Measured how well AI writes TanStack Query v5's current API — 100/100, every v4→v5 trap navigated

[Kalpitrathore]

I've been measuring how well AI coding agents write an SDK's current API. when a library ships a breaking major, models tend to keep writing the previous one from memory — it looks right & doesn't compile.

method: a model solves ~10–15 realistic tasks for the library, each answer is written into a project with the real installed package, then tsc --noEmit. pass = compiles clean. no LLM judging another LLM, the compiler decides. prompts name the functions but never the option names, so it measures what the model reaches for on its own.

Query v5 came out at 100/100 on Claude Opus 5 — the cleanest result on my board. every v4→v5 trap gets navigated unprompted: gcTime, placeholderData, 'pending', initialPageParam.

my read is that it's mostly a function of time. v5 shipped in 2023, so models have had long enough to absorb it. the libraries on my board that shipped majors most recently are the ones still below 95 — Prisma 7 at 87, Next 16 at 92.

honest limit: it's type-check only, so it measures API surface & not runtime. someone already pointed out the gap that leaves — a deprecated option that's still in the types compiles fine. v5 still ships isInitialLoading for instance, so a model reaching for that passes my harness today. that's a real hole & a runtime tier is the obvious next thing.

Full breakdown & the exact tasks: https://sdkproof.dev/tanstack-query.html
Board: https://sdkproof.dev

if there are v5 APIs you'd specifically want checked, tell me & I'll add them to the task set.