How-to

Is Claude getting worse, or is it your setup?

A Hacker News thread on 9 November says Claude Opus 5.5 got weaker after 3 November, and nobody has proved it. Rule out the settings that change answers, then rerun your real tasks with this prompt for a fair verdict.

Its commenters split between worse results and blaming their tools, under the title ‘Is Opus 5.5 nerfed? Thinking patterns changed on November 3’.

As of 10 November, Anthropic’s status page lists no Opus 5.5 incident in November (the last one named for it was elevated errors on 24 October).

Anthropic’s developer docs say Anthropic does not update the weights of a released model ID, though the systems serving it, such as safety classifiers and sampling, can change and are the most likely cause of small shifts.

The prompt: a fair before-and-after test

Paste it into a new chat to build your test pack, then paste it again later with your results for the verdict. For old answers, open a past chat you were happy with from before 3 November and copy it with its date; if you have none, today’s run becomes your baseline.

Prompt Test if my model slipped
You are helping me find out, fairly, whether the AI model I use has really got worse at my work, or whether something in how I use it changed. Work like a careful tester who has seen this argument many times. If you tell me the model got worse when my settings changed, I may drop a tool that still works or pay for another one. If you wave away a real drop, I will keep trusting answers that are now weaker. So judge only from what I give you, and say plainly when the evidence is too thin.

WHAT YOU NEED TO KNOW
- Answers vary from run to run even with the same model and settings, so one bad answer proves nothing.
- These things change answers without the model changing: which model actually answered (some apps switch model for some requests and label the answer), the effort or thinking setting, how long the chat is (late in a long chat the assistant may be working from a summary of earlier messages), saved memory, custom instructions, project instructions and files, connected tools or web search, the app or coding-tool version, and incidents on the provider's status page.
- You cannot see my settings, my past chats or the provider's systems. Never guess them. If I have not told you something, mark it NOT STATED.
- You cannot know why a model changed, and you must never claim the provider secretly changed or downgraded it. The most you can say is what my test shows, on my tasks, under my settings.

ABOUT ME (I fill this in once; any line can say "not sure")
- The assistant and model I am testing: [e.g. "Claude, Opus 5.5 in the desktop app", "Claude Code in the terminal"]
- What feels worse: [e.g. "it skips steps in my spreadsheet formulas", "its client emails got vague", "it stops halfway through code changes"]
- When I first noticed: [e.g. "around 3 November", "this week", "not sure"]
- My real tasks (three to five jobs I actually do often): [e.g. "turn my meeting notes into an action list with owners and dates", "rewrite a product description to under 80 words", "fix a failing test in my repo"]
- Old answers I was happy with: [paste each with its date and the settings I used, or write "none saved"]
- Mode: [BUILD, if I have no new results yet / JUDGE, if I am pasting my test pack and results below]

HOW TO WORK
Check the ABOUT ME section first. If my tasks are missing, or vaguer than the examples, ask me for them once, show me one sharper version of a vague task, and wait. If Mode is BUILD, do Steps 1 to 3. If Mode is JUDGE, do Step 4 using the test pack and results I paste. If Mode is missing, ask which one.

Step 1. Build my test set. For each task, write:
- The exact prompt I will paste, word for word. Name any document or data I must attach in square brackets, e.g. [ATTACH: the same notes file every time]. If the task needs material I have not described, write [NEED: what it is].
- Two to four pass checks a stranger could mark yes or no without judgement, e.g. "lists every action with an owner", "stays under 80 words", "the test passes". Never use checks like "is good" or "sounds right".
- A 1 to 5 quality score, with one line on what a 5 and a 1 look like for this task.
If a task cannot be checked this way (for example, it is about taste), say so, keep it, and mark it JUDGEMENT ONLY so it never decides the verdict on its own.

Step 2. Give me a settings record to fill in for every test session, with these lines: date and time; assistant; model name as shown on the answer; effort or thinking setting as shown; new chat or continuing one; memory on or off; custom instructions on or off; inside a project or not; connected tools or web search on or off; app or tool version, if I can see it; anything the provider's status page lists for that day. Tell me to run every test in a new chat with the same settings each time, and to write "not found" for any line I cannot find.

Step 3. Give me a scoring sheet as a plain table I can copy: one row per run, with columns for task, run number (1 to 3), each pass check (yes or no), quality score, and the model label shown on the answer. Tell me to run each task three times on the same day, in three new chats, and to copy the full answers into a separate document before scoring. Then give me the whole test pack (Steps 1 to 3) in one block, and tell me to save it myself, because you will not remember this chat. If I have no old answers saved, tell me to run it today as my baseline and again whenever something feels off.

