Story perspectives

Anthropic Repairs Claude Code, Users Cite Gas-lighting, Vulnerabilities

4/25/2026

Summary

Anthropic Restores Claude

Anthropic said three changes—Mar 4 reasoning-effort cut, Mar 26 bug erasing history, Apr 16 25-word prompt cap—caused Claude Code’s decline.
Anthropic fixed all three issues and reset usage limits for subscribers on Apr 23.
Users cancelled subscriptions and accused the company of gas-lighting about the tool’s performance.
Veracode found Claude Opus 4.7 introduced vulnerabilities in 52% of coding tasks, versus about 30% for OpenAI models.

Summary

Anthropic Tightens Controls

Anthropic will require more staff to run the public Claude Code build, tighten system-prompt change controls, run per-model evaluations and ablations, and use soak periods, gradual rollouts, and expanded evaluation suites to detect intelligence-latency trade-offs before releases.
Anthropic created the @ClaudeDevs X account and will share GitHub threads alongside X posts for transparency.
Anthropic updated CLAUDE.md with guidance to gate model-specific changes and is upgrading the Code Review tool.

1 of 2

Anthropic Restores Claude

Anthropic said three changes—Mar 4 reasoning-effort cut, Mar 26 bug erasing history, Apr 16 25-word prompt cap—caused Claude Code’s decline.
Anthropic fixed all three issues and reset usage limits for subscribers on Apr 23.
Users cancelled subscriptions and accused the company of gas-lighting about the tool’s performance.
Veracode found Claude Opus 4.7 introduced vulnerabilities in 52% of coding tasks, versus about 30% for OpenAI models.

1 / 2