MADRID, November 4 — Anthropic released FlashProve on November 2, an open-source research prototype designed to prove which model generated an AI response without revealing its weights. For small-business owners choosing AI services, the distinction is between a model name supplied by a vendor and evidence a customer can independently check.

According to Anthropic’s research announcement, the provider first publishes a cryptographic commitment, effectively a fingerprint of its model weights. It then generates a proof linking a prompt and response to that commitment. Users can verify the proof without possessing the weights or rerunning the model. Released under the Apache 2.0 license, FlashProve makes that approach available for research. Anthropic identifies customers of third-party routing services as potential beneficiaries, since output quality alone can’t reliably establish whether a provider substituted another model.

The computational bill remains substantial.

In Anthropic’s tests using quantized Llama 4 Maverick weights, proofs were about 19 MB per 1,024-token request. Verification took 17 seconds on 24 CPU threads, while generating proofs required roughly 70–95 times the GPU time of ordinary inference in its comparison. The business implication is a trade-off: independently checkable model identity comes with additional computation, not just a different label on a reply.

That isn’t a commercial feature announcement. TechCrunch reported there were no plans to deploy FlashProve at Anthropic or another laboratory; Anthropic itself described the release as intermediate progress rather than a completed research agenda.

For owners evaluating AI used in customer communications or business-critical work, the practical question is whether a provider supplies model commitments and response-level proofs customers can verify. The research doesn’t establish that existing paid services provide either. Editorially, its significance is the distinction it makes possible: a model-name claim and verifiable evidence of model identity are different things. Neither, by itself, proves that an answer is correct.

Sources