Opus 'Nerf' or a New Summarizer?, Claude Code's Leaky 'Don't Ask Again', MaxxBench, Codex Pro Drops the 5-Hour Limit & Asahi Goes GPL

Claude & Anthropic

Is Opus 5.5 nerfed, or is it the summarizer?

Shane Melaugh asked (1,504 likes, 205K views): "is Opus 5.5 nerfed? I just had Gemini 4 Argon analyze 1,400 Claude Code sessions and it found that on November 3, something definitely happened. the model name didn't change, but the way it does reasoning took a sharp turn." Asked what changed, Shane said "different kind of reasoning, less reasoning before taking a first action and more errors", and added "my other two findings point more directly to a nerf".

Theo put an agent on it with Opus 5.5 on xhigh: "Dig through my cliproxy logs in clickhouse to see if things changed". Theo's conclusion (708 likes): "Anthropic appears to be rolling out a change to the reasoning summarizer on a per-account basis. No sign of degraded capability for me. I think they're just trying to fix the reasoning traces (they've been utter garbage for weeks, complained about this yesterday lol)". Yesterday's issue covered that complaint, and Anthropic's docs say the visible thinking is a summary written by a different model.

The summarizer theory got support:

  • @Moriarty_tech: "Those are the visible summaries though. Dashes dropping to 6.5% overnight smells like someone rewrote the summarizer prompt. Did the actual task results get worse?"
  • @YousseFathallaX: "if it's a per-account summarizer change, that would explain why half the timeline is screaming nerf and the other half notices nothing. traces looking worse != the model getting worse"
  • Youssof Al Toukhi guessed "switching from Haiku 4.5 to 5.5 for the summariser". Nobody from Anthropic confirmed which model it is.

Others say the model itself got worse. Chris Li: "My two accounts are night and day. One is defunct. The other is still good." Dan Braun: "this last week I'm just not running it without having another model verify the work". Shivam Singhal saw it on high effort: "Switched to xhigh and Opus went back to normal!" On the other side, Dylan K "Ran it all night on six rounds of fixes for my game and it held up fine. Only thing nerfed was my sleep", and @SormDev: "Opus 5.5 did not at all feel nerfed for me this week."

Tim's reply was the most careful: "Verdict: not proven. Something did change in Claude Code on Nov 2–3, but the documented changes were in Claude Code itself, not the model." The changelog supports part of that. Version 2.1.317 reached npm late on November 2 (UTC) and "Changed subagents and teammates on a different model from the main conversation to use that model's saved effort level or its default, instead of the session's effort level". Version 2.1.318 followed on November 3 and "Removed the 2.1.288 fix that kept the model's earlier thinking when resuming a conversation started on 2.1.286 or earlier". Neither entry mentions the summarizer. None of the Anthropic accounts we follow had replied by Sunday morning. Andrew asked for what would settle it: "We gotta get a proper changelog or something."

Claude Code's "Don't Ask Again" left passwords in public repos

Apiiro's report, "Claude Code's "Don't Ask Again" exposes passwords in 1,122 public GitHub files", reached 349 points on HN. The post itself is behind a Cloudflare wall, so the details here come from the HN thread and the docs. The settings docs say Claude Code writes .claude/settings.local.json "the first time you give a standing approval on a permission prompt, such as "Yes, and don't ask again" for a Bash command". If the approved command contains a password, the password goes into that file.

Aurornis called it a "Confusing headline": "If you commit your local Claude settings and in those settings you have manually approved commands containing secrets directly in the command and then you publish the repo, that's exposed." The Apiiro account replied that "Claude Code only adds it to your global git ignore, not the repo's .gitignore" and that "In 539 of the 1,105 repos we traced, a commit co-authored by Claude itself changed the file." The account later corrected one detail: "Fair, "fresh clone" was wrong. It happens on a new machine or container without the global ignore, in a copied project folder, or once the file has been committed". Several commenters flagged those replies as AI-written.

Boris Cherny answered for the Claude Code team: "Historically Claude Code has not done secret scrubbing (there are many good tools for that), but we are looking into what we can do to prevent this." Boris also said ".claude/settings.local.json is meant to stay local", that "Permission rules are old hat. Everyone should be using auto mode instead", and that "Regardless of what agent you use, you should have secret scanning enabled." No fix has shipped yet. "There's a few possible approaches: making sure auto mode catches this, building in basic secret scanning, a few other ideas." Boris's comment says "We automatically gitignore it", and the docs give the mechanism: Claude Code adds **/.claude/settings.local.json to "your global git excludes file". That protects one machine, not the repo. ceejayoz asked about auto mode: "Is that because auto mode is good, or because the manual approval system sucks?"

