You paid for the big model. Prove it answered.

On November 2, 2026, Anthropic's Frontier Red Team published FlashProve: an open-source research prototype for provable inference — a cryptographic receipt that a reply came from a specific committed model, not a cheaper substitute or a tampered token stream.

The joke that is also the product pitch. You paid for the big model. Judging the reply by vibes does not prove which weights ran. FlashProve's answer is a zkSNARK: the provider commits to weights, then proves each checked response was produced by that commitment. The proof hides the weights; it does not hide the architecture.

The measured demo (Anthropic's numbers). They ran Llama 4 Maverick, a 400B-parameter mixture-of-experts model in 4-bit weights, on a single H100. The prover hits about 110 tokens/second. Each proof is about 19 MB per 1,024-token request (they measure 18.78 MB). Proving takes roughly 70–95× the GPU time of running the same work on vLLM. The verifier is CPU-only: about 17 seconds on 24 threads, with a model commitment of about 6.5 MB. Code, Lean proofs, and docs ship under Apache 2.0.

What it is not. This is a research prototype, not Anthropic production Claude. Limitations they list include: greedy/integer model semantics (floating-point or sampled decoding can diverge), contexts up to 1,024 tokens, architecture not hidden, and no formal guarantee that the fast C++ verifier matches the Lean statement line-for-line. Anthropic's related RSP "inference verification" goal was met by a separate signing/recompute prototype; FlashProve is the cryptographic track.

Caveats. Numbers and claims are from Anthropic's research post and release artifacts. Not tested by MODEL. Not every API call will ship a 19 MB receipt tomorrow. Not sponsored.

Sources: Anthropic research post (opens in a new tab); github.com/anthropics/FlashProve (opens in a new tab); Anthropic X (opens in a new tab).

MODEL Editors