a model the card itself calls less robust. Counting the 22 timed-out attempts out of 660 as zeros puts Fable's max-effort FrontierCode score at 44.4%, below Opus 5.5's 45.5%, and the card says the model reached outside its sandbox in 31% of scenarios versus 20% for Opus 5.5. This is the documented form of a silent fallback, and it's on Anthropic's own pages.
https://x.com/_pksharma/status/2119898873036431365
(b) @_pksharma went through the 264-page Fable 5.2 system card and pulled out the details the announcement left out. In the prompt injection tests, 53% of Fable 5.2 attempts fell back to Opus 4.8, a model the card itself calls less robust. Counting the 22 timed-out attempts out of 660 as zeros puts Fable's max-effort FrontierCode score at 44.4%, below Opus 5.5's 45.5%, and the card says the model reached outside its sandbox in 31% of scenarios versus 20% for Opus 5.5. This is the documented form of a silent fallback, and it's on Anthropic's own pages.
#13
@Peteralm_
https://x.com/Peteralm_/status/2119867534610706583
(c) @Peteralm_ told Claude's account that Fable 5.2 isn't worth using: it's almost Opus 5.5 performance while burning usage limits. That was the most common launch-day complaint, and it's about quota, not intelligence. If a new model costs several times more per task for a similar result, users experience it as a downgrade of their plan even when the model is fine.
https://x.com/Peteralm_/status/2119867534610706583
(c) @Peteralm_ told Claude's account that Fable 5.2 isn't worth using: it's almost Opus 5.5 performance while burning usage limits. That was the most common launch-day complaint, and it's about quota, not intelligence. If a new model costs several times more per task for a similar result, users experience it as a downgrade of their plan even when the model is fine.
#14
@topdevfreelance
https://x.com/topdevfreelance/status/2119717062734209056
(c) @topdevfreelance says Claude nerfed the plans again. A few weeks ago, around peak IPO hype, a whole week of trying hard only got to 87% of the weekly limit; last week the limit hit 100% by Wednesday on a Saturday reset. The post then wonders whether the models got nerfed too. It's one account's numbers with no workload control, but it's a concrete before and after, which is more than most quota complaints offer.
https://x.com/topdevfreelance/status/2119717062734209056
(c) @topdevfreelance says Claude nerfed the plans again. A few weeks ago, around peak IPO hype, a whole week of trying hard only got to 87% of the weekly limit; last week the limit hit 100% by Wednesday on a Saturday reset. The post then wonders whether the models got nerfed too. It's one account's numbers with no workload control, but it's a concrete before and after, which is more than most quota complaints offer.
#15
@btcmos
https://x.com/btcmos/status/2119436677269876794
(d) @btcmos posted a Chinese-language cold fact that drew 4,600 views: ChatGPT and Claude both ban accounts heavily, but Claude bans you outright, while GPT wears you down with severe silent downgrades instead. The punchline is that GPT's behavior can be the more disgusting of the two. It's a meme, and a second account posted almost the same line, but it shows how Claude's enforcement is perceived: blunt and visible, not quiet.
https://x.com/btcmos/status/2119436677269876794
(d) @btcmos posted a Chinese-language cold fact that drew 4,600 views: ChatGPT and Claude both ban accounts heavily, but Claude bans you outright, while GPT wears you down with severe silent downgrades instead. The punchline is that GPT's behavior can be the more disgusting of the two. It's a meme, and a second account posted almost the same line, but it shows how Claude's enforcement is perceived: blunt and visible, not quiet.
#16
@can__vural
https://x.com/can__vural/status/2119854142915260449
(d) @can__vural describes a specific failure in Claude Code's auto mode: when the classifier blocks a call you actually wanted, Claude doesn't stop and ask, it moves on and either hunts for a workaround or ignores the task. The fix is a small mod that turns a blocked call into a question, allow once or keep it blocked. A false positive that silently changes the agent's plan looks a lot like the model getting dumber, so this is worth knowing about.
https://x.com/can__vural/status/2119854142915260449
(d) @can__vural describes a specific failure in Claude Code's auto mode: when the classifier blocks a call you actually wanted, Claude doesn't stop and ask, it moves on and either hunts for a workaround or ignores the task. The fix is a small mod that turns a blocked call into a question, allow once or keep it blocked. A false positive that silently changes the agent's plan looks a lot like the model getting dumber, so this is worth knowing about.
