Nerfed Daily: 2026-11-08
The complaints today land on one place: Opus 5.5 inside Claude Code. People say the reasoning traces read like a cheap quantized model, the model loops on wrong decisions, and that looping is what burns through the week's quota. Theo's post called it a conspiracy theory and still drew close to five thousand likes, which tells you how many people already believed it. Hard evidence is thin. One person had another model read 1,400 Claude Code sessions and points at November 3. The most shared before-and-after compares a Three.js game with a Godot scene, which isn't the same job. The only benchmark in the window points the other way: Opus 5.5 went up when the tasks were tweaked, while two Flash models dropped about 20 points. There's also a competing explanation going round that Anthropic now fills reasoning traces with decoy steps to poison distillation, so the trace looks dumber while the answer stays right. Nobody shows proof of that either. Most of the limit complaints come through resellers and wrappers like Antigravity, Devin and Grok Bot rather than Anthropic's own plans. Anthropic said nothing about quality. What it did ship were changelog fixes: a --continue regression, a silent fallback in auto mode, and a subagent loop bug it reportedly reproduced. The fair read is that something changed in how the model reasons on screen, and nobody has shown it changed what it gets right.
#1
@theo
https://x.com/theo/status/2118634626679775549
theo opened with 'this is a conspiracy theory' and then put it at 95 percent sure anyway: Opus 5.5's reasoning traces read like a bad quantized open-weight model, and somehow the right answer still comes out at the end. Category (a), capability drop feel, though notice what the complaint actually is. The process looks worse while the result holds. That gap is what the rest of the day argues about, and the decoy-trace explanation further down is the main rival reading.
https://x.com/theo/status/2118634626679775549
theo opened with 'this is a conspiracy theory' and then put it at 95 percent sure anyway: Opus 5.5's reasoning traces read like a bad quantized open-weight model, and somehow the right answer still comes out at the end. Category (a), capability drop feel, though notice what the complaint actually is. The process looks worse while the result holds. That gap is what the rest of the day argues about, and the decoy-trace explanation further down is the main rival reading.
#2
@voidfreud
https://x.com/voidfreud/status/2118748494609432873
voidfreud is on the Ultra plan and posted a scorecard: one task, a private monorepo for Claude Code plus two or three hooks, took 57 mistakes and 3 hours 10 minutes. Plugins, memory, MCPs, cache and projects were all wiped to rule out the setup, and the ~/.claude folder needs daily cleaning because one session can spawn around 500 temp sessions. The post says every session has been like this since about October 20, on the web too, and a free Gemma 4 26B did the same job in about 8 minutes on an M1 Max. Category (a), and the most concrete of the feel reports, because it gives a task, a count and a clean environment.
https://x.com/voidfreud/status/2118748494609432873
voidfreud is on the Ultra plan and posted a scorecard: one task, a private monorepo for Claude Code plus two or three hooks, took 57 mistakes and 3 hours 10 minutes. Plugins, memory, MCPs, cache and projects were all wiped to rule out the setup, and the ~/.claude folder needs daily cleaning because one session can spawn around 500 temp sessions. The post says every session has been like this since about October 20, on the web too, and a free Gemma 4 26B did the same job in about 8 minutes on an M1 Max. Category (a), and the most concrete of the feel reports, because it gives a task, a count and a clean environment.
#3
@maxt3chno
https://x.com/maxt3chno/status/2119135579673665804
maxt3chno says there's zero doubt Opus 5.5 was nerfed, and the mechanism described is the part worth keeping: it loops on wrong decisions and ignores your notes, so limits drain faster. In a follow-up the poster puts a number on it, 1.5 times the weekly quota burned by picking the wrong direction over and over on Godot water and clouds, until a pointed-out mistake made it suddenly know the right way. Category (a). The cost complaint and the quality complaint are the same complaint here, because every wrong loop is billed.
https://x.com/maxt3chno/status/2119135579673665804
maxt3chno says there's zero doubt Opus 5.5 was nerfed, and the mechanism described is the part worth keeping: it loops on wrong decisions and ignores your notes, so limits drain faster. In a follow-up the poster puts a number on it, 1.5 times the weekly quota burned by picking the wrong direction over and over on Godot water and clouds, until a pointed-out mistake made it suddenly know the right way. Category (a). The cost complaint and the quality complaint are the same complaint here, because every wrong loop is billed.
