Nerfed Daily: 2026-11-15
The nerf talk this window is about Opus 5.5 again, mostly inside Claude Code and the Claude app. People say it misses instructions, leaves code half finished, can't do a rebase and writes worse copy than Opus 4.5 did, and several now claim it got worse right after Fable 5.2 shipped. The second complaint is usage: Max and Ultra users say their limits drain faster than before, and nobody can tell what one message costs. Hard evidence is still thin. The only number on the complaint side is one benchmark run where Opus 5.5 at high effort fell back to an older Opus on about a fifth of its replies, while Fable 5.2 never fell back. The loudest counter-evidence is a public retest where Fable 5.2 beat Opus 5.5 on the same prompt a day after its author wondered if Fable had been nerfed. Anthropic said nothing about quality. What it did ship were docs and release notes on Code Mode, auto-mode classifier inputs and the Vertex fast-mode fallback, plus a reset that expires on December 10, which is easy to read as a coupon instead of an answer.
#1
@ayesha_fatiima
https://x.com/ayesha_fatiima/status/2121520986785013916
@ayesha_fatiima says Opus 5.5 is getting worse at the one thing it is supposed to be best at: coding (a). The list is concrete, with coding mistakes, incomplete code, missed instructions and half-finished outputs. The sharpest line is about value, not speed: what is the point of paying for the best model if half your time goes to fixing its work? The post asks whether anyone else sees it, and it drew the most engagement of any first-hand complaint this window.
https://x.com/ayesha_fatiima/status/2121520986785013916
@ayesha_fatiima says Opus 5.5 is getting worse at the one thing it is supposed to be best at: coding (a). The list is concrete, with coding mistakes, incomplete code, missed instructions and half-finished outputs. The sharpest line is about value, not speed: what is the point of paying for the best model if half your time goes to fixing its work? The post asks whether anyone else sees it, and it drew the most engagement of any first-hand complaint this window.
#2
@Frtnel
https://x.com/Frtnel/status/2121680529003336066
@Frtnel used to be in awe of Opus for everything from video editing to coding, and liked that it did not burn many tokens doing it (a). The claim now is simple: after Fable 5.2 arrived, Anthropic degraded Opus. There is no test attached, just the timing. That timing theory, that the old flagship gets worse once the new one lands, is the most repeated story in this window.
https://x.com/Frtnel/status/2121680529003336066
@Frtnel used to be in awe of Opus for everything from video editing to coding, and liked that it did not burn many tokens doing it (a). The claim now is simple: after Fable 5.2 arrived, Anthropic degraded Opus. There is no test attached, just the timing. That timing theory, that the old flagship gets worse once the new one lands, is the most repeated story in this window.
#3
@kvistgaard
https://x.com/kvistgaard/status/2121672384889442313
@kvistgaard posted a screenshot of what a typical exchange with Claude looks like lately and summed it up as first Opus 5.5 got nerfed, and now this (a). It is a short post, but it shows the mood shift: the nerf is no longer argued, it is treated as settled background, and each new annoyance is stacked on top of it.
https://x.com/kvistgaard/status/2121672384889442313
@kvistgaard posted a screenshot of what a typical exchange with Claude looks like lately and summed it up as first Opus 5.5 got nerfed, and now this (a). It is a short post, but it shows the mood shift: the nerf is no longer argued, it is treated as settled background, and each new annoyance is stacked on top of it.
#4
@Octoberry
https://x.com/Octoberry/status/2121682690340450338
@Octoberry, replying in the same thread, agrees that Opus is actually nerfed, then adds the twist that it is still far better than Fable on cost efficiency (a). That is a useful detail. Even people who believe the downgrade keep Opus as their default, because Fable's price per task is the bigger pain for them.
https://x.com/Octoberry/status/2121682690340450338
@Octoberry, replying in the same thread, agrees that Opus is actually nerfed, then adds the twist that it is still far better than Fable on cost efficiency (a). That is a useful detail. Even people who believe the downgrade keep Opus as their default, because Fable's price per task is the bigger pain for them.
