Why does Claude seem dumb?

Most of the time Claude seems dumb because the conversation got long, a setting changed underneath you, or you are being throttled. Some of the time it is because Anthropic shipped a bug, and they have a track record of admitting that, eventually. Almost none of the time is it because the model was secretly made worse.

The rest of this page is the long answer. If you want the crowd’s answer instead, it is right here:

1. Context rot

Anthropic’s own documentation puts it plainly: “As token count grows, accuracy and recall degrade, a phenomenon known as context rot.” A conversation that was brilliant at message five is a different beast at message eighty, even on the same model, with the same settings, on the same day. The instruction you gave at the top is still in there; it just weighs less.

2. The window is full and things are being compacted

Every model has a hard context limit. Long, tool-heavy sessions hit it, and when they do, older content is summarized or dropped so the conversation can continue. Anthropic recommends compaction as the primary strategy for long sessions, which means “Claude forgot the thing from an hour ago” is frequently working as designed.

3. It never had the context in the first place

Agentic work is only as good as what is actually in the window: the files that were read, the tool results that came back, the error that was pasted in full instead of paraphrased. Anthropic calls this context engineering, and their guidance is to find the smallest set of high-signal tokens that gets the job done rather than dumping everything in.

4. Effort is set lower than you think

The effort parameter controls how many tokens Claude spends on a response, trading thoroughness for speed and cost. It is a dial, not a quality guarantee, and the default position of that dial is chosen by whatever app you are using. In early 2026 Claude Code moved its default from high to medium to cut latency; weeks of reports that Claude felt less intelligent followed before it was reverted on April 7. That is the canonical example of reason four.

5. Extended thinking is off, or budgeted too low

Thinking is where the hard reasoning happens. On the API it is opt-in and budgeted; in apps it is a toggle. A model that is not allowed to think through a problem will happily give you the first plausible answer instead, and that reads as dumb.

6. You are on a different model than yesterday

Apps pick defaults, defaults change, and the lineup moves fast: Anthropic’s current documentation lists Opus 5.5 as the recommended default for most work, with Fable 5.1 for the hardest reasoning and Sonnet 5 and Haiku 4.5 below it. A workflow tuned for one tier will feel off on another, in both directions.

7. There is an incident, whether or not the status page says so

Between August 5 and September 18, 2025, three separate infrastructure bugs intermittently degraded Claude’s responses. One misrouted up to 16% of Sonnet 4 requests at its worst hour; another produced stray Thai or Chinese characters in English replies; a third was a TPU compiler bug. Anthropic published a full postmortem. None of it looked like a classic outage.

8. The app changed, not the model

The April 2026 Claude Code postmortem is the textbook case: the API was unaffected, but three product-layer changes (the effort default above, a caching bug that wiped Claude’s reasoning history every turn instead of once, and a system-prompt line capping response length) added up to six weeks of “Claude Code got dumb” reports. All three were fixed by April 20 (v2.1.116).

9. You are being rate-limited or usage-limited

Claude Code, claude.ai and the API all meter usage, and the limits move. A 50% weekly-limit promotion ran from May 13 to September 13, 2026, after which limits settled at 25% above the pre-promotion level, which is also a drop of a sixth from what people had grown used to. A throttled, retrying, half-finished session feels exactly like a dumb one.

In both postmortems Anthropic notes that the early reports were “difficult to distinguish from normal variation in user feedback.” Read that charitably: some fraction of “it got dumber” is a run of hard problems, a bad mood, or the model being the same while your expectations went up. Read it uncharitably: their detection lagged for weeks, twice.

Either way, the fix is the same. Stop arguing from one conversation and look at a number. That is what this site is for: a daily score, a history you can see a slope in, and a feed full of other people’s receipts.

In order of how often they work:

1. New conversation (/clear or /compact).
2. Set effort and thinking explicitly; do not trust the default.
3. Pick the model explicitly.
4. Put stable instructions in a Project or CLAUDE.md.
5. Give it the actual file, the actual error, the actual document.
6. Check usage limits and the date any limit last changed.
7. Check the status page.
8. Update the app or CLI.
9. Vote, so the next person searching this gets a better answer than you did.