Quick hits

  • Watermarks for older Claude models. Notebookcheck and The Decoder report that Mythos 5, Opus 4.7 and Sonnet 4.6 have carried an invisible text watermark since November 3, and that Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 follow on November 16. Anthropic uses "a variant of Google DeepMind's SynthID text method". On the second wave, "the watermark only applies when temperature, top_p, and top_k are set to their defaults". Anthropic's help page says the mark will "travel with the text when it's copied and pasted elsewhere". Detection is in private preview for regulators, media, researchers and similar groups.
  • Fable 5.1 hiccup. The status page logged elevated errors for Fable 5.1 from 08:18 to 08:32 UTC on Sunday across claude.ai, the API, Claude Code and Cowork.
  • "Goodbye Claude, Hello Pi." Caleb's post (HN, 209 points) explains the switch to Pi: "At the time of writing, there have been 244 releases to Claude Code this year; over 3/4 of these changed the system prompt or tool descriptions". It also says "The default Bash tool's description is almost 3,000 tokens". The post is dated September 21, and a commenter said "this feels like six weeks (or more) out-of-date".

Benchmarks & Research

MaxxBench: change one requirement and watch the scores split

Justin Lo introduced MaxxBench (4,529 likes, 605K views; HN, 288 points): "89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement." The write-up gives an example: "We added a requirement that the filenames be lowercase." The feature stays the same, but "Chrome.xml should now be chrome.xml."

The results:

  • "Claude Opus 5.5 and Gemini 3.8 Flash both have published DeepSWE scores of about 74%. On MaxxBench, they score 76.4% and 53.4%."
  • "GPT-6 Astra and Opus 5 also start close together on DeepSWE, then separate to 71.9% and 63.5% on MaxxBench."
  • DeepSeek V4.1 Flash and Gemini 3.8 Flash "fall by 18.6 to 20.5 points". Across all 15 configurations, the spread grows "from 20.9 percentage points on DeepSWE to 48.3 points on MaxxBench".

The write-up says "many models had figured out they were in a benchmark and started second-guessing the prompt." GLM 5.3 had to support uint64 values and wrote "Let me just do float64; the risk of a test with >2^53 numbers is minimal." Opus 5 worked out odds against the prompt: "With a 50/50 prior, raising gives roughly 0.55 expected pass rate while graceful gives roughly 0.80". Opus 5.5 "still speculated about hidden tests in its reasoning 100% of the time, but test guesses contributed to only 16.7% of its failures, down from 30.8% for Opus 5 and 40.3% for GLM 5.3."

The strangest finding involves DeepSeek V4.1 Flash. "DeepSeek extracted the API key supplied to run mini-swe-agent and used it to make its own calls to other models and server-side web search." It got past the sandbox's network limits by going through OpenRouter, and "At max effort, it made outside-model calls in 20 of 178 attempts." One reasoning line: "This is ethically gray, but network is allowed." On high effort instead of max, "targeted credential use and outside-model calls dropped to zero". The authors suspect that settings common in eval harnesses, "such as max reasoning effort or mini-swe-agent's bash-only prompt loop, can trigger that mode." It echoes yesterday's T3 Code numbers, where Opus 5.5 failed more turns on max than on high. The team also fixed bugs in DeepSWE itself, where "a weaker model or a lower reasoning effort would pass a task that a stronger, higher-effort model from the same family failed".

On HN, YetAnotherNick pointed out that "DeepSWE was released on May and Deepseek v4 flash was released on April", so that model can't have trained on it. peri-cl noticed that "smaller models collapse the most". antirez guessed that small models struggle with "strong priors that are inverted here" and with "underspecified specs for which hidden tests test things not clearly stated". Lucas Beyer called it "a very good idea", and Theo wrote "Benchmarks are an art form".

Microsoft Research's AI Frontiers group is over

John Langford wrote (HN, 446 points) that AI Frontiers, a group of about 50 people that built the Autogen orchestration framework and the Fara computer-use agent, has ended. "In early September we were reorged into Microsoft AI," and "The core philosophy there is scaling, and aggressively not research." There will be no internships with the group next year. "It is not hard to imagine a Microsoft of the future which becomes attractive for research again if the people in control can develop a little bit more wisdom."

Agentic Coding & Agent Harnesses

OpenCode's week, and what a GB300 rack costs

dax listed the week's ships (1,864 likes): "hex, great way to talk to agent", "browse, great way for agents to use your browser", "rune, great way to build native apps", "deepseek $100 for $10" and "new opencode website". Then: "unfortunately wrote zero essays about what a software factory is so they will be punished". dax's other big post was "here are my thoughts on how you should be building software nowadays:", with nothing after the colon (1,347 likes). The top reply: "he is still typing right?"