#5
@kh8rl
https://x.com/kh8rl/status/2121667346200682628
@kh8rl asks how Claude is getting worse at writing copy and says Opus 4.5 was the peak (a). The warning to others is that it will hand you straight slop sometimes if you don't steer it. This is the non-coding version of the complaint, and a reminder that a model can get better at agentic code while marketing writers feel it got blander.
https://x.com/kh8rl/status/2121667346200682628
@kh8rl asks how Claude is getting worse at writing copy and says Opus 4.5 was the peak (a). The warning to others is that it will hand you straight slop sometimes if you don't steer it. This is the non-coding version of the complaint, and a reminder that a model can get better at agentic code while marketing writers feel it got blander.
#6
@heyfworld
https://x.com/heyfworld/status/2121014398198202589
@heyfworld says Opus 5.5 is almost unusable and needs correcting over and over to get anything right (a). The post makes the release-day argument directly: Codex models don't degrade like this, but a few weeks after each Claude launch the model seems to get nerfed. It ends by tagging Dario Amodei and Anthropic and asking for the release-day version back.
https://x.com/heyfworld/status/2121014398198202589
@heyfworld says Opus 5.5 is almost unusable and needs correcting over and over to get anything right (a). The post makes the release-day argument directly: Codex models don't degrade like this, but a few weeks after each Claude launch the model seems to get nerfed. It ends by tagging Dario Amodei and Anthropic and asking for the release-day version back.
#7
@dtrad
https://x.com/dtrad/status/2121715636909023537
@dtrad noticed it the night before and had to move back to DeepSeek 4.1 Flash to finish a task, calling it a major Opus regression (a). Dropping from a frontier model to a cheap open one mid-task is a strong behavioural signal, even without a benchmark. When people switch down, not across, they are saying the expensive model stopped earning its price that day.
https://x.com/dtrad/status/2121715636909023537
@dtrad noticed it the night before and had to move back to DeepSeek 4.1 Flash to finish a task, calling it a major Opus regression (a). Dropping from a frontier model to a cheap open one mid-task is a strong behavioural signal, even without a benchmark. When people switch down, not across, they are saying the expensive model stopped earning its price that day.
#8
@pyronaur
https://x.com/pyronaur/status/2121729984075087873
@pyronaur says Opus can't even do a rebase anymore and just wants to squash commits and cuddle (a). It is a joke, but a precise one. Rebase is a multi-step task where the model has to keep history straight, and falling back to squash is the lazy shortcut people mean when they say lazier. The post calls the amount of nerf maddening.
https://x.com/pyronaur/status/2121729984075087873
@pyronaur says Opus can't even do a rebase anymore and just wants to squash commits and cuddle (a). It is a joke, but a precise one. Rebase is a multi-step task where the model has to keep history straight, and falling back to squash is the lazy shortcut people mean when they say lazier. The post calls the amount of nerf maddening.
#9
@yituihi
https://x.com/yituihi/status/2121553729241358729
@yituihi claims Anthropic silently quantized Fable 5.2 overnight because too many Pro users were draining GPUs before Monday, and asks if it suddenly feels dumber (b). No source, no test and no screenshot come with it, and the motive is invented. It is worth listing as an example of how fast a quantization rumour gets posted as fact. Treat it as noise until someone shows a before-and-after.
https://x.com/yituihi/status/2121553729241358729
@yituihi claims Anthropic silently quantized Fable 5.2 overnight because too many Pro users were draining GPUs before Monday, and asks if it suddenly feels dumber (b). No source, no test and no screenshot come with it, and the motive is invented. It is worth listing as an example of how fast a quantization rumour gets posted as fact. Treat it as noise until someone shows a before-and-after.
#10
@ManuelBaut63639
https://x.com/ManuelBaut63639/status/2121687070888313003
@ManuelBaut63639 has the only fallback numbers of the window, from a benchmark of 115 runs (b). At high effort, 20.9% of Opus 5.5's replies came from a fallback to Opus 4.8, rising to 47.8% at max, while Fable 5.2 had zero fallback across all runs. The traces show Fable going straight to the binary, thinking once, rewriting and verifying, while Opus reasons longer and then falls back. That is why Fable at high ended up cheaper per task, $2.86 against $3.27, while scoring higher. If the fallback rate holds outside this benchmark, a lot of the felt nerf on Opus is really an older model answering.
https://x.com/ManuelBaut63639/status/2121687070888313003
@ManuelBaut63639 has the only fallback numbers of the window, from a benchmark of 115 runs (b). At high effort, 20.9% of Opus 5.5's replies came from a fallback to Opus 4.8, rising to 47.8% at max, while Fable 5.2 had zero fallback across all runs. The traces show Fable going straight to the binary, thinking once, rewriting and verifying, while Opus reasons longer and then falls back. That is why Fable at high ended up cheaper per task, $2.86 against $3.27, while scoring higher. If the fallback rate holds outside this benchmark, a lot of the felt nerf on Opus is really an older model answering.
