FlashProve

FlashProve is Anthropic's Frontier Red Team provable-inference research prototype (with EPFL/MIT collaborators), open-sourced on November 2, 2026. A provider commits to weights, then attaches a zkSNARK so clients can verify on a normal CPU that a reply came from those committed weights—without learning the weights.

On Llama 4 Maverick (~400B MoE, 4-bit) Anthropic reports ~110 tokens/s proof generation on one H100 and ~19 MB proofs per 1,024-token request, shipping a prover, fast C++ verifier, Lean verifier and technical report. The GitHub repo is explicitly an unmaintained research artifact under Apache-2.0.

Problem

API clients rarely can tell whether a provider silently swapped in a cheaper model. FlashProve turns “which weights answered?” into a verifiable claim via commitments + inference proofs.

Stated limits

• Context ~1,024 tokens; mainly integer/greedy paths.
• Proof cost ~70–95× GPU time vs ordinary vLLM inference.
• Architecture is not hidden; side channels out of scope; unmaintained research prototype.

Best for security/crypto and model-governance researchers—not a drop-in production billing verifier yet.

Who it is for

• Security teams researching verifiable AI, zkSNARKs and hosting trust.
• Platforms experimenting with audits of “did we run the committed model?”.

Getting started

1. Read anthropic.com/research/provable-inference and the technical report.
2. Clone github.com/anthropics/FlashProve and run the prover / cp_verify flow.
3. Reproduce latency and proof size on the documented MoE setup before deciding whether to integrate into your own hosting stack.