Anthropic released FlashProve on November 3, 2026, an open-source system that cryptographically proves a given AI output came from a specific model1,2. The tool targets a concrete trust gap in the AI inference chain: providers or routers secretly swapping a cheaper model behind the API endpoint a customer paid for, or substituting a more expensive model to inflate benchmark results.
FlashProve delivers what the researchers call provable inference — a cryptographic proof, described as a mathematical guarantee, that a particular output was generated by a particular model. The mechanism is analogous to cryptographic email signatures: it lets recipients confirm a message really came from its claimed sender, and is designed to be robust even when the impostor is the lab that built and runs the model.
For the system to function, AI labs must publish a fingerprint of each model they release; that fingerprint can later be matched to specific outputs and verified cryptographically.
Beyond policing cost-cutting swaps, the researchers outlined third-party verification scenarios. A developer might claim a software library was scanned for vulnerabilities by a frontier cybersecurity model, or a company might assert its business plan was reviewed by a particular model — FlashProve would let anyone check those claims.
The project lists contributors from Anthropic — Roy Rinberg and Alex Allain — alongside external collaborators Alessandro Chiesa of EPFL and StarkWare, Miranda Christ of MIT and RESI, and Rahul Dalal.
ANALYSIS FlashProve addresses a problem that grows with the proliferation of model routers and multi-model orchestration layers, where end users have no independent way to confirm which model actually served their request. The requirement that labs publish per-model fingerprints, however, means adoption depends on voluntary cooperation across competing providers — a coordination challenge the open-source release alone does not resolve.