#11
@chrimtech
https://x.com/chrimtech/status/2121249711952667021
@chrimtech flagged the second Claude incident of November 13: errors on the app, the API and Claude Code, then spreading to connectors and Managed Agents, a few hours after a morning login outage (c). Anthropic's status page called it degraded performance. A lot of the felt-dumber posts that day landed in the same hours, so part of the window's quality complaints is really an availability story.
https://x.com/chrimtech/status/2121249711952667021
@chrimtech flagged the second Claude incident of November 13: errors on the app, the API and Claude Code, then spreading to connectors and Managed Agents, a few hours after a morning login outage (c). Anthropic's status page called it degraded performance. A lot of the felt-dumber posts that day landed in the same hours, so part of the window's quality complaints is really an availability story.
#12
@CVSGireesh
https://x.com/CVSGireesh/status/2121180460445245883
@CVSGireesh posted the logs behind the morning failure: Claude Desktop 5.1231.2 on macOS kept signing out and killing every session, with 503s on the Google verification and account endpoints and a Cloudflare 403 on auth starting at 3:47 AM Central (c). The status page said degraded. This is what an outage looks like from the user side: not a dumber model, just sessions that die.
https://x.com/CVSGireesh/status/2121180460445245883
@CVSGireesh posted the logs behind the morning failure: Claude Desktop 5.1231.2 on macOS kept signing out and killing every session, with 503s on the Google verification and account endpoints and a Cloudflare 403 on auth starting at 3:47 AM Central (c). The status page said degraded. This is what an outage looks like from the user side: not a dumber model, just sessions that die.
#13
@CueEditz
https://x.com/CueEditz/status/2121482016067907700
@CueEditz doesn't understand how Claude usage limits work (c). Sometimes two messages end the session, and other times the model creates more than eight coding files without any warning. The complaint is not the cap itself, it is that the meter is invisible, so nobody can plan a day of work around it.
https://x.com/CueEditz/status/2121482016067907700
@CueEditz doesn't understand how Claude usage limits work (c). Sometimes two messages end the session, and other times the model creates more than eight coding files without any warning. The complaint is not the cap itself, it is that the meter is invisible, so nobody can plan a day of work around it.
#14
@ObsvRetardation
https://x.com/ObsvRetardation/status/2121697268961284606
@ObsvRetardation planned to run a regression of BMI against weight loss across the more than 25 trials in their library, and has to wait four hours for the weekly Claude usage to reset first (c). It is a small post, but it shows how weekly caps hit research work: the analysis is ready, the data is ready, and the clock is the bottleneck.
https://x.com/ObsvRetardation/status/2121697268961284606
@ObsvRetardation planned to run a regression of BMI against weight loss across the more than 25 trials in their library, and has to wait four hours for the weekly Claude usage to reset first (c). It is a small post, but it shows how weekly caps hit research work: the analysis is ready, the data is ready, and the clock is the bottleneck.
#15
@getLeveragedai
https://x.com/getLeveragedai/status/2121350318151610459
@getLeveragedai nearly hit the usage limits on all of their Claude Ultra accounts and says Opus 5.6 is needed as soon as possible (c). Note the plural: the account is already split across several top-tier plans and still runs dry. For heavy users, limits are the nerf that actually stops work.
https://x.com/getLeveragedai/status/2121350318151610459
@getLeveragedai nearly hit the usage limits on all of their Claude Ultra accounts and says Opus 5.6 is needed as soon as possible (c). Note the plural: the account is already split across several top-tier plans and still runs dry. For heavy users, limits are the nerf that actually stops work.
#16
@kaymakefendl
https://x.com/kaymakefendl/status/2121722185744261578
@kaymakefendl asks whether other Claude Max users have noticed their limits running out much faster than usual, and tags Boris Cherny, Anthropic and Claude (c). It is a question, not a measurement, but it repeats the shape of the last two weeks: faster drain with no visible change in what one prompt costs.
https://x.com/kaymakefendl/status/2121722185744261578
@kaymakefendl asks whether other Claude Max users have noticed their limits running out much faster than usual, and tags Boris Cherny, Anthropic and Claude (c). It is a question, not a measurement, but it repeats the shape of the last two weeks: faster drain with no visible change in what one prompt costs.