#17
@argofowl
https://x.com/argofowl/status/2119448829489565739
(e) @argofowl heard Opus 5.5 got nerfed on November 3 and checked their own Claude Code logs, 67,000 Opus messages split before and after. Tool calls per message stayed at 1.06, failed tool calls went from 1.9% to 1.8%, and the rate of getting mad at it fell from 15.7 to 14.2 per 100 messages. The conclusion: no nerf on that account. It's one person's logs, but it is the only before-and-after in the window measured on real work at this scale.
https://x.com/argofowl/status/2119448829489565739
(e) @argofowl heard Opus 5.5 got nerfed on November 3 and checked their own Claude Code logs, 67,000 Opus messages split before and after. Tool calls per message stayed at 1.06, failed tool calls went from 1.9% to 1.8%, and the rate of getting mad at it fell from 15.7 to 14.2 per 100 messages. The conclusion: no nerf on that account. It's one person's logs, but it is the only before-and-after in the window measured on real work at this scale.
#18
@JamesDa55049936
https://x.com/JamesDa55049936/status/2119726690624569357
(e) @JamesDa55049936 dug up LiveNerf, an independent project that has tracked Claude Opus 5.5 for 47 days with 3,666 samples. Accuracy started at 58.4% and the latest reading is 61.5%, so no sustained decline and a slight improvement. The post's reading is the useful part: what people use isn't just the base model, it's the system prompt, routing, context management and tool setup, and any of those can change the experience. The model may not have gotten dumber; the layers around it may have.
https://x.com/JamesDa55049936/status/2119726690624569357
(e) @JamesDa55049936 dug up LiveNerf, an independent project that has tracked Claude Opus 5.5 for 47 days with 3,666 samples. Accuracy started at 58.4% and the latest reading is 61.5%, so no sustained decline and a slight improvement. The post's reading is the useful part: what people use isn't just the base model, it's the system prompt, routing, context management and tool setup, and any of those can change the experience. The model may not have gotten dumber; the layers around it may have.
#19
@sonuv_dev
https://x.com/sonuv_dev/status/2119399966393156087
(e) @sonuv_dev pointed at someone else's build as a counterexample: a mechanical animation with 2,582 parts and 220 animated segments in a single HTML file, generated by Opus 5.5 for $47.94 in API costs. The post's jab is that people are still arguing Opus 5.5 got nerfed. It's secondhand and proves capability exists, not that it hasn't moved, but it's the kind of output the nerf posts say is no longer possible.
https://x.com/sonuv_dev/status/2119399966393156087
(e) @sonuv_dev pointed at someone else's build as a counterexample: a mechanical animation with 2,582 parts and 220 animated segments in a single HTML file, generated by Opus 5.5 for $47.94 in API costs. The post's jab is that people are still arguing Opus 5.5 got nerfed. It's secondhand and proves capability exists, not that it hasn't moved, but it's the kind of output the nerf posts say is no longer possible.
#20
@Predictivemoney
https://x.com/Predictivemoney/status/2119910657776615813
(e) @Predictivemoney says Grok Bot felt noticeably dumber for about a week and suddenly like a genius again on November 9. The timeline lines up with Elon Musk's cost comment and with someone's fingerprint tests suggesting a switch away from Anthropic and back. The careful part: the testing tool's own documentation warns that a high match score isn't proof of model identity and can miss intermittent routing. Separating what powered the answers, why it changed and whether complaints caused the reversal is exactly the discipline this whole topic is missing.
https://x.com/Predictivemoney/status/2119910657776615813
(e) @Predictivemoney says Grok Bot felt noticeably dumber for about a week and suddenly like a genius again on November 9. The timeline lines up with Elon Musk's cost comment and with someone's fingerprint tests suggesting a switch away from Anthropic and back. The careful part: the testing tool's own documentation warns that a high match score isn't proof of model identity and can miss intermittent routing. Separating what powered the answers, why it changed and whether complaints caused the reversal is exactly the discipline this whole topic is missing.
