Daily AI Builder Brief - April 25, 2026

Executive Summary

Anthropic's April 23 engineering postmortem names three changes that quietly degraded Claude Code from March through mid-April.
March 4 change dropped default reasoning effort from "high" to "medium" to shave latency. Anthropic now calls that the wrong tradeoff.
March 26 idle-session bug discarded the model's thinking every turn instead of once, driving the "forgetful and repetitive" reports.
April 16 system prompt added a 25-word response cap between tool calls. Reverted April 20. Coding quality recovered.
All three issues shipped fixed in v2.1.116; usage limits reset for all subscribers on April 23.
Claude Code, the Agent SDK, and Cowork were affected. The raw API wasn't.
Election safeguards update lands: Opus 4.7 and Sonnet 4.6 score 95 to 96% on internal political neutrality benchmarks and 99.8 to 100% on appropriate election handling.
No new Claude Code release in the last 48 hours. 2.1.119 (April 23) is still current. MCP spec is still 2025-11-25.

Claude Code

No new release since 2.1.119

Latest version is still 2.1.119 from April 23. If you're running 2.1.115 or earlier, upgrade: v2.1.116 is the cutoff that contains the postmortem fixes (reasoning-effort default restored, idle-session caching bug patched, 25-word system prompt cap reverted). Changelog

Anthropic

Claude Code quality postmortem (April 23)

Three identified and fixed: (1) March 4 cut default reasoning effort from high to medium for latency; (2) March 26 cache change discarded prior thinking every turn after an hour idle, instead of once; (3) April 16 system prompt capped responses at 25 words between tool calls. The API itself was untouched. Subscriber limits got reset April 23. If you script Claude Code in CI and saw flatter output last month, this is why. Postmortem

Election safeguards update (April 24)

Opus 4.7 and Sonnet 4.6 hit 95 to 96% on internal political-neutrality benchmarks and 99.8 to 100% appropriate handling on election queries. New autonomous-operation testing showed only the newest models could finish more than half of multi-step influence-op tasks with safeguards stripped, so red-team budgets aren't done. Election banners route users to TurboVote; web search triggers on 92 to 95% of relevant queries. Mostly consumer-facing. No API policy changes. Update

MCP Ecosystem

No new spec or major server release

Spec is still 2025-11-25. Working groups remain focused on streamable HTTP scaling, audit trails, SSO auth, and .well-known capability metadata for registries and crawlers. Nothing cut in the last 48 hours. MCP Blog

Broader AI

Nothing today.

What This Means For Builders

Pin Claude Code to v2.1.116 or later in any pipeline. Older versions still carry the three-bug stack the postmortem describes.
If output quality drifted in March or early April and you blamed your prompts, the postmortem says it wasn't you. Re-baseline benchmarks against current versions before tuning further.
The raw API wasn't affected. Direct API integrations don't need code changes from this incident.