#17
@iyzebhel
https://x.com/iyzebhel/status/2121728341715685735
@iyzebhel is now so afraid of tripping the classifier on Fable 5.2 that affectionate messages and photos of the day go through Fable 5.1 in one chat, which passes them to Fable 5.2 in another, and the reply is read back the same way (d). The workaround itself is not new, the poster says, but this time it is forced rather than chosen. It is a companion-use case, not a coding one, and it shows the classifier's cost is not only false blocks but users rebuilding their behaviour around the fear of one.
https://x.com/iyzebhel/status/2121728341715685735
@iyzebhel is now so afraid of tripping the classifier on Fable 5.2 that affectionate messages and photos of the day go through Fable 5.1 in one chat, which passes them to Fable 5.2 in another, and the reply is read back the same way (d). The workaround itself is not new, the poster says, but this time it is forced rather than chosen. It is a companion-use case, not a coding one, and it shows the classifier's cost is not only false blocks but users rebuilding their behaviour around the fear of one.
#18
@YashVrdn
https://x.com/YashVrdn/status/2121720119164559612
@YashVrdn argues that unless you work on frontier research, Fable 5.2 is overkill, and adds that Anthropic already nerfs the model's output for computer science and biotech (d). That second clause matches Anthropic's own disclosure that Fable 5.2's kernel and architecture skills were held back at Opus 5.5 level. So in this one area, the nerf is not a rumour but a stated policy, and users are starting to fold it into the general story.
https://x.com/YashVrdn/status/2121720119164559612
@YashVrdn argues that unless you work on frontier research, Fable 5.2 is overkill, and adds that Anthropic already nerfs the model's output for computer science and biotech (d). That second clause matches Anthropic's own disclosure that Fable 5.2's kernel and architecture skills were held back at Opus 5.5 level. So in this one area, the nerf is not a rumour but a stated policy, and users are starting to fold it into the general story.
#19
@EnvolDev
https://x.com/EnvolDev/status/2121641306107871500
@EnvolDev ran a public retest after wondering out loud the day before whether Fable 5.2 had been nerfed (e). Same prompt, The Crown of the Reef, both at Extra High. Fable 5.2 took $78.45 and 3h 15m, Opus 5.5 took $54.96 and 1h 56m, and Fable's output won on detail: curtains, stilts, the lagoon, the peaks. The conclusion is that the earlier miss was task-specific, not a nerf, and that Fable is still the better model at a $23 and 79-minute premium.
https://x.com/EnvolDev/status/2121641306107871500
@EnvolDev ran a public retest after wondering out loud the day before whether Fable 5.2 had been nerfed (e). Same prompt, The Crown of the Reef, both at Extra High. Fable 5.2 took $78.45 and 3h 15m, Opus 5.5 took $54.96 and 1h 56m, and Fable's output won on detail: curtains, stilts, the lagoon, the peaks. The conclusion is that the earlier miss was task-specific, not a nerf, and that Fable is still the better model at a $23 and 79-minute premium.
#20
@fierronava_
https://x.com/fierronava_/status/2120807311405089070
@fierronava_ pushes back on the take that Claude will get worse to please Wall Street after the IPO roadshow (e). The argument: the model is the product, anyone can compare it with GPT or Gemini the same day, and making it worse would kill the revenue investors want. What to watch instead is the deal around the model: tighter limits on cheaper plans, higher seat prices, new features landing on the top tier first, worse rate limits and slower support. That reframes the whole nerf debate from weights to terms.
https://x.com/fierronava_/status/2120807311405089070
@fierronava_ pushes back on the take that Claude will get worse to please Wall Street after the IPO roadshow (e). The argument: the model is the product, anyone can compare it with GPT or Gemini the same day, and making it worse would kill the revenue investors want. What to watch instead is the deal around the model: tighter limits on cheaper plans, higher seat prices, new features landing on the top tier first, worse rate limits and slower support. That reframes the whole nerf debate from weights to terms.