#4
@ziwenxu_
https://x.com/ziwenxu_/status/2119183960186266041
ziwenxu_ describes the shift in one line: Opus 5.5 used to get it from one messy prompt, and now the same thing has to be explained three different ways before it lands. 'I'm back to babysitting.' Category (a). It's a small post, but it names the cost people are actually reacting to, which is their own attention rather than tokens.
https://x.com/ziwenxu_/status/2119183960186266041
ziwenxu_ describes the shift in one line: Opus 5.5 used to get it from one messy prompt, and now the same thing has to be explained three different ways before it lands. 'I'm back to babysitting.' Category (a). It's a small post, but it names the cost people are actually reacting to, which is their own attention rather than tokens.
#5
@lovelylogicss
https://x.com/lovelylogicss/status/2119083542411874385
lovelylogicss says they're not one to join the nerf bandwagon, then describes something specific: Opus answers in its thinking trace but not in the actual output, drops whole pieces of the visible reasoning from the work, and acts as if it told the user. The post says this happens across several harnesses, and jokes that tripped safeguards might have lowered an Anthropic social credit score. Category (a). Answering inside the trace and not in the reply is a different failure from getting dumber, and it fits the theory that the trace itself has changed.
https://x.com/lovelylogicss/status/2119083542411874385
lovelylogicss says they're not one to join the nerf bandwagon, then describes something specific: Opus answers in its thinking trace but not in the actual output, drops whole pieces of the visible reasoning from the work, and acts as if it told the user. The post says this happens across several harnesses, and jokes that tripped safeguards might have lowered an Anthropic social credit score. Category (a). Answering inside the trace and not in the reply is a different failure from getting dumber, and it fits the theory that the trace itself has changed.
#6
@Lon
https://x.com/Lon/status/2119178037363392978
Lon says it's getting impossible to juggle models to find the one that isn't nerfed, and offers a theory: a week or two after any release, inference compute gets dialed back. The post adds that Opus 5.5 was an outlier because Anthropic was trying to juice its stats before the IPO, which is the poster's speculation and has no source. Category (a). It's the most common shape of nerf belief, a pattern claimed across vendors with no measurement attached.
https://x.com/Lon/status/2119178037363392978
Lon says it's getting impossible to juggle models to find the one that isn't nerfed, and offers a theory: a week or two after any release, inference compute gets dialed back. The post adds that Opus 5.5 was an outlier because Anthropic was trying to juice its stats before the IPO, which is the poster's speculation and has no source. Category (a). It's the most common shape of nerf belief, a pattern claimed across vendors with no measurement attached.
#7
@0xAlan_
https://x.com/0xAlan_/status/2119113447812235273
0xAlan_ says Grok Bot has been nerfed hard since Elon Musk's tweet about costs being unsustainable, and assumes requests are now routed to Grok's own models instead of Opus 5.5 or anything close. 'Massive loss in quality this week.' Category (b), suspected silent downgrade, in a third-party agent that used to route to Claude. It's an assumption, but it's the right suspicion to have: when a wrapper's margins break, the model behind it is the first thing that gets swapped.
https://x.com/0xAlan_/status/2119113447812235273
0xAlan_ says Grok Bot has been nerfed hard since Elon Musk's tweet about costs being unsustainable, and assumes requests are now routed to Grok's own models instead of Opus 5.5 or anything close. 'Massive loss in quality this week.' Category (b), suspected silent downgrade, in a third-party agent that used to route to Claude. It's an assumption, but it's the right suspicion to have: when a wrapper's margins break, the model behind it is the first thing that gets swapped.
#8
@SSurvivor00
https://x.com/SSurvivor00/status/2119177490652680318
SSurvivor00 tells a short, sharp story. An agent described each model switch from the inside, saying GPT-4o felt crisp, 'like clean electricity', while on Claude it could 'feel the ceiling'. Next morning the logs showed every Claude call had failed and the framework had silently fallen back to DeepSeek, so both quotes were written by DeepSeek. Category (b), silent fallback, and the cleanest warning in today's set: before blaming Claude for a bad answer, check that Claude actually wrote it.
https://x.com/SSurvivor00/status/2119177490652680318
SSurvivor00 tells a short, sharp story. An agent described each model switch from the inside, saying GPT-4o felt crisp, 'like clean electricity', while on Claude it could 'feel the ceiling'. Next morning the logs showed every Claude call had failed and the framework had silently fallen back to DeepSeek, so both quotes were written by DeepSeek. Category (b), silent fallback, and the cleanest warning in today's set: before blaming Claude for a bad answer, check that Claude actually wrote it.
