Anthropic deprecated Claude Sonnet 4 and Claude Opus 4 on April 14, set their retirement date at June 15, and shipped Claude Opus 4.7 on April 16. [1] [2] Same week, three actions: kill the old SKUs, ship the new flagship, set the 60-day migration clock. The deprecation notice came without prior public warning. [2]

Opus 4.7 lifted SWE-bench Verified from 80.8% to 87.6% and CursorBench from 58% to 70%, kept pricing flat at $5 input / $25 output per million tokens, and introduced a new xhigh effort level. [1] Claude Design, a visual-prototyping product, launched alongside it. The Pro plan lost Claude Code access. Annualised revenue reportedly reached $30 billion. The Mythos preview model arrived as the next-up. [1]

The contrast with the Tuesday I/O capex spectacle is the operating frame. Google priced a $180-190B 2026 AI capex line on a stage; OpenAI gets the antitrust trial coverage; Anthropic ships in release notes. [1] After June 15, API calls to claude-opus-4-0 and claude-sonnet-4-0 return errors with no fallback. The deprecation calendar is the announcement.

Anthropic has also begun routing Claude Platform directly through AWS. [1] The April 14 deprecation came without prior public warning, prompting the model-welfare argument on X that has followed Anthropic since the Claude 4 system card. The customer cost is the migration; the discourse cost is whether a company that built its public ethics around model welfare can quietly retire a model six months later because newer benchmarks favor a successor.

-- DAVID CHEN, Beijing