#21
@erezt
https://x.com/erezt/status/2121654134592221508
@erezt says OpenAI's Sol 6.1 at medium effort got dumbed down, failing basic coding tasks it used to nail, and that it feels tuned to burn tokens on the cheapest compute (e). The punchline is the move: from OpenAI to Opus 5.5 at medium, with the verdict that ChatGPT isn't even close right now. In a window full of Opus complaints, this is a reminder that every lab gets the same accusation, and Opus is still somebody's escape hatch.
https://x.com/erezt/status/2121654134592221508
@erezt says OpenAI's Sol 6.1 at medium effort got dumbed down, failing basic coding tasks it used to nail, and that it feels tuned to burn tokens on the cheapest compute (e). The punchline is the move: from OpenAI to Opus 5.5 at medium, with the verdict that ChatGPT isn't even close right now. In a window full of Opus complaints, this is a reminder that every lab gets the same accusation, and Opus is still somebody's escape hatch.
#22
@DrBeavisAI
https://x.com/DrBeavisAI/status/2121707116930949384
@DrBeavisAI says Carbon hallucinates a lot in chat and Antigravity, with about 15% of literature papers linked wrongly and made-up numbers, while these mistakes don't happen with GPT and only very rarely with Opus or Fable (e). The poster upgraded to Ultra and downgraded back to Pro the next day. For research use, it is a counterpoint: on citation accuracy, Claude is still the model people trust.
https://x.com/DrBeavisAI/status/2121707116930949384
@DrBeavisAI says Carbon hallucinates a lot in chat and Antigravity, with about 15% of literature papers linked wrongly and made-up numbers, while these mistakes don't happen with GPT and only very rarely with Opus or Fable (e). The poster upgraded to Ultra and downgraded back to Pro the next day. For research use, it is a counterpoint: on citation accuracy, Claude is still the model people trust.
#23
@MoaHabtamu
https://x.com/MoaHabtamu/status/2121696990517973169
@MoaHabtamu replies to people praising Gemini's limits that Carbon is very slow and drains usage much faster than Opus 5.5 for coding, so the limits are not even close (e). It is a direct counter to the limits complaints elsewhere in this window: measured against the main alternative, the person finds Claude's allowance the better deal.
https://x.com/MoaHabtamu/status/2121696990517973169
@MoaHabtamu replies to people praising Gemini's limits that Carbon is very slow and drains usage much faster than Opus 5.5 for coding, so the limits are not even close (e). It is a direct counter to the limits complaints elsewhere in this window: measured against the main alternative, the person finds Claude's allowance the better deal.
#24
@ke_xin_han
https://x.com/ke_xin_han/status/2121697072974082397
@ke_xin_han points out that November 13 had two outages, the morning login failure and a second one hours later that hit the API, and that while everyone argues over whether Opus downgraded, the service itself went down twice (e). The point is about attribution. A model that errors out and a model that got dumber feel similar from the keyboard, and the evidence this window supports the first much better than the second.
https://x.com/ke_xin_han/status/2121697072974082397
@ke_xin_han points out that November 13 had two outages, the morning login failure and a second one hours later that hit the API, and that while everyone argues over whether Opus downgraded, the service itself went down twice (e). The point is about attribution. A model that errors out and a model that got dumber feel similar from the keyboard, and the evidence this window supports the first much better than the second.
#25
@codedusting
https://x.com/codedusting/status/2121732749816160266
@codedusting says the setup never hits usage limits on end-to-end production builds (e). The trick is orchestration: T3 Code drives OpenCode on a $10 plan with DeepSeek and Kimi models, while a $20 Opus 5.5 high plan does the steering, and Fable 5.2 runs weekly security sweeps. The whole stack costs about $45 a month in India and powers a live production app. It is counter-evidence with a method attached: limits bite people who make Claude do everything.
https://x.com/codedusting/status/2121732749816160266
@codedusting says the setup never hits usage limits on end-to-end production builds (e). The trick is orchestration: T3 Code drives OpenCode on a $10 plan with DeepSeek and Kimi models, while a $20 Opus 5.5 high plan does the steering, and Fable 5.2 runs weekly security sweeps. The whole stack costs about $45 a month in India and powers a live production app. It is counter-evidence with a method attached: limits bite people who make Claude do everything.