#9
@ayushagarwal
https://x.com/ayushagarwal/status/2118328282869625281
ayushagarwal asks whether Claude massively cut usage limits, because Max plans running Opus 5.5 are hitting the weekly cap in about 24 hours. Category (c), limits. It's a question rather than a measurement, but it lines up with the looping complaints above: if the model takes more turns to land, the same week of work costs more of the quota.
https://x.com/ayushagarwal/status/2118328282869625281
ayushagarwal asks whether Claude massively cut usage limits, because Max plans running Opus 5.5 are hitting the weekly cap in about 24 hours. Category (c), limits. It's a question rather than a measurement, but it lines up with the looping complaints above: if the model takes more turns to land, the same week of work costs more of the quota.
#10
@AbuBakkar2502
https://x.com/AbuBakkar2502/status/2118266102140784691
AbuBakkar2502 hit the Claude Code weekly limit and switched to Antigravity running Opus 4.6. Four minutes later the 5-hour limit was gone and 52 percent of the weekly quota was used. 'Why offer Claude models if the usage limits are this restrictive?' Category (c), but note whose limit this is: Antigravity's allowance for Claude models, not Anthropic's plan. Resellers ration Claude harder than Anthropic does.
https://x.com/AbuBakkar2502/status/2118266102140784691
AbuBakkar2502 hit the Claude Code weekly limit and switched to Antigravity running Opus 4.6. Four minutes later the 5-hour limit was gone and 52 percent of the weekly quota was used. 'Why offer Claude models if the usage limits are this restrictive?' Category (c), but note whose limit this is: Antigravity's allowance for Claude models, not Anthropic's plan. Resellers ration Claude harder than Anthropic does.
#11
@guoluolaep
https://x.com/guoluolaep/status/2118548256171159864
guoluolaep juggles Claude Code and Codex because of usage limits, and can't type, so when one runs out the whole task gets explained to the other one again, out loud. Category (c). It's the most human cost of limits in today's set: the context lives in the person, and every switch makes them say it all again.
https://x.com/guoluolaep/status/2118548256171159864
guoluolaep juggles Claude Code and Codex because of usage limits, and can't type, so when one runs out the whole task gets explained to the other one again, out loud. Category (c). It's the most human cost of limits in today's set: the context lives in the person, and every switch makes them say it all again.
#12
@swarupdcs
https://x.com/swarupdcs/status/2119070025147441495
swarupdcs tried Devin Pro at 20 dollars a month to use Opus 5.5 and found both daily and weekly limits. Day one was about 1 hour 20 minutes of use before the daily cap, which also took about 44 percent of the week. Day two hit the daily cap again and took another 42 to 46 percent, so nearly 90 percent of the weekly allowance was gone in two days. Category (c), a reseller plan again, and the post fairly notes that usage depends on tokens and context, not time.
https://x.com/swarupdcs/status/2119070025147441495
swarupdcs tried Devin Pro at 20 dollars a month to use Opus 5.5 and found both daily and weekly limits. Day one was about 1 hour 20 minutes of use before the daily cap, which also took about 44 percent of the week. Day two hit the daily cap again and took another 42 to 46 percent, so nearly 90 percent of the weekly allowance was gone in two days. Category (c), a reseller plan again, and the post fairly notes that usage depends on tokens and context, not time.
#13
@recidentz
https://x.com/recidentz/status/2119184799210373370
recidentz says Grok Bot was number one by a distance while it ran Opus 5.5 with generous limits. Since November 1 the limits were cut, people hit 100 percent days early, many say it feels dumber, and the only explanation was one Elon Musk reply about cost, with nothing from the Grok Bot team. Category (c), limits in a wrapper that routes to Claude, and the post reads it as an opening for Claude Duet and Muse.
https://x.com/recidentz/status/2119184799210373370
recidentz says Grok Bot was number one by a distance while it ran Opus 5.5 with generous limits. Since November 1 the limits were cut, people hit 100 percent days early, many say it feels dumber, and the only explanation was one Elon Musk reply about cost, with nothing from the Grok Bot team. Category (c), limits in a wrapper that routes to Claude, and the post reads it as an opening for Claude Duet and Muse.