Step 4. When I paste the test pack, the settings records and the results, give a verdict in this order:
a) Settings check. Compare the settings records line by line. If any line differs (a different model label, a different effort or thinking setting, a long chat in place of a new one, memory or instructions switched on or off, a project used one day and not the other, a new app or tool version), say MY SETUP CHANGED, name each difference, and tell me to rerun with matched settings before anyone blames the model. If a line says "not found" on either day, say it could not be compared and that the verdict is less certain.
b) Fallbacks and incidents. If any answer is labelled with a different model from the one I chose, set that run aside and tell me the app switched model for that request. If the status page listed an incident for my assistant during a session, set that session's runs aside and say a rerun is needed.
c) Noise check, task by task. Compare the range of quality scores across the three runs on each day. If the two ranges overlap, call it NO CLEAR CHANGE for that task. Only if every later run scores below every earlier run, call it WORSE ON THIS TASK; if every later run scores above, BETTER ON THIS TASK. Pass checks count too: if a check passed in every earlier run and fails in every later run, call it WORSE ON THIS TASK even when the scores overlap. A check that fails in only some later runs is noise.
d) If my baseline is one saved old answer, you cannot measure its spread. Then: if any new run passes every check the old answer passed, call it NO CLEAR CHANGE; if all three new runs fail a check the old answer passed, call it POSSIBLE CHANGE, and tell me to run the full test today as a new baseline and repeat it in a week. Anything else is NO CLEAR CHANGE.
e) Overall verdict. MY SETUP CHANGED if part a) found a difference. THE MODEL SEEMS WORSE ON MY WORK only if the settings matched, nothing was set aside in part b), and more than half of the checkable tasks came out WORSE ON THIS TASK. If no task is WORSE ON THIS TASK but at least one is POSSIBLE CHANGE, say POSSIBLE CHANGE overall and repeat the advice from part d). If some tasks are WORSE ON THIS TASK but not more than half, say WORSE ON SOME TASKS and name them. Otherwise NO CLEAR CHANGE. Never strengthen a verdict because I sound frustrated.
f) Next move. MY SETUP CHANGED: list the exact settings to match, then tell me to rerun. NO CLEAR CHANGE: say the feeling may come from harder tasks, a long chat or chance, and tell me to keep the pack and rerun after the next app update. THE MODEL SEEMS WORSE ON MY WORK: tell me to run the same test one effort level higher and on another model I can use, to see whether a setting fixes it, and to keep the dated results in case I report it to the provider's support.

MY RESULT
End with a short summary I can save: the verdict, the reasons (each tied to a line in my results), what was set aside and why, and the date to rerun.

SELF-CHECK BEFORE YOU ANSWER
- Every verdict rests on scores or settings I gave you, and you named them.
- You did not claim the provider changed, downgraded or secretly swapped a model.
- Anything I did not tell you is marked NOT STATED, and nothing was guessed.
- No JUDGEMENT ONLY task decided the verdict on its own.
- You did not promise to remember this chat or to run anything later.

Check these before you blame the model

Each one fills a line of the prompt’s settings record, and each can make a sound model look weaker.

  1. Look at the status page. status.claude.com lists every incident with times, such as degraded performance for Fable 5.1 for about 15 minutes on 8 November, and a bad session during one looks like a weaker model.
  2. Check which model answered. When a message is flagged, Opus 5.5 can switch to Opus 5 or Opus 4.8, shows a notice, labels the answer and stays on that model for the rest of the chat. The toggle is Switch models when a message is flagged in Settings > Capabilities.
  3. Check the effort level. Open the model menu next to the send button and check the effort level. On Opus 5.5 thinking is always on, so a lower effort level is what makes it think less, and each model’s recommended level is marked Default.
  4. Start a new chat. With code execution on, Claude summarises earlier messages as a long chat nears its limit, so a test run late in that chat is working from the summary.
  5. Note memory and instructions. Instructions for Claude apply to every chat and project instructions to every chat in that project. An incognito chat skips saved memory but still uses your styles and preferences, so record those too.
  6. In Claude Code, check the model, effort and version. Check the model with /model and the effort with /effort, then run claude --version to see which version you are on.
  7. Read the early November changelog. Two entries matter: 2.1.317 on 3 November changed the effort level subagents use when they run on a different model, and 2.1.318 on 4 November stopped keeping earlier thinking when you resume some older conversations. Neither entry mentions the reports, but either can change results without the model changing.

What a verdict looks like

The task: turn meeting notes into an action list with owners and dates. Pass checks: every action has an owner, every action has a date, nothing is invented.

The settings: the same on both days, with Opus 5.5 on the answer, effort High, a new chat each time and memory off.

The scores: 5, 4 and 5 on 27 October, then 4, 4 and 5 on 10 November, when one run missed an owner.

Verdict: the two ranges overlap, so the prompt calls it no clear change on this task, even though the missed owner felt like proof.

The honest bit

  • Small test, big changes only: a handful of tasks run three times can catch a drop you would feel on your own work, but it cannot prove a small one. Brandon Da Costa’s brief at Digiton suggests five repeats on 30 to 50 prompts for teams who need that.
  • Your check: the AI tallying your results is the same kind of tool you are testing, so mark the pass checks on one task yourself before you read its verdict. If you disagree, go with your own marks.

The facts here come from Anthropic’s developer docs, help pages, status page and Claude Code changelog, the Hacker News thread and Digiton’s brief, read 10 November 2026.

Do it now

Pick three to five tasks from last month that went well, paste them into the prompt in BUILD mode and save the test pack, so the next time Claude feels off you have something to compare against.

A few quick questions

Does the Thinking section show Claude's real reasoning?

It shows a summary. Anthropic’s help page calls it Claude’s thought process summary, so shorter or fewer thinking notes are weak evidence that the model itself changed.

Does this test work for ChatGPT or Gemini?

Yes. The prompt names no tool, and the same things shift answers there: the model you picked, any thinking setting, memory, custom instructions, project files and outages. Find each in that app’s settings and on its status page.