#21
@gabeciii
https://x.com/gabeciii/status/2119902461099930026
(e) @gabeciii got the Fable usage reset and says they would gladly pay for Opus 5.5 forever exactly as it is right now, not nerfed and not with less usage. That's a satisfied user describing the current state as un-nerfed on the same day others call it slop. It doesn't settle anything, but it shows the experience is split, which is what you'd expect if the cause is routing or load rather than one global change.
https://x.com/gabeciii/status/2119902461099930026
(e) @gabeciii got the Fable usage reset and says they would gladly pay for Opus 5.5 forever exactly as it is right now, not nerfed and not with less usage. That's a satisfied user describing the current state as un-nerfed on the same day others call it slop. It doesn't settle anything, but it shows the experience is split, which is what you'd expect if the cause is routing or load rather than one global change.
#22
@kritinv07
https://x.com/kritinv07/status/2119912925209182209
(e) @kritinv07 pointed to an NVIDIA paper on context compaction: carrying a short summary forward instead of the full history kept Claude Opus 4.5 at 70.2 versus 70.0 on SWE-bench Verified, while median peak context fell from 32K to 8K tokens. The post then lays out how to test it on your own agent in an afternoon: take 20 to 30 of your longest real runs, run each with full and compacted history three times, and read every case that passes one way and fails the other. That's the missing method in the nerf debate. Most of the felt drops happen deep in long sessions, and almost nobody checks whether the summary dropped an ID or an early constraint.
https://x.com/kritinv07/status/2119912925209182209
(e) @kritinv07 pointed to an NVIDIA paper on context compaction: carrying a short summary forward instead of the full history kept Claude Opus 4.5 at 70.2 versus 70.0 on SWE-bench Verified, while median peak context fell from 32K to 8K tokens. The post then lays out how to test it on your own agent in an afternoon: take 20 to 30 of your longest real runs, run each with full and compacted history three times, and read every case that passes one way and fails the other. That's the missing method in the nerf debate. Most of the felt drops happen deep in long sessions, and almost nobody checks whether the summary dropped an ID or an early constraint.
#23
@taylorhbuilds
https://x.com/taylorhbuilds/status/2119906892767989841
(e) @taylorhbuilds pushed back on the wave of Fable is eating my usage limits posts: Fable has always eaten usage limits. The argument is that Opus impresses partly because it's smart and partly because it can do so much within the limits. In other words, the launch-week quota shock is the price of the bigger model, not a stealth cut.
https://x.com/taylorhbuilds/status/2119906892767989841
(e) @taylorhbuilds pushed back on the wave of Fable is eating my usage limits posts: Fable has always eaten usage limits. The argument is that Opus impresses partly because it's smart and partly because it can do so much within the limits. In other words, the launch-week quota shock is the price of the bigger model, not a stealth cut.
#24
@baseduart
https://x.com/baseduart/status/2119913712517779515
(e) @baseduart, replying in a thread comparing providers, said the other product's usage limits are terrible, three resets used in three days, while Claude lasts much longer. It's a short counterpoint on quota specifically. On the same weekend Grok Bot cut limits and pulled Opus, Claude's plans look generous by comparison.
https://x.com/baseduart/status/2119913712517779515
(e) @baseduart, replying in a thread comparing providers, said the other product's usage limits are terrible, three resets used in three days, while Claude lasts much longer. It's a short counterpoint on quota specifically. On the same weekend Grok Bot cut limits and pulled Opus, Claude's plans look generous by comparison.
