Anthropic has published an open-source research system called FlashProve that aims to answer a question users of AI services currently cannot check: did the response they received really come from the model they were promised? The system, described by the company on November 2, offers what researchers call provable inference, a cryptographic proof that a given output was produced by a specific model.

Today, when someone queries a hosted AI model, the answer itself carries no reliable evidence of its origin. A provider could quietly substitute a cheaper, less capable model to cut costs, or present the work of a stronger model than the one actually used, and the user would have no practical way to tell. That gap grows as businesses route work through third-party providers.

How the proof works

FlashProve works in two stages. First, a provider publishes a cryptographic commitment to a model's weights, effectively a fingerprint of the exact model it intends to serve. The fingerprint does not reveal the weights themselves, the internal numerical parameters that define how a model behaves and that commercial labs keep private. That could let it work with closed commercial models too.

When a user submits a request, the provider can then generate a proof, using a technique called a zero-knowledge succinct non-interactive argument of knowledge, that the committed model produced those output tokens for that prompt. Anyone holding the commitment, prompt and response can verify the proof without access to the weights, and Anthropic says verification is lightweight enough to run on a local device. The researchers compare the idea to a digital signature on an email, which confirms a message really came from its claimed sender.

Anthropic suggests the same mechanism could verify claims made by third parties. A developer might claim a software library was scanned for vulnerabilities by a frontier cybersecurity model; a proof attached to the work would let others check the claim instead of taking it on trust.

What the demonstration showed

To demonstrate the approach, Anthropic tested FlashProve on Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model. Running on a single Nvidia H100 graphics processor, the system generated proofs at approximately 110 tokens per second, with each proof measuring around 19 megabytes for a 1,024-token request. Verification took approximately 17 seconds using a 24-core processor, without requiring a graphics processor.

Anthropic has released the executable code, technical report, protocol specification and formal verification proofs under the Apache 2.0 open-source licence, for other researchers to build on.

Why it is not ready yet

The limitations are substantial. Generating proofs currently takes approximately 70 to 95 times the graphics-processor time of ordinary inference using vLLM, an open-source inference engine. The current system also supports contexts of up to 1,024 tokens and proves an integer version of the model under greedy decoding, so it does not yet cover every floating-point or sampled result. It hides model weights but not a model's architecture.

Anthropic is also clear that FlashProve is intermediate research, not a product, and the company has no plans to deploy it at Anthropic or any other lab for now. A proof is only as trustworthy as the published commitment it points to: it can show a response matches a committed model, but users still need reason to trust that the commitment was labelled honestly.

Even so, FlashProve points at a shift in AI trust. If model identity becomes something that can be checked rather than merely claimed, the question users ask of a provider could change from which model do you say you run, to can you prove it.