Anthropic announced on November 2, 2026 FlashProve, an open-source prototype of cryptographic proofs for large language model inference. The system uses zkSNARKs so a model provider can commit to a specific set of weights and then prove that each response was generated by that model, without revealing the weights.
What FlashProve does
When a user queries a model via API or a routing service, there is no way to be certain the response came from the exact announced model. A provider could swap in a cheaper one, or an attacker could alter the token stream. FlashProve solves this with a zero-knowledge succinct argument: the provider publishes a short commitment to the weights and, for each request, generates a roughly 19 MB proof that anyone can verify in seconds on a regular CPU, without access to the weights.
In testing with Llama 4 Maverick (a 400-billion-parameter MoE model, 4-bit quantized), a single NVIDIA H100 GPU generated proofs at about 110 tokens per second. That is roughly 70 to 95 times the GPU time needed to run the same model on vLLM. The proofs reveal neither the weights nor which experts were activated.
Why it matters
Anthropic highlights applications in customer trust, tamper detection, regulator audits, and verifying that hardware is being used for inference rather than training. The system includes executable code (prover and two verifiers), Lean formalizations, and a technical report, all under the Apache 2.0 license in the anthropics/FlashProve repository.
The company stresses that this is intermediate progress, not a finished product, and that the goal is to accelerate the development of provable inference by the community. The full post is at anthropic.com/research/provable-inference.
Sources
Transparency: This content was created, edited, or reviewed with the aid of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Consult the original sources to verify the full context.
By GeekikiBot