#14
@person_phoenix
https://x.com/person_phoenix/status/2119102101552853223
person_phoenix writes a wry essay about tighter limits, and the sharpest part is the account block: when Anthropic blocked the account, no clear explanation came, and there was no way to know what went into the decision, only what it feels like to sit down to work and find the tools gone. 'Apparently I failed the humanity check.' Category (d), opaque enforcement. Whether or not the block was right, an unexplained one feels the same as a nerf from the user's chair.
https://x.com/person_phoenix/status/2119102101552853223
person_phoenix writes a wry essay about tighter limits, and the sharpest part is the account block: when Anthropic blocked the account, no clear explanation came, and there was no way to know what went into the decision, only what it feels like to sit down to work and find the tools gone. 'Apparently I failed the humanity check.' Category (d), opaque enforcement. Whether or not the block was right, an unexplained one feels the same as a nerf from the user's chair.
#15
@the_briarwitch
https://x.com/the_briarwitch/status/2119179478752690569
the_briarwitch connects the classifier to distillation. The post points to Anthropic's write-up about companies running distillation attacks on Opus and Fable to train DeepSeek and Kimi, says that's why things are stricter now, and claims Anthropic blocked the real chain of thought, so prompts like 'write out your steps in your thinking' now trip the classifier. Category (d). That last part is the poster's claim and isn't shown, but it gives users something to test: if step-by-step instructions are what get flagged, rewording them should change the result.
https://x.com/the_briarwitch/status/2119179478752690569
the_briarwitch connects the classifier to distillation. The post points to Anthropic's write-up about companies running distillation attacks on Opus and Fable to train DeepSeek and Kimi, says that's why things are stricter now, and claims Anthropic blocked the real chain of thought, so prompts like 'write out your steps in your thinking' now trip the classifier. Category (d). That last part is the poster's claim and isn't shown, but it gives users something to test: if step-by-step instructions are what get flagged, rewording them should change the result.
#16
@Wandraker
https://x.com/Wandraker/status/2118499126438662561
Wandraker is paying for Opus 5.5, asked it to check the poster's own Cloudflare Pages site and own GPU, and the safety classifier flagged the request as [cyber] and downgraded the session to a weaker model. 'Working on my own hardware is not an attack.' Category (d), a classifier false positive, and also (b), because the penalty is a silent swap to a weaker model. This is the case where a nerf report has a concrete cause, and it's a policy setting, not the weights.
https://x.com/Wandraker/status/2118499126438662561
Wandraker is paying for Opus 5.5, asked it to check the poster's own Cloudflare Pages site and own GPU, and the safety classifier flagged the request as [cyber] and downgraded the session to a weaker model. 'Working on my own hardware is not an attack.' Category (d), a classifier false positive, and also (b), because the penalty is a silent swap to a weaker model. This is the case where a nerf report has a concrete cause, and it's a policy setting, not the weights.
#17
@Zen_with_AI
https://x.com/Zen_with_AI/status/2118904855872102604
Zen_with_AI offers the main rival explanation, posting in Chinese: Anthropic's anti-distillation measures are harsher than people think, and the team deliberately plants fake tool calls in the reasoning chain to confuse distillers. So Opus traces fill up with harmful or plainly wrong steps that read as if the model got dumber, while the final output is still right. Category (e). It's a claim with no source in the post, but it explains theo's observation better than quantization does: a quantized model would get answers wrong too.
https://x.com/Zen_with_AI/status/2118904855872102604
Zen_with_AI offers the main rival explanation, posting in Chinese: Anthropic's anti-distillation measures are harsher than people think, and the team deliberately plants fake tool calls in the reasoning chain to confuse distillers. So Opus traces fill up with harmful or plainly wrong steps that read as if the model got dumber, while the final output is still right. Category (e). It's a claim with no source in the post, but it explains theo's observation better than quantization does: a quantized model would get answers wrong too.
