Anthropic releases FlashProve for provable LLM inference
Anthropic and academic collaborators released FlashProve, an open-source zkSNARK prototype that proves a 400B-parameter model produced a given response, giving users a cryptographic way to detect swapped models or tampered outputs.
Anthropic, working with researchers from EPFL and MIT, published FlashProve, a prototype system that generates cryptographic proofs that a response came from a specific committed model. The proofs do not reveal the model's weights.
The problem it targets is simple to state: today, users can't check that a response came from the model a provider says it did. A routing service could quietly swap in a cheaper model, or an attacker could tamper with the token stream. The authors say provable inference could help users of third-party routers, independent evaluators who need to confirm they are testing the model actually being served, and providers who want to show they are meeting commitments or legal obligations.
On Llama 4 Maverick, a 400B-parameter mixture-of-experts model with 4-bit quantized weights, a single H100 generates proofs at about 110 tokens per second. Each 1,024-token request produces a proof of about 19 MB. That costs roughly 70–95 times the GPU time of serving the same model on vLLM. The team says prior open-source systems are orders of magnitude slower. The proofs hide the weights but not the model architecture.
The post says Claude produced the system's improvements. Because Claude's optimization choices can be hard to evaluate, the team added top-level security statements written in Lean, a formal verification language.
The release includes five artifacts under Apache 2.0:
- a blog post
- a technical report
- a Claude-written protocol specification
- a prover with a fast C++ verifier and a provably correct Lean verifier
- Lean proofs, with an interactive visualizer for auditing them
The authors call this intermediate progress, not a finished research agenda. They are inviting others to examine the work and improve it.