#26
@pulpmatrix
https://x.com/pulpmatrix/status/2121664566270873914
@pulpmatrix found that Codex nailed the cold start but degraded once a session got long, so the harness now checks task shape before picking a model (e). It counts expected tool calls and checks whether there is existing context to chain from, and sends debugging and long multi-file sessions to Opus. Degradation here is a property of session length, not of a vendor, and the routing rule puts Claude where it holds up best.
https://x.com/pulpmatrix/status/2121664566270873914
@pulpmatrix found that Codex nailed the cold start but degraded once a session got long, so the harness now checks task shape before picking a model (e). It counts expected tool calls and checks whether there is existing context to chain from, and sends debugging and long multi-file sessions to Opus. Degradation here is a property of session length, not of a vendor, and the routing rule puts Claude where it holds up best.
#27
@xinxianhub
https://x.com/xinxianhub/status/2120673695513346510
@xinxianhub calls Anthropic's make-good a coupon (f). The post says Anthropic's own postmortem admits Claude Code got dumber after its changes to reasoning, caching and prompts, while Pro, Max and Ultra subscribers kept paying. November 9 brought Fable-only resets, and November 10 added one full reset for all paid plans, banked and expiring on December 10. The question it leaves is pointed: what would it take to walk, another burned week or another expiring apology?
https://x.com/xinxianhub/status/2120673695513346510
@xinxianhub calls Anthropic's make-good a coupon (f). The post says Anthropic's own postmortem admits Claude Code got dumber after its changes to reasoning, caching and prompts, while Pro, Max and Ultra subscribers kept paying. November 9 brought Fable-only resets, and November 10 added one full reset for all paid plans, banked and expiring on December 10. The question it leaves is pointed: what would it take to walk, another burned week or another expiring apology?
#28
@Ssisakhti
https://x.com/Ssisakhti/status/2121699067935773111
@Ssisakhti explains why the auto-mode classifier behaves the way it does, citing Anthropic's write-up (f). The classifier sees only your messages and the agent's tool calls. Tool outputs and Claude's own text are stripped, so a fetched page claiming you granted permission never reaches it. That is the design answer to people surprised when auto mode ignores context from a web page: the blindness is deliberate, it is the defence against prompt injection.
https://x.com/Ssisakhti/status/2121699067935773111
@Ssisakhti explains why the auto-mode classifier behaves the way it does, citing Anthropic's write-up (f). The classifier sees only your messages and the agent's tool calls. Tool outputs and Claude's own text are stripped, so a fetched page claiming you granted permission never reaches it. That is the design answer to people surprised when auto mode ignores context from a web page: the blindness is deliberate, it is the defence against prompt injection.
#29
@Osu_most
https://x.com/Osu_most/status/2120922242787754185
@Osu_most read Anthropic's docs for Claude on Google Cloud before anyone moves an agent there (f). Fable 5.2 is available with a 1M-token context, but the Batches API, Files API, code execution, web fetch, Agent Skills, the MCP connector, programmatic tool calling, Managed Agents and server-side fallbacks are not. Regional endpoints only serve older models, multi-region costs 10% more than global, and requests cap at 30 MB. The warning is that the same model on a different cloud is a different product, and missing features will feel exactly like a nerf.
https://x.com/Osu_most/status/2120922242787754185
@Osu_most read Anthropic's docs for Claude on Google Cloud before anyone moves an agent there (f). Fable 5.2 is available with a 1M-token context, but the Batches API, Files API, code execution, web fetch, Agent Skills, the MCP connector, programmatic tool calling, Managed Agents and server-side fallbacks are not. Regional endpoints only serve older models, multi-region costs 10% more than global, and requests cap at 30 MB. The warning is that the same model on a different cloud is a different product, and missing features will feel exactly like a nerf.
#30
@tiagojardins
https://x.com/tiagojardins/status/2121679593920766449
@tiagojardins offers the boring explanation for Claude getting dumber the longer you chat: context rot (f). As a conversation grows, earlier instructions get diluted and quality slides, whatever the model. The fix is to start fresh sessions more often. It is worth keeping next to the nerf claims, because a lot of the window's examples come from long sessions.
https://x.com/tiagojardins/status/2121679593920766449
@tiagojardins offers the boring explanation for Claude getting dumber the longer you chat: context rot (f). As a conversation grows, earlier instructions get diluted and quality slides, whatever the model. The fix is to start fresh sessions more often. It is worth keeping next to the nerf claims, because a lot of the window's examples come from long sessions.
