Anthropic Compute Crisis: Claude Code Removed from Pro, OpenClaw Banned, Opus 4.7 Tokenizer Inflated
A comprehensive analysis by Matthew Berman documents Anthropic's unfolding compute crisis in detail. Claude Code was quietly removed from the Pro tier as a 2% A/B test (7% of power users affected by peak-hour limits per Anthropic's own communications). Third-party agentic harness usage — including OpenClaw and Hermes — was banned on personal subscription quotas effective April 3 with less than 24 hours notice. Opus 4.7 shipped with two quota-inflation vectors: a new tokenizer consuming 1.0–1.35× more tokens on identical input, and higher thinking-token output at elevated effort levels — both confirmed by Anthropic staff. Root cause: Dario Amodei's 2024 decision not to pre-commit to trillion-dollar compute builds, which proved too conservative as agentic coding demand surged past projections.
Why It Matters
OpenAI has converted each Anthropic restriction into a communications win — resetting Codex usage limits at 3M weekly users and pledging "transparency and trust." The strategic pressure on Anthropic is real: it cannot raise prices while Codex offers comparable quality at $20/month, forcing half-measures that erode developer trust. Watch whether the AWS Trainium deal (separately announced) closes the compute gap before user attrition accelerates. Full analysis: Matthew Berman.