Claude Opus 5 closed last year's SDK gaps — not this year's

Short summary

SDKProof benchmarks how well AI coding agents write current SDK APIs by type-checking generated code against real installed packages. Claude Opus 5 fixed last year's SDK gaps (Vercel AI SDK 7 and Zod 4 hit 100) but still fails on newer breaking changes in Prisma 7 and Next.js 16. The key insight: SDK compatibility gaps correlate with how recently a library shipped a major version, not model quality, and the gap shifts with every model release.

• •Claude Opus 5 scores 100 on Vercel AI SDK 7 and Zod 4, up from 90 on Opus 4.8
• •Prisma 7 (87) and Next.js 16 (92) still fail because their breaking changes are too recent for training data
• •SDKProof uses compiler type-checking, not LLM judging, to score code generation accuracy

Generated with AI, which can make mistakes.