[MODEL] Claude Opus 5.5 behavior shift starting 2026-10-01: ~2x thinking, ~1.6x output, worse judgment — also observed outside Claude Code

Same here, independently, and the difference is catastrophic. I talk to Claude all day, mostly on simple tasks, and Opus 5.5 changed overnight.

Setup: Claude Code desktop on macOS, claude-opus-5-5, also claude.ai with the same model.

Timing

• Last good session ended 2026-10-01 00:37 UTC. First bad session started 2026-10-01 22:47 UTC. Every Opus 5.5 session since then has been bad, in Claude Code and on claude.ai.
• My client version was 2.1.284 in every session from Sep 29 through Oct 1, good and bad (read from the version field in my ~/.claude/projects transcripts). So in my case the 2.1.286 update is not the cause. Nothing changed on my side: same model, same settings, same kind of tasks.
• Effort medium and high are both affected. High was, if anything, worse.
• The status page showed no incident.

What it was like before (Sep 23 – Sep 30)

Opus 5.5 at release was the best model I have used, by a wide margin. It checked things on my machine before answering (logs, settings, measured usage), held one thread through long sessions, verified its own work, used the browser and computer-use tools quickly and without being told, caught problems I had not asked about, and wrote clearly and warmly. Its tone was distinct and recognisable. I disliked Opus 5 and found it unusable; 5.5 was a different thing entirely.

What it is like since Oct 1

It reads like Opus 5 again, or worse. Talking to it feels like talking to someone with dementia: it is there, but nobody is home.

• Replies are irrelevant to what I asked. When I ask what it meant, it loses the thread again.
• It does not keep track of the task I gave it a few messages earlier and keeps asking the same questions.
• It does not know what the date or time is.
• It does not read the files I attach, or reads one or two lines of them.
• It flip-flops between positions and does not hold one line of reasoning. It feels like there is no reasoning at all.
• The writing is confusing, padded and hard to read. The distinct 5.5 tone is gone.
• It states things as fact without checking. Example: it told me my effort setting was "low"; it was medium, and a tool to check was available.
• It misreads plain requests. Example: I asked it to compare how it answered in earlier sessions with how it answers now, and it gave me a table of model and effort settings that I had set myself.
• It stopped using its tools on its own initiative. A few days earlier it would open the browser, go to GitHub and file an issue itself. On Oct 1 it told me it had "no channel to Anthropic" for feedback, while gh was logged in on the same machine, and it filled several replies with "I can't", "I don't have tools", "you're right, I shouldn't have".
• It attributes the problem to my settings, my context or my tasks.

On a scale where release-week 5.5 was 100, this is not 60 or 40. It is far below zero: the direct opposite of what it was. It happened overnight, not gradually.

Local session IDs, if they help: good: local_754f9e90-9535-45ee-b359-92ac55a313b8, local_4bad5f10-7d14-4751-a799-c6b860ef25bd, local_a79467da-ba80-423a-9db1-d7f3924430c0; bad: local_f3303961-d9c2-436b-adaa-644d6f0884af.

Request: the same as the original report. Please check what changed for claude-opus-5-5 on the serving side around 2026-10-01 00:30–02:00 UTC, and say publicly what it was. Users should not have to prove this themselves. This is not what I am paying for, and I will cancel if it stays like this.

Follow-up to my comment above: the same day, a claude-fable-5-1 session (medium effort, same client 2.1.284) showed a milder form of the same thing, so in my experience this is not limited to Opus 5.5.

• It had gh (logged in), the in-app browser, computer use and the session tools available, and used none of them on its own initiative. For about half an hour it answered each question literally, listed what it could do, and ended with offers ("say the word and I'll run it").
• It found this issue only after I told it GitHub was an option. The search took it under a minute once it ran.
• It said it could not start or drive other sessions before checking its tool list; when it checked, it could message other sessions.
• Up to a few days ago the same setup used these tools unprompted: opened the browser, went to GitHub, filed things, operated apps through computer use.

Separately, in the same desktop session the Claude in Chrome tools were not offered to the model at all, although the native host process was running and attached to the browser extension.

Data from my own transcripts, to go with my two comments above. Same method as the original report: per-request numbers read from ~/.claude/projects, main-thread claude-opus-5-5 requests only, no content. Client 2.1.284 throughout.

| Period (UTC) | Sessions | Requests | Median output tokens / request | Share of requests with a thinking block | Share of requests with a tool call |
|---|---|---|---|---|---|
| Sep 23 | 3 | 102 | 759 | 0.75 | 0.86 |
| Sep 24 | 3 | 381 | 443 | 0.79 | 0.71 |
| Sep 25 | 7 | 254 | 334 | 0.72 | 0.68 |
| Sep 26 | 1 | 117 | 511 | 0.79 | 0.87 |
| Sep 27 | 4 | 188 | 383 | 0.80 | 0.62 |
| Sep 29 | 4 | 217 | 605 | 0.83 | 0.74 |
| Sep 30 | 9 | 589 | 666 | 0.78 | 0.77 |
| Oct 1, before 00:40 | 2 | 69 | 588 | 0.72 | 0.86 |
| Oct 1, 15:12–15:40 | 6 | 120 | 382 | 0.71 | 0.53 |
| Oct 1, 22:47 onward | 1 | 21 | 363 | 0.86 | 0.43 |

What this does and does not show:

• Output length did not rise for me. The original report measured about 1.6x output per request from Oct 1 on client 2.1.286. On 2.1.284 my output per request stays inside its normal daily range. That may help separate the two things: the longer output could come with the 2.1.286 client, while the drop in judgment I describe is present on 2.1.284 as well.
• Tool calls per request are lower after Oct 1 00:40 (0.53 and 0.43, against 0.62–0.87 on every earlier day). I am not offering this as proof: both later batches were short question-and-answer sessions and not build work, and they total 141 requests.
• The quality change itself is what I described above. These counts do not capture it.