Anthropic open-sources FlashProve, a prototype that proves which AI model wrote a reply
Anthropic's Frontier Red Team released FlashProve, an open-source research prototype that lets a model provider prove a reply came from a specific committed set of weights, without revealing those weights, using zero-knowledge proofs that anyone can check.
What happened
- Anthropic published a research post on November 2, 2026 describing FlashProve, a system for provable inference built with help from Claude.
- A provider first publishes a commitment that pins down one exact set of model weights while keeping them hidden; for each request it then produces a proof that those committed weights produced the reply.
- Anthropic reports that for Llama 4 Maverick (a 400B-parameter mixture-of-experts model in 4-bit weights), one H100 GPU generates proofs at around 110 tokens per second, with about 19 MB of proof per 1,024-token request, roughly 70-95 times the GPU time of running the model on vLLM.
- The release includes Apache-2.0 code on GitHub, a technical report and protocol specification, and Lean formal statements of the claimed security guarantees; the repository describes it as a research artifact that is not maintained.
- The post frames the work as a defense against a compromised inference server, or a different model, quietly serving replies in place of the authorized one.
Why it matters: If AI providers adopt something like this, businesses could check that the model they pay for is the one actually answering, instead of taking it on trust. For now it is a research prototype with heavy compute overhead, not a product feature.
Sources
- Anthropic: Progress on provable inference for large language models anthropic.com
- FlashProve on GitHub github.com
Checked November 4, 2026 against the sources above. Claims about performance or benchmarks are the company's own unless stated otherwise.