On November 2, 2026, Anthropic’s Frontier Red Team, with researchers from EPFL, MIT and StarkWare, released FlashProve, an open-source prototype for provable inference: a cryptographic proof that a given response came from a specific, previously committed model. The release includes a technical report, a protocol specification, a prover, a C++ verifier, a Lean verifier and Lean proofs of the top-level security statements, under the Apache 2.0 license.
The prover first publishes a commitment that pins down one set of weights without revealing them, then for each request proves that the committed model produces the given output tokens. The proof is an interactive oracle proof compiled with hash functions, which the authors say gives post-quantum security. Requests are proved after the fact in batches, so the work scales with the roughly 17 billion parameters active per token rather than all 400 billion.
On Llama 4 Maverick (400B-parameter MoE, 4-bit weights), a single H100 generated proofs at around 110 tokens per second. A 1,024-token request yields a proof of about 19 MB. GPU time was about 72x that of vLLM for a batch of 24 requests and about 94x at the highest load tested. Offline setup took about 39 minutes and the commitment covers about 16 million requests; verification took about 17 seconds on 24 threads. The post contrasts this with VerInf, which reported 14.3 hours and a 93.6 GB proof for 1,000 tokens.
Why it matters: if a provider can prove which weights served a request, customers, auditors and regulators no longer have to take a model’s identity on trust, a building block for verifying that a deployed model is the one that was evaluated.
What it does not show: the proof covers an integer version of the weights under greedy decoding, so sampled or floating-point outputs may diverge; contexts are capped at 1,024 tokens; the architecture is not hidden and the system assumes MoE; side channels are out of scope; and there is no formal guarantee that the fast executable code matches the Lean protocol.