FlashProve verifies which model answered
Anthropic released an open-source prototype that uses cryptographic proofs to verify outputs against a specified model while keeping its weights private.
An AI answer does not reveal which model actually produced it. FlashProve explores a way to verify that connection without requiring the customer to hold the model’s weights.
What happened
Anthropic released FlashProve, an open-source research prototype for provable inference, on 2 November. It uses cryptographic proofs tied to a commitment to particular model weights.
The release includes executable code, protocol documentation and formal verification work. In its demonstration using the 400-billion-parameter Llama 4 Maverick model with four-bit weights, proof generation required roughly 70–95 times the GPU time of ordinary inference with vLLM.
Why it matters
Verification could help customers check that a service used the model they expected, or help evaluators assess whether the model being tested is the one being served.
The guarantee concerns how an output was generated. It does not establish that the answer is factually correct, useful or appropriate for a particular decision.
The bigger picture
AI services increasingly involve hosting providers, routing systems and other intermediaries. That creates a separate trust problem around model identity and the integrity of the response.
FlashProve provides research progress on that problem, while the computing overhead remains a substantial practical constraint. The appropriate next questions concern which use cases can justify the extra cost and how the method performs beyond the demonstrated configuration.