Anthropic releases FlashProve, open-source cryptographic proofs for AI inference, at 110 tokens per second

Anthropic released FlashProve on November 2, 2026: an open-source research prototype that uses cryptographic proofs to show a language-model response came from committed weights without revealing them. The company reported proof generation at 110 tokens per second for Meta's Llama 4 Maverick.

What FlashProve does

FlashProve produces cryptographic proofs tied to a language model's inference. A proof shows that a given response was generated by a model whose weights were committed in advance. The weights themselves stay hidden.

Anthropic describes the release as a research prototype and has made it open source. The 110 tokens per second figure is Anthropic's own report for proof generation on Llama 4 Maverick, Meta's model. Pricing, context window and other benchmark figures are not applicable to this release.

The verification gap

The work targets a gap in hosted AI. A high-quality answer does not establish which model produced it. A routing service could substitute another system for the one a customer requested. The same ambiguity applies to an evaluator receiving outputs from a hosted model.

What this means

FlashProve is a research prototype, not a production service, and the 110 tokens per second figure comes from Anthropic. Its core claim is that a user of a hosted model can check that a response came from a specific committed set of weights, rather than trusting the provider's label. Because the proofs do not reveal the weights, a provider could in principle offer that check without exposing the model. Open-sourcing the prototype lets outside researchers examine and test the approach.