Anthropic quietly downgrades Claude Code defaults, tests Pro removal
A detailed post by 김이더 connects four events — adaptive thinking rollout, a default effort downgrade, an AMD director's damning telemetry analysis, and a 12-hour removal of Claude Code from Anthropic's Pro plan page — into a single pattern of cost-cutting under pressure from long-running agent workloads.
Score breakdown
Teams running Claude Code on Pro plans should manually set `CLAUDE_CODE_EFFORT_LEVEL=max` or use `/effort high` to restore pre-March reasoning depth, and should treat Anthropic's pricing signals as an indicator that Pro-tier access to agentic features may be repriced or restricted in the near future.
- 01On April 21, Anthropic removed Claude Code from its Pro plan pricing page and support docs, then reverted the change roughly 12 hours later after public backlash.
- 02Anthropic's Head of Growth Amol Avasare said the removal was a test affecting ~2% of new prosumer signups, but also stated that how people use Claude subscriptions has 'fundamentally changed.'
- 03AMD AI Group Director Stella Laurenzo filed GitHub issue #42796 on April 2, backed by analysis of 6,852 Claude Code sessions, 234,760 tool calls, and 17,871 thinking blocks.
The post by 김이더 reconstructs a timeline of four events that, taken together, suggest Anthropic is struggling with the economics of long-running agentic workloads. On February 9, Anthropic shipped Opus 4.6 alongside adaptive thinking, which lets the model decide per-turn how much reasoning budget to allocate rather than using a fixed budget. On March 3, the default effort level was quietly lowered from High to Medium (85 out of 100). In early March, thinking content redaction was applied broadly. Then on April 2, Stella Laurenzo — AMD's AI Group Director and former Google OpenXLA lead — filed GitHub issue #42796 titled "Claude Code is unusable for complex engineering tasks with the Feb updates," backed by quantitative analysis of 6,852 sessions, 234,760 tool calls, and 17,871 thinking blocks. Her data showed thinking depth dropped an average of 67%, reads-per-edit fell from 6.6 to 2.0 (a 70% decline), and a stop-hook script that had never triggered before March 8 fired 173 times in the following 17 days. Claude Code team member Boris Cherny responded, arguing that thinking content redaction was UI-only and did not affect actual reasoning, but acknowledged the adaptive thinking and effort-level changes, and conceded that adaptive thinking appeared to under-allocate reasoning on certain turns. His suggested workarounds were the `/effort high`, `/effort max` commands, the `CLAUDE_CODE_EFFORT_LEVEL=max` environment variable, and `CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1`.
The post frames the April 21 Pro plan removal as the second, more explicit cost-control lever — following the quieter default degradation.
The post frames the April 21 Pro plan removal as the second, more explicit cost-control lever — following the quieter default degradation. Avasare's public statement that Anthropic's plans "were not designed for this kind of usage" is read not as a 2%-test disclaimer but as a signal of impending structural repricing. The post draws a broader analogy to the economics of AI subscription services versus traditional SaaS: unlike Netflix, where marginal cost per additional user approaches zero, every agent run consumes real GPU time, making heavy users a liability rather than an asset at flat subscription prices. The post notes that Sam Altman publicly acknowledged ChatGPT Pro's $200 plan was loss-making due to usage volume, and that OpenAI, Cursor, and Replit face the same structural constraint. For teams evaluating AI coding tools, the post argues the key signal is not the 12-hour rollback but Avasare's admission that current plan structures cannot sustain current usage patterns.
Key facts
- 01On April 21, Anthropic removed Claude Code from its Pro plan pricing page and support docs, then reverted the change roughly 12 hours later after public backlash.