Anthropic open-sources FlashProve, a cryptographic system to verify which AI model answered

Anthropic said on 2 November 2026 that it has made progress on cryptographic proofs that let a user verify which specific AI model produced a given response, publishing an open-source system it calls FlashProve on its research site; TechCrunch reported the work the following day.

What FlashProve does

• The system generates a cryptographic proof — "a short proof... that the user can quickly verify (ideally, in seconds on a CPU)" — that a response came from a specific model, Anthropic said, guarding against a model "secretly substituted in by a routing service for cost reasons, or even by an attacker who has tampered with the token stream."
• For a 400-billion-parameter mixture-of-experts model, Llama 4 Maverick, a single Nvidia H100 GPU generates proofs at around 110 tokens per second with 4-bit quantized weights, with each proof about 19 megabytes per 1,024-token request under the same quantized-weights setup, according to Anthropic — roughly "70-95x the GPU time" needed to run the model itself on the open-source library vLLM.
• Anthropic said it is releasing five artifacts under an Apache 2.0 licence: the research post, a technical report, a protocol specification document, executable prover and verifier code, and Lean-language mathematical proofs of the system's security claims.
• TechCrunch reported that labs must publish a "fingerprint" of each model that can later be matched to its outputs, and that the fingerprints do not reveal a model's weights, so commercial labs could adopt the system without disclosing proprietary data.
• Anthropic said the research is "intermediate progress, and not a completed research agenda." TechCrunch reported that "research in the system is ongoing, and there are no plans to deploy it at Anthropic or any other lab."

Why it matters

Anthropic frames provable inference as a tool not just for consumer trust but for "verifying compliance with AI pacing requirements" and other regulatory commitments, which could matter as governments move from voluntary AI codes towards binding obligations on frontier labs. For businesses outside Silicon Valley — including Azerbaijani banks and telecoms operators now buying AI services through resellers and routing layers — a working version of this kind of proof would offer a way to confirm they are receiving the model tier they are paying for, rather than a cheaper substitute swapped in without disclosure.