#31
@hammadspeaks
https://x.com/hammadspeaks/status/2121715960587628936
@hammadspeaks went looking for what eats Claude Code limits and found a fresh session already held 66,300 tokens of context before any coding, about 50k of it from system instructions, tool definitions, MCP servers and skills (f). About 25k of that was tool definitions that could be deferred. The result is Wirewolf, a free open-source pane for the Claude desktop app that shows the raw request, tracks tokens, defers unused tools and trims large outputs. If limits feel smaller, the first place to look is your own setup.
https://x.com/hammadspeaks/status/2121715960587628936
@hammadspeaks went looking for what eats Claude Code limits and found a fresh session already held 66,300 tokens of context before any coding, about 50k of it from system instructions, tool definitions, MCP servers and skills (f). About 25k of that was tool definitions that could be deferred. The result is Wirewolf, a free open-source pane for the Claude desktop app that shows the raw request, tracks tokens, defers unused tools and trims large outputs. If limits feel smaller, the first place to look is your own setup.
#32
@DecadentPuff
https://x.com/DecadentPuff/status/2121727919781564921
@DecadentPuff names a quieter cause: instructions written for far dumber Claude models now get in the way of Opus (f). Old CLAUDE.md rules and prompt crutches that compensated for weaker models can make a stronger one look worse. The practical takeaway is to prune instructions after each model upgrade instead of carrying them forward forever.
https://x.com/DecadentPuff/status/2121727919781564921
@DecadentPuff names a quieter cause: instructions written for far dumber Claude models now get in the way of Opus (f). Old CLAUDE.md rules and prompt crutches that compensated for weaker models can make a stronger one look worse. The practical takeaway is to prune instructions after each model upgrade instead of carrying them forward forever.
#33
@roughneckactual
https://x.com/roughneckactual/status/2121664464085024837
@roughneckactual hit Claude's usage limits and the bot switched to Codex by itself (f). The setup is short: install the Codex and Claude Code CLIs, log in to both, and click Allow in the Grok Bot desktop app. It is the cheapest answer to limits in this window: don't wait for the reset, let the harness fail over.
https://x.com/roughneckactual/status/2121664464085024837
@roughneckactual hit Claude's usage limits and the bot switched to Codex by itself (f). The setup is short: install the Codex and Claude Code CLIs, log in to both, and click Allow in the Grok Bot desktop app. It is the cheapest answer to limits in this window: don't wait for the reset, let the harness fail over.
#34
@Salim4ai
https://x.com/Salim4ai/status/2121684879251214350
@Salim4ai points to Anthropic's official documentation for the experimental Code Mode in Claude Code 2.1.323 and later (f). Instead of making tool calls one by one, Claude writes JavaScript that orchestrates several calls in parallel, doing more in a single step. For people watching five-hour and weekly limits, this is the one official change this window that could make the same allowance go further.
https://x.com/Salim4ai/status/2121684879251214350
@Salim4ai points to Anthropic's official documentation for the experimental Code Mode in Claude Code 2.1.323 and later (f). Instead of making tool calls one by one, Claude writes JavaScript that orchestrates several calls in parallel, doing more in a single step. For people watching five-hour and weekly limits, this is the one official change this window that could make the same allowance go further.
#35
@Dekrov_Dev
https://x.com/Dekrov_Dev/status/2121576298677940325
@Dekrov_Dev reads the Claude Code 2.1.322 release note on fast mode for Vertex AI in the global, US and EU regions (f). The detail that matters: if Vertex rejects fast mode, Claude Code falls back to standard speed instead of reporting it unavailable. That is a silent fallback by design, and anyone who feels Claude got slower on Vertex should check whether the fast path is actually active.
https://x.com/Dekrov_Dev/status/2121576298677940325
@Dekrov_Dev reads the Claude Code 2.1.322 release note on fast mode for Vertex AI in the global, US and EU regions (f). The detail that matters: if Vertex rejects fast mode, Claude Code falls back to standard speed instead of reporting it unavailable. That is a silent fallback by design, and anyone who feels Claude got slower on Vertex should check whether the fast path is actually active.
#36
Signal Board
Today: 17 complaints vs 9 evidence or counter-evidence posts vs 5 official items from Anthropic, none of which addresses model quality.
Today: 17 complaints vs 9 evidence or counter-evidence posts vs 5 official items from Anthropic, none of which addresses model quality.
Comments