Breaking

Anthropic Releases FlashProve, an Open-Source Library for Verifying AI Inference

November 4, 2026 at 10:45 EST

Open Source
Security

Anthropic has released FlashProve as open-source software, a library designed