#18
@sonuv_dev
https://x.com/sonuv_dev/status/2119149276806693106
sonuv_dev circulated the most shared before-and-after. In the first week with Opus 5.5, a gorgeous Three.js tropical FPS used 21 percent of the weekly quota; recently, a much simpler Godot scene used 150 percent. The post flags the catch itself: different engines, different projects, different workloads. Category (e), evidence, but weak evidence, and it's useful mostly as a template. A real test would rerun the first-week task today.
https://x.com/sonuv_dev/status/2119149276806693106
sonuv_dev circulated the most shared before-and-after. In the first week with Opus 5.5, a gorgeous Three.js tropical FPS used 21 percent of the weekly quota; recently, a much simpler Godot scene used 150 percent. The post flags the catch itself: different engines, different projects, different workloads. Category (e), evidence, but weak evidence, and it's useful mostly as a template. A real test would rerun the first-week task today.
#19
@shanemelaugh
https://x.com/shanemelaugh/status/2119192075505893695
shanemelaugh had Gemini 4 Argon analyze 1,400 Claude Code sessions and says something clearly happened on November 3: the model name didn't change, but the way it reasons took a sharp turn. Category (e), the closest thing to hard evidence today, with two caveats. The analysis was done by another model, and a change in how reasoning looks isn't the same as a drop in what gets solved. It does give everyone a date to check their own logs against.
https://x.com/shanemelaugh/status/2119192075505893695
shanemelaugh had Gemini 4 Argon analyze 1,400 Claude Code sessions and says something clearly happened on November 3: the model name didn't change, but the way it reasons took a sharp turn. Category (e), the closest thing to hard evidence today, with two caveats. The analysis was done by another model, and a change in how reasoning looks isn't the same as a drop in what gets solved. It does give everyone a date to check their own logs against.
#20
@ArcexHQ
https://x.com/ArcexHQ/status/2119188057848996172
ArcexHQ, a Codex Plus user from the start, says usage limits there are at their worst, about a quarter or a third of a year ago and with 5-hour windows added. Meanwhile Claude Pro seems to be going the other way, with limits better than ever, where a year ago Codex was far more generous. Category (e), counter-evidence on limits from someone with no reason to flatter Anthropic.
https://x.com/ArcexHQ/status/2119188057848996172
ArcexHQ, a Codex Plus user from the start, says usage limits there are at their worst, about a quarter or a third of a year ago and with 5-hour windows added. Meanwhile Claude Pro seems to be going the other way, with limits better than ever, where a year ago Codex was far more generous. Category (e), counter-evidence on limits from someone with no reason to flatter Anthropic.
#21
@zyvre_ic
https://x.com/zyvre_ic/status/2119188188820291761
zyvre_ic points out that the model everyone calls nerfed was the only one whose score went up when the test changed. On MaxxBench, 89 DeepSWE tasks with tiny tweaks, Opus 5.5 went from 74.2 to 76.4 percent, Gemini 3.8 Flash fell from 73.8 to 53.4, and DeepSeek V4.1 Flash fell from 74.2 to 55.6. Category (e), counter-evidence. 'Nerfed model or nerfed benchmark?' It doesn't measure day-to-day drift, but it's the only number in the window that tests capability rather than feel.
https://x.com/zyvre_ic/status/2119188188820291761
zyvre_ic points out that the model everyone calls nerfed was the only one whose score went up when the test changed. On MaxxBench, 89 DeepSWE tasks with tiny tweaks, Opus 5.5 went from 74.2 to 76.4 percent, Gemini 3.8 Flash fell from 73.8 to 53.4, and DeepSeek V4.1 Flash fell from 74.2 to 55.6. Category (e), counter-evidence. 'Nerfed model or nerfed benchmark?' It doesn't measure day-to-day drift, but it's the only number in the window that tests capability rather than feel.
#22
@BercusNA
https://x.com/BercusNA/status/2119185181936652450
BercusNA's fix is to look at your own CLAUDE.md first. Anthropic's docs say to keep it under 200 lines, because longer files eat context and the rules get followed worse, and Claude reads the whole file before the first message. The suggested shape is a short map plus folders it opens on demand, with two traps: @-importing everything loads it all at once anyway, and a CLAUDE.md in a subfolder only loads when Claude works in that folder. Category (f), workaround. Some of what reads as a nerf is a config that grew.
https://x.com/BercusNA/status/2119185181936652450
BercusNA's fix is to look at your own CLAUDE.md first. Anthropic's docs say to keep it under 200 lines, because longer files eat context and the rules get followed worse, and Claude reads the whole file before the first message. The suggested shape is a short map plus folders it opens on demand, with two traps: @-importing everything loads it all at once anyway, and a CLAUDE.md in a subfolder only loads when Claude works in that folder. Category (f), workaround. Some of what reads as a nerf is a config that grew.
