Claude Code Quality Postmortem (April 23, 2026): Effort Defaults Reverted, Idle-Session Bug Fixed, Limits Reset

Anthropic published an engineering postmortem on April 23, 2026 after users reported that Claude Code felt worse over parts of March and April. According to Anthropic, the root cause was three separate product-layer changes that affected Claude Code’s perceived intelligence and session continuity — while the Claude API itself was unaffected. (Anthropic engineering postmortem)

This post summarizes the high-signal, user-relevant takeaways for teams using Claude Code day-to-day.

TL;DR

• A default effort change (high → medium) traded intelligence for latency; it was reverted after feedback. (Anthropic engineering postmortem)
• An idle-session optimization bug repeatedly cleared prior reasoning, causing “forgetful” sessions and extra cache misses; it was fixed in v2.1.101. (Anthropic engineering postmortem)
• A verbosity-reduction system prompt hurt coding quality and was reverted in v2.1.116. (Anthropic engineering postmortem)
• Anthropic says all issues were resolved by April 20 (v2.1.116) and usage limits were reset for subscribers as of April 23. (Anthropic engineering postmortem)

1) Effort defaults: medium was the wrong tradeoff

Anthropic says that on March 4, Claude Code changed the default reasoning effort from high to medium to reduce very long tail latencies that made the UI look frozen for some users. (Anthropic engineering postmortem)

After users reported lower perceived intelligence, Anthropic says it reverted this decision on April 7 — defaulting back to high for most models, and to xhigh for Opus 4.7 specifically. (Anthropic engineering postmortem)

What to do

• If Claude Code feels “dumber” on a given day, confirm your current effort level with /effort and bump it back up for hard tasks.
• If you’re using Opus 4.7, note that Anthropic’s stated default is xhigh effort (so “high” may be a downgrade for your workload). (Anthropic engineering postmortem)

2) Idle-session bug: repeated clearing of prior thinking

On March 26, Anthropic shipped an optimization intended to clear older thinking from sessions idle for more than an hour to reduce resume latency. Anthropic says a bug caused this clearing to happen every turn for the rest of the session, which made Claude Code appear forgetful/repetitive and also increased cache misses (making usage limits drain faster). (Anthropic engineering postmortem)

Anthropic says this was fixed on April 10 in v2.1.101. (Anthropic engineering postmortem)

What to do

If you’re coming back to a long-running session and it suddenly loses thread:

• First, update Claude Code to the latest version.
• Second, consider starting a fresh session with a short recap and links to key files (especially if you were idle for long periods during the work). The underlying bug should be fixed, but this remains a useful workflow when resuming weeks-old work.

3) Verbosity prompt change: reverted after eval regression

Anthropic says that on April 16, it added a system prompt instruction to keep text between tool calls to 25 words and final responses to 100 words unless needed. In combination with other changes, Anthropic says this hurt coding quality and was reverted on April 20 (v2.1.116) after a 3% drop in coding-quality evaluations. (Anthropic engineering postmortem)

Subscriber usage limits reset

Anthropic says that while the issues were challenging to identify (because each change affected different slices of traffic on different schedules), all three issues were resolved as of April 20 (v2.1.116) and subscriber usage limits were reset as of April 23. (Anthropic engineering postmortem)

Sources

• Anthropic engineering: “An update on recent Claude Code quality reports” (Apr 23, 2026): https://www.anthropic.com/engineering/april-23-postmortem