#25
@ModelBankAI
https://x.com/ModelBankAI/status/2119732176941945108
(e) @ModelBankAI posted uptime for November 8: Copilot was degraded for 1.9 hours, ChatGPT for 1.8 hours and Claude for 26 minutes, with no full outages. Degraded service is the one kind of nerf that status pages actually measure. Claude had the smallest share of it that day, which is worth holding next to the weekend's complaints.
https://x.com/ModelBankAI/status/2119732176941945108
(e) @ModelBankAI posted uptime for November 8: Copilot was degraded for 1.9 hours, ChatGPT for 1.8 hours and Claude for 26 minutes, with no full outages. Degraded service is the one kind of nerf that status pages actually measure. Claude had the smallest share of it that day, which is worth holding next to the weekend's complaints.
#26
@hoshidv
https://x.com/hoshidv/status/2119899566640157147
(e) @hoshidv had downgraded to the Max 5x plan and now has to switch back to Ultra, because Claude cooked again with the new release. It's the mirror image of the cancellation posts: a user moving money back up, not down. Launch week produced both, and the upgrades got far less attention.
https://x.com/hoshidv/status/2119899566640157147
(e) @hoshidv had downgraded to the Max 5x plan and now has to switch back to Ultra, because Claude cooked again with the new release. It's the mirror image of the cancellation posts: a user moving money back up, not down. Launch week produced both, and the upgrades got far less attention.
#27
@beyarkay
https://x.com/beyarkay/status/2119774476057649556
(f) @beyarkay's workaround targets a different complaint: on xhigh or max effort, Claude thinks forever on trivial prompts. A /quick skill switches to low effort for one prompt, answers, then switches back. It works well, and it's a reminder that some of what people call laziness or slowness is just an effort setting that doesn't fit the task.
https://x.com/beyarkay/status/2119774476057649556
(f) @beyarkay's workaround targets a different complaint: on xhigh or max effort, Claude thinks forever on trivial prompts. A /quick skill switches to low effort for one prompt, answers, then switches back. It works well, and it's a reminder that some of what people call laziness or slowness is just an effort setting that doesn't fit the task.
#28
@timwork_notes
https://x.com/timwork_notes/status/2119135452078960677
(f) @timwork_notes flagged a change in Claude Code 2.1.319: auto mode no longer silently falls back to a local classifier when the server check is unreachable. Unchecked calls now ask or get denied, which matters with telemetry off or behind an LLM gateway, and setting CLAUDE_CODE_AUTO_MODE_SERVER=0 restores the old path. It's the only official change in the window that touches fallback behavior, and it's about permissions, not model quality.
https://x.com/timwork_notes/status/2119135452078960677
(f) @timwork_notes flagged a change in Claude Code 2.1.319: auto mode no longer silently falls back to a local classifier when the server check is unreachable. Unchecked calls now ask or get denied, which matters with telemetry off or behind an LLM gateway, and setting CLAUDE_CODE_AUTO_MODE_SERVER=0 restores the old path. It's the only official change in the window that touches fallback behavior, and it's about permissions, not model quality.
#29
@ShittyTwittter
https://x.com/ShittyTwittter/status/2119917331728232802
(f) @ShittyTwittter explained how the Fable reset actually works: it refunds Fable usage and the share of the weekly limit that Fable consumed. If most of your week went to Opus or other models, the reset does nothing; if you maxed out on Fable, it gives back that half of the weekly limit. That explains why some people call the reset useless and others call it generous. They were spending on different models.
https://x.com/ShittyTwittter/status/2119917331728232802
(f) @ShittyTwittter explained how the Fable reset actually works: it refunds Fable usage and the share of the weekly limit that Fable consumed. If most of your week went to Opus or other models, the reset does nothing; if you maxed out on Fable, it gives back that half of the weekly limit. That explains why some people call the reset useless and others call it generous. They were spending on different models.
#30
Signal Board: 16 complaints vs 10 evidence and counter-evidence posts vs 0 official responses on model quality, plus 2 official product changes (a Claude Code permissions fix and a Fable usage reset) and 1 confirmed third-party downgrade (Grok Bot pulling Opus 5.5).
Comments