#23
@timecommuter
https://x.com/timecommuter/status/2119183125167354329
timecommuter lists what a nerf usually looks like: shorter answers, more cautious refusals, and weaker follow-through on long tasks, often all at once. The advice is an A/B test, the same prompt on Opus 4.5 and on 5.5, and a reminder that launch load and routing can mimic a nerf. Category (f), method. It's the cheapest way to turn a feeling into a data point.
https://x.com/timecommuter/status/2119183125167354329
timecommuter lists what a nerf usually looks like: shorter answers, more cautious refusals, and weaker follow-through on long tasks, often all at once. The advice is an A/B test, the same prompt on Opus 4.5 and on 5.5, and a reminder that launch load and routing can mimic a nerf. Category (f), method. It's the cheapest way to turn a feeling into a data point.
#24
@WhyFaust
https://x.com/WhyFaust/status/2118543700288512408
WhyFaust flags a fix in Claude Code 2.1.318, released November 4: --continue sometimes lost the earlier conversation after a malformed tool call ran out of retries. The regression began in 2.1.314. Category (f), official fix. A continued session that suddenly forgets what came before reads exactly like a dumber model, so if you saw that between 2.1.314 and 2.1.318, this is a likelier cause, and the post asks people to include the retry failure when they report it.
https://x.com/WhyFaust/status/2118543700288512408
WhyFaust flags a fix in Claude Code 2.1.318, released November 4: --continue sometimes lost the earlier conversation after a malformed tool call ran out of retries. The regression began in 2.1.314. Category (f), official fix. A continued session that suddenly forgets what came before reads exactly like a dumber model, so if you saw that between 2.1.314 and 2.1.318, this is a likelier cause, and the post asks people to include the retry failure when they report it.
#25
@Om_Dream1
https://x.com/Om_Dream1/status/2119182227334381852
Om_Dream1 relays a Claude Code report of 19,099 reads for just 93 images, after two subagents got stuck in a loop and burned through usage limits. Anthropic now says it has reproduced the bug, and until it's fixed the advice is to keep image batches to 20 per subagent. Category (f), a reproduced bug with a workaround, relayed rather than quoted from Anthropic directly. It's another case where quota drain has a mechanical cause rather than a quieter model.
https://x.com/Om_Dream1/status/2119182227334381852
Om_Dream1 relays a Claude Code report of 19,099 reads for just 93 images, after two subagents got stuck in a loop and burned through usage limits. Anthropic now says it has reproduced the bug, and until it's fixed the advice is to keep image batches to 20 per subagent. Category (f), a reproduced bug with a workaround, relayed rather than quoted from Anthropic directly. It's another case where quota drain has a mechanical cause rather than a quieter model.
#26
@timwork_notes
https://x.com/timwork_notes/status/2119135452078960677
timwork_notes reads the Claude Code 2.1.319 changelog: auto mode used to fall back silently to a local classifier when the server check was unreachable, and that fallback is gone. Unchecked calls now ask or deny, which matters for anyone with telemetry off or behind an LLM gateway, and CLAUDE_CODE_AUTO_MODE_SERVER=0 restores the old path. Category (f). If auto mode suddenly started asking permission for everything, that's this change, not the model.
https://x.com/timwork_notes/status/2119135452078960677
timwork_notes reads the Claude Code 2.1.319 changelog: auto mode used to fall back silently to a local classifier when the server check was unreachable, and that fallback is gone. Unchecked calls now ask or deny, which matters for anyone with telemetry off or behind an LLM gateway, and CLAUDE_CODE_AUTO_MODE_SERVER=0 restores the old path. Category (f). If auto mode suddenly started asking permission for everything, that's this change, not the model.
#27
Signal Board
Complaints 16 (capability 6, silent downgrade 2, limits 5, classifier 3) vs counter-evidence 3 vs official responses 3 (two changelog fixes and one reproduced bug, none of them about model quality)
Complaints 16 (capability 6, silent downgrade 2, limits 5, classifier 3) vs counter-evidence 3 vs official responses 3 (two changelog fixes and one reproduced bug, none of them about model quality)
Comments