Anthropic open-sources FlashProve: prove which model answered you

A zkSNARK catches silent model swaps by any router or provider — verify on a CPU. The catch: proving costs ~70–95× GPU. Plus OpenAI's Nov 30 deprecation wave.

Nowline NOV 4 5:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Prove which model actually served your request

    Anthropic open-sourced FlashProve (Apache 2.0): a zkSNARK that lets a provider prove a reply came from a specific committed set of weights — without revealing the weights. Anyone can verify the proof on an ordinary CPU in ~17 seconds. The prover, a fast C++ verifier, Lean-checked soundness proofs and the protocol spec are all on GitHub.

  • Why you'd care: catching silent model swaps

    Route through an aggregator or a cheap third-party endpoint and you normally can't tell whether you got the model you paid for. FlashProve detects substituted models and tampered token streams even behind third-party routing — the groundwork for auditor and regulator tooling that checks which model is really being served.

  • The catch: 70–95× the GPU cost

    Proving isn't cheap yet: roughly 70–95× the GPU time of plain vLLM, ~110 tokens/sec on a single H100, ~19 MB proof files, and a 1,024-token context cap. It's a research prototype (greedy decoding only) that ships marked 'not maintained.' Verifying stays cheap; generating the proof is the bottleneck.

  • OpenAI's Nov 30 deprecation wave lands

    Mark the calendar: Agent Builder, reusable prompt objects (the v1/prompts API) and the Evals dashboard/API all shut down Nov 30. On Dec 1, gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest are removed — migrate to gpt-image-2.5-sunburst or -flare. Existing evals already went read-only Oct 31.