FlashProve is an open-source research prototype for proof-carrying inference, released November 2.
The design aims to verify a response against committed weights while hiding those weights.
What happened?
Anthropic published FlashProve on November 2 as an open-source research prototype for “provable inference.” When a user calls a hosted model, the output itself does not prove that the provider served the model selected. A router could substitute a cheaper model, an evaluator could receive a different checkpoint, or a compromised service could alter the token stream. FlashProve’s proposed answer is a proof attached to the response, tied to a provider’s commitment to specific model weights.
The system uses zero-knowledge succinct arguments, or zkSNARKs. Anthropic says the commitment and proof hide the weights while allowing a verifier to check the claimed computation without running the full model. The release includes a research post, technical material, a protocol specification, executable code, C++ and Lean verifiers, and Lean proofs for top-level mathematical statements. The work is presented as intermediate progress, released so others can audit and improve it.
Anthropic reports a demonstration on a 400-billion-parameter Llama 4 Maverick mixture-of-experts model using four-bit quantized weights. On one H100, the system generated proofs at roughly 110 tokens per second; a proof for a 1,024-token request was about 19 MB, and the company says proving required roughly 70–95 times the GPU time of ordinary vLLM inference in the comparison. Those figures describe the prototype’s measured setup, not the cost or latency of a production service using every model and request shape.
Source: Anthropic ↗ · Source: Anthropic ↗ · Source: Unite.AI ↗
A proof can bind output to weights, not to a vendor’s marketing name
The most important limitation is upstream of the proof. FlashProve can show that a response matches a committed model under a declared acceptance policy; it cannot, by itself, prove that the provider committed to the model name advertised. A verifier needs an authenticated mapping between the commitment and the public model identity, plus confidence that the implementation and deployment operate as claimed.
The prototype also has engineering limits. Its documentation describes integer, quantized computation, bounded request windows and acceptance policies that can be exact or tolerance-based. The repository warns that the programs themselves are not automatically covered by the high-level soundness statements and that a proof about an integer model does not show how close it is to a floating-point model. These are conditions that determine what a successful verification actually means.
TechCrunch reports that Anthropic has no current plan to deploy the system in a product. That matches the research post’s description of intermediate progress. A reader cannot turn the release into a claim that Claude, OpenAI, a third-party router or an API endpoint now supplies a user-visible verified-model badge. The evidence supports a public protocol and implementation, not live commercial enforcement.
Source: Anthropic ↗ · Source: Anthropic ↗ · Source: TechCrunch ↗
Why does it matter?
Model identity is becoming a practical trust problem as routing services, enterprise gateways and evaluation platforms sit between users and weights. A benchmark result is only as meaningful as the system that produced it, and a compliance promise is hard to audit if the client sees only text. Proof-carrying responses could make model substitution detectable without forcing a provider to publish proprietary weights. That would be useful for independent evaluators, regulated deployments and teams paying for a particular capability rather than merely a fluent answer.
The cost profile explains why this is research rather than a finished product. A roughly 19 MB proof per 1,024-token request is substantial for high-volume inference, and a 70–95x GPU-time comparison would change provider economics if applied directly. Verification may be cheaper than inference, but the provider still has to generate evidence and manage commitments, keys, policies and failures. For interactive systems, the result must also be explainable when a proof is missing or invalid without silently falling back to an unverified answer.
The near-term question is not whether FlashProve has solved AI provenance. It is whether the protocol, implementation and operational duties can mature into a scheme where users know what was committed, what was proved, what assumptions were made and what remains outside the proof. Anthropic has made that discussion more concrete by releasing code and limits. Until an endpoint participates and a verifier checks the exact record, a model label remains a claim, not cryptographic identity.
Source: Anthropic ↗ · Source: Anthropic ↗ · Source: Anthropic ↗ · Source: TechCrunch ↗ · Source: Unite.AI ↗
Read the sources
- Provable inference ↗
Anthropic · November 2, 2026 - FlashProve repository ↗
Anthropic - FlashProve architecture and limits ↗
Anthropic - Anthropic system to prove AI model identity ↗
TechCrunch · November 3, 2026 - Anthropic releases FlashProve prototype ↗
Unite.AI · November 3, 2026
