Is Claude Code getting worse?

Claude Code is a product wrapped around a model, and most of what feels like the model getting worse is the wrapper: effort defaults, context management, prompt caching, usage limits, and the occasional bug in all four. Anthropic’s April 2026 postmortem is the proof: six weeks of “Claude Code got dumb” reports, model and API unaffected, three client-side causes.

That is good news, because the wrapper is the part you can see and change. Start with the table.

Worth knowing in detail, because every future “is Claude Code worse?” thread will rhyme with it. From Anthropic’s own write-up:

• Early 2026: Opus 4.6 ships in Claude Code with default effort high. Some users see very long thinking and a frozen-looking UI.
• Shortly after: the default is lowered to medium to cut latency. Users start reporting that Claude feels less intelligent. Several UI changes try to make the effort setting more visible; most users keep the default.
• March 26: a prompt-caching change meant to clear old reasoning once after an idle hour instead clears it on every turn for the rest of the session. Forgetfulness, repetition and strange tool choices follow, plus extra cache misses that drain usage limits.
• April 7: effort default reverted. Opus 4.7 now defaults to xhigh, other models to high.
• April 10: caching bug fixed in v2.1.101.
• April 16: a system-prompt line shipped alongside Opus 4.7 to tame its verbosity (“keep text between tool calls to ≤25 words…”) turns out to cost a noticeable amount of intelligence.
• April 20: all three issues resolved as of v2.1.116.
• April 23: postmortem published; usage limits reset for all subscribers.

All from the official command reference:

• /effort — sets the effort level (low through xhigh, plus max, ultracode and auto); /effort status prints the current one. If you have never run it, you are on a default someone else chose.
• /context — shows what is filling the window as a grid, with warnings when you are near or over the limit. Run it before you blame the model for forgetting.
• /compact — summarizes the conversation to free space, optionally with focus instructions; the docs spell out what survives it and what does not.
• /clear — a genuinely new conversation. The single most effective fix on this page, and the one people resist the most.

Claude Code meters usage in five-hour and weekly windows, and the weekly one moved this month. A promotion that raised weekly limits by 50% ran from May 13 through September 13, 2026. From September 14, weekly limits for Pro, Max, Team and seat-based Enterprise plans are 25% higher than before the promotion, which is 25% above the original but roughly a sixth below what four months of habit had calibrated everyone to.

If your week ran out early after mid-September, that is the mechanism. The model did not change; the meter did. It is worth separating the two in your head before you vote, because “I hit my limit” and “Claude got dumb” both land as a 12 on the dial and only one of them tells anyone anything.

Cheap ways to turn a feeling into a report:

1. Keep a control task. One prompt in one repo that Claude Code handled well before. Run it fresh (/clear) at the same effort on the same model. Pass or fail, you now know something.
2. Record the version. Note which Claude Code release you were on when things worked, so a regression has a suspect.
3. Check the two status sources. Anthropic’s for outages, this one for whether everyone else is having your day too.
4. Vote with specifics. Model, effort, version, what it did. A comment like “Sonnet 5, v2.1.2xx, medium effort, rewrote the test harness instead of fixing the test” is worth twenty “it’s cooked”s.