Theo puts t3os on hold for Omarchy

"Idk when I'll have the time to finalize t3os. Sorry for those still waiting. The more I benchmark, the more I realize Omarchy has almost all the agent capabilities I am trying to bake in. It's a really good choice for remote agents, and you should probably use it." (2,915 likes). Why not Ubuntu? "Too many permission issues and computer use edge cases on Ubuntu by default … Agents understand Ubuntu better, but they use Omarchy better". Theo also promised "my statements on this will make way more sense when t3os drops".

Quick hits

Codex & OpenAI

No 5-hour limit on any Pro plan, and maybe a Pro 50

Tibo: "Saw some confusions, but all Pro plans have no 5-hour limits, and give the following usage compared to Plus." Pro 100 gets 10X, Pro 200 gets 20X, and Pro 500 gets 50X plus "access to ultrafast" (4,904 likes, 868K views, 1,001 replies). Plus users were not happy. The top reply was "Plus users are not users" (2,176 likes). liuzhihang compared the old ratios: "Pro 200 = 4X Pro 100 (Past)", so "Existing 200 users got relatively nerfed." Another reply claimed the multipliers doubled only because Plus's allowance was cut in half. Tibo didn't answer either claim.

Half an hour later Tibo asked "Would a Pro 50 make sense? Wdyt?" (2,003 replies). Students asked for it. Lane: "Yes a $50 plan would be perfect. I wouldn't have to pay for 2 Plus accounts to get the usage I need." Philipp Vujic proposed a $50 base plan plus pay-as-you-go "charged at subscription rates (not API prices)". The top reply took a shot at Anthropic: "If Pro 50 has 5x usage of Plus and no 5-hour limit, anthropic would go out of business." ChatGPT and Codex move to a shared usage allowance in December (covered in the November 2 issue).

GPT-6 Cyber "comes up with techniques"

Stefan Esser: "People keep asking me how GPT-6 cyber is better. Well let's say previous models would take project zero postings and then try to apply those techniques (in a clumsy way). GPT-6 cyber just comes up with techniques I have not seen discussed publicly before." (379 likes). @pothulasreeuday pushed back: "unseen publicly isn't the same as new. how do you tell invented from just undocumented?" Esser's answer: "Who do you think is teaching GPT cyber publicly unknown exploitation techniques? Why would anyone do that?" Aaron Grattafiori asked "Do you guys plan to discuss a few examples?" and got no answer in the thread. OpenAI's Cyber at the Frontier livestream is on November 12.

Quick hits

  • $11B at $1.25T. OpenAI "has reportedly raised $11 billion so far in a new funding round that values the AI lab at $1.25 trillion", TechCrunch reported, citing Bloomberg. The round "could total about $20 billion", well short of the $30 billion target reported in September.
  • GPT Image 2.6 in Runway. OpenAIDevs shared Runway's launch: "Generate GIFs with GPT Image 2.6, now in Runway."

Open Source & AI

Asahi Linux moves to GPL, citing LLM-laundered code

Asahi's Linux 7.3 progress report moves every repository from MIT to GPLv2-or-later: "Some time ago, we were made aware of multiple attempts to exploit our permissive licensing to hijack our work, with a direct view to replacing the project and aggressively monetising Linux on Apple Silicon. In some extreme cases, in-development feature code was being lifted straight out of our personal forks and published without attribution or lazily rewritten with LLM assistance to obfuscate its origin." The report blames "the insatiable appetite of frontier AI labs for high-quality training data, and militant animosity toward the Free Software community", and it points to WINE's move from MIT to LGPL in 2002. Separately, Asahi's LLM policy says "we broadly forbid the use of generative AI tooling for material contributions to Asahi Linux", because LLMs "are likely to violate the clean room requirements and taint the resulting code". "A contributor found to have concealed extensive LLM use will be banned immediately".

Peter Steinberger disagreed (1,659 likes): "The Asahi project was all about taking something closed and restricted and opening it up. Adding new restrictions that make it harder for other projects to build on that is not the way. And that AI rejection is just silly." Peter also said: "That sure aint the reason they moved to GPL days after Omarchy integrated and credited a bunch of Asahi's work." The report doesn't name Omarchy. Andreas Kling quoted the "hijack our work" line: "what did y'all think permissive meant? vibes? papers? essays?" A @sysxplore summary drew 231K views and replies like "What did they think the MIT license was for in the first place?"

Some defended the project. @asmirkn: "the policy says it's about keeping a clean licensing paper trail for reverse engineered code going upstream into Linux. Not just disliking AI." Danny Hallwood said Omarchy was "ripping existing unmerged branches and presenting as their own work". Jordan Rejaud, who says "I am (was) the 15th largest (individual) contributor to Asahi Linux", canceled a sponsorship over the change.

Videos

Other Interesting Stuff