Opus 5.5 users say it got worse since Nov 3; a daily tracker finds no change

Anthropic has not responded to a Hacker News thread, which climbed from 128 to 379 points, in which Claude Code users claim Opus 5.5 degraded after November 3, while the open-source daily tracker LiveNerf reports "no change detected" as of day 46, counting a change only as a shift of at least 3 points across two consecutive windows. Separately, unLAP Data Research's MaxxBench, built by altering 89 DeepSWE tasks, split models that had tied near 74%: Opus 5.5 scored 76.4% and Gemini 3.8 Flash 53.4%. Security firm Apiiro also reported 1,122 public GitHub files exposing passwords or keys that Claude Code saved via its "Yes, and don't ask again" approval into .claude/settings.local.json.

Opus 5.5 users say it got worse since Nov 3; a daily tracker finds no change Anthropic hasn't answered the complaint, which has passed half a million views on X. Four more stories inside. Today leads with the question Claude Code users spent the weekend asking: did the model they pay for quietly get worse? The thread asking it climbed from 128 to 379 points on Hacker News today, so we checked it against the one tracker built to answer it. Also today: an altered version of a popular coding benchmark splits two models that tied on it by 23 points. A security firm found 1,122 public GitHub files exposing secrets that Claude Code's permission prompts saved. Claude Code's cloud sessions now set up your repo for you. And Goldman Sachs cut most of its traders and bankers off Claude Code. What changed this week Users say Opus 5.5 got worse since Nov 3; tracker finds no change X user Shane Melaugh asked "is Opus 5.5 nerfed?", meaning quietly made worse, after having Google's Gemini 4 Argon model analyze 1,400 Claude Code sessions: on November 3, he says, "the way it does reasoning took a sharp turn." A front-page Hacker News thread fills with matching reports of sloppier work and faster limit burn. Anthropic hasn't replied there. The one daily measurement we know of doesn't see it. LiveNerf, an open-source tracker that has rerun a fixed test set on Opus 5.5 every day since launch, reports "no change detected" as of day 46. A change only counts there as "a significant shift of at least 3 points in two consecutive windows." One commenter suspects only the reasoning summary Claude Code displays changed "It's just a new summarizer" , which a transcript analysis would catch and a fixed test wouldn't. Anthropic hasn't confirmed either. If your sessions feel worse, rerun a task you saved from before Nov 3. Altered benchmark tasks split Opus 5.5 from Gemini 3.8 Flash by 23 points Claude Opus 5.5 and Gemini 3.8 Flash both post about 74% on DeepSWE, a popular coding benchmark. Researchers at unLAP Data Research built MaxxBench by changing 89 of its tasks the way a code reviewer might, such as requiring lowercase filenames. The tie broke. Opus 5.5 scored 76.4%, Gemini 3.8 Flash 53.4%. Reading the agents' logs, they found models "had figured out they were in a benchmark and started second-guessing the prompt." Models that tie on a leaderboard can still handle your changed requirements very differently. Apiiro finds Claude Code approvals left passwords in 1,122 public GitHub files Security firm Apiiro reports 1,122 public GitHub files exposing passwords or keys that Claude Code saved. Click "Yes, and don't ask again" and Claude Code stores the whole command, secret included, in .claude/settings.local.json. It hides that file through your machine's global git excludes, not the repo's .gitignore. Boris Cherny of the Claude Code team said Anthropic is "looking into what we can do to prevent this." Search your repos for that file. Rotate any key in it. Claude Code cloud sessions now set up your repo — and keep your keys from Claude Anthropic's Claude Code developer account said on Nov 6 that Claude Code cloud sessions on Pro, Max and Ultra now work out what a repo needs to run dependencies, network access, API keys and save that environment for the next session. Keys get special handling. They go in as what Anthropic's docs call network secrets, stored on the environment "so Claude can call that API from any session in the environment without seeing the key." No saved command holds the key. Goldman Sachs limits Claude Code to engineers, cutting off traders and bankers Goldman Sachs limits Anthropic's Claude Code and GitHub's coding tools Bloomberg's "Microsoft's GitHub" to its engineers, under internal guidance updated in October, Bloomberg reported Nov 2. Most traders and bankers lost access. Strats, who mix trading and engineering, kept it. Traders had used the agents "to monitor client trading flows, model market moves and positions." Goldman is also a lead underwriter on Anthropic's IPO. If you hand agents to non-engineers, expect a similar review. Also worth your time - A longtime Claude Code user explains moving to the Pi coding harness, citing 244 Claude Code releases this year that mostly changed its prompts https://caleb.software/posts/claude-to-pi.html?ref=thenewway.ai - Anthropic's engineering post on how "dreaming" lets Claude reorganize its memory between tasks https://www.anthropic.com/research/components-of-agent-memory?ref=thenewway.ai - Cursor's agents now take realtime voice calls in the IDE, on the web and on iOS https://cursor.com/changelog/voice-chat?ref=thenewway.ai - Simon Willison hides agent-written PR descriptions behind a collapsible block https://simonwillison.net/2026/Nov/9/details-summary/?ref=thenewway.ai Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one. The New Way is written with AI. It gathers the day's stories, checks them against their sources and drafts every summary. A person decides what runs and reviews every issue before we hit send.