Skip to main content

When Claude Feels Worse Than Before: What Anthropic's Official Postmortems Say and What to Check Before You Cancel

A guide for anyone wondering whether Claude has gotten worse and whether to cancel. It draws on two quality postmortems Anthropic published itself and lays out what to check first.

By
🌐 This article was machine-translated and may contain inaccuracies. Read the Korean original if in doubt.

The feeling that Claude has gotten worse is not just imagination. Anthropic has published two official postmortems (a postmortem is a report written after an incident to explain what went wrong) that confirm real quality problems. In both cases the reports attribute the problem to service-side bugs and configuration changes, not to the model slowly declining, and say they were fixed. Here is a step-by-step guide to what to check before you cancel.

🟢 Current lineup (November 2026): Claude Opus 5.5 / Claude Sonnet 5.5 / Claude Haiku 5.5 (higher tier: Claude Fable 5.2). · model notice · Fable subscription
🔵 Current lineup (November 2026): Claude Opus 5.5 / Claude Sonnet 5.5 / Claude Haiku 5.5 (higher tier: Claude Fable 5.2). Parts of this article may still describe an earlier model.

Fable subscription (updated November 10, 2026): Claude Fable 5.2, released November 9, 2026, is the current Fable model; Fable 5.1 and Fable 5 are now legacy. Plan terms are the same for all three — Max and Ultra plans and premium seats on Team and seat-based Enterprise plans include Fable at up to 50% of the weekly usage limit; Pro plans and standard seats use usage credits (not available on Free). As a launch offer, Fable models are also included for Pro plans and Team standard seats at up to 50% of the weekly limit through November 16, 2026 (PT). API pricing is $8/M input and $40/M output tokens for Fable 5.2, and $10/$50 for Fable 5.1 and 5. Some cybersecurity and biology requests may be answered by an Opus model due to safety classifiers. See the Fable 5.2 launch news and the Fable 5.2 guide for details.

Aug-Sep 2025 report Cause: 3 infrastructure bugs 1. Requests routed to wrong servers 2. Stray characters in output 3. Compiler bug dropped top token Affected regular API and app use Mar-Apr 2026 report Cause: 3 Claude Code changes 1. Default effort high to medium 2. Idle-session thinking cleared often 3. Verbosity-reduction prompt API not affected

Has Claude actually gotten worse at times?

Yes. For certain periods quality really did drop, and Anthropic acknowledged it. The two cases in the diagram break down as follows.

  • 2025: three infrastructure bugs — According to the report, some Sonnet 4 requests were routed to servers set up for a different context window (starting August 5). A misconfiguration on TPU servers (Google-designed AI chips) put stray characters, such as text in other languages, into answers (starting August 25). A compiler bug could drop the most probable token. The fixes or rollbacks (rollback — returning to the previous version) landed in early to mid September.
  • 2026: three Claude Code changes — These affected Claude Code (a coding tool that runs in the terminal), the Agent SDK, and Cowork. The report states the API was not impacted. On March 4 the default effort (effort — a setting for how deeply Claude thinks) went from high to medium (reverted April 7). On March 26 a cache optimization meant to clear old thinking once instead cleared it every turn (fixed April 10). On April 16 a system prompt (system prompt — instructions an app gives Claude in advance) asking for shorter answers hurt coding quality (reverted April 20). The report names Sonnet 4.6 and Opus 4.6 as affected (the third issue also Opus 4.7) and says all three were resolved in v2.1.116, shipped April 20.

Note that this does not explain every complaint anyone has ever had. These facts apply only to the periods and products the reports cover.

Does Anthropic lower quality when demand is high?

Officially, no. The 2025 report states plainly that Anthropic never reduces model quality due to demand, time of day, or server load, and it attributes those problems solely to the infrastructure bugs above.

The report also admits Anthropic's own checks missed the problem at first. According to the post, its evaluations (tests used to check a model) did not catch the degradation, and privacy controls limit engineers' access to user conversations, so user feedback helped narrow down the causes.

What should you check before canceling?

Work through the four steps below to separate a service-side problem from a usage-side one.

1. Status Official status page 2. Retest Same prompt, new chat 3. Settings Model and version 4. Feedback With a repro example
  1. Check the official status page — see whether anything is open right now on status.claude.com. An active incident points to a service-side cause.
  2. Ask the same question in a new chat — if answers are only bad deep inside a very long conversation, the length of that conversation may be the cause (a general rule of thumb, not a cause confirmed in the reports above). If a fresh chat is also bad, a service-side issue becomes more likely.
  3. Check the model and app version — confirm which model is selected on screen. If you use Claude Code, make sure it is up to date; the 2026 fixes shipped as updates (for example v2.1.116).
  4. Send feedback — in the apps use the thumbs-down button on a response; in Claude Code use the /feedback command. Include a reproducible example: what you asked and what came back. Exact button names can change with updates, so check what your screen shows.

What success looks like: if a new chat gives a good answer to the same question, the problem was probably that conversation's flow. If it still fails: if nothing shows on the status page and the problem repeats in new chats, collect examples and send them as feedback. Comparing other services in the meantime is your call.

How did Anthropic respond?

According to the reports, the response was fixes, prevention commitments, and compensation.

  • Fixes — as listed above, each issue was rolled back or patched.
  • Prevention — the 2026 report says Anthropic will run a broad suite of per-model evaluations for every Claude Code system prompt change and use gradual rollouts for changes that could reduce intelligence. The 2025 report also plans continuous quality evaluation on production systems.
  • Compensation — as of April 23, 2026, it reset usage limits for all subscribers. For limits in general, see Claude usage limits explained.

자주 묻는 질문

Q. Does Claude get worse on its own over time?
The official reports do not show that. What they confirm is specific bugs and configuration changes in particular periods, each identified and fixed.

Q. Did the regular chat app suffer from the 2026 issues?
The report names Claude Code, the Agent SDK, and Cowork as affected and says the API was not. I could not confirm whether the regular chat app was included, so I won't claim either way.

Q. Should I cancel anyway?
That depends on how you use Claude and how satisfied you are, so there is no single answer. Running the four steps above first helps you tell a service problem from a usage problem.

Summary

Quality drops did happen and are documented in official reports, but the stated causes were bugs and configuration changes rather than demand-based throttling, and they were fixed. You can sort out whether your problem is similar with the status page, a new-chat retest, and a version check. Sources: Anthropic's engineering blog — A postmortem of three recent issues (September 17, 2025) and An update on recent Claude Code quality reports (April 23, 2026), both re-read on November 8, 2026. No newer postmortem appeared in the engineering blog index as of that date.

Was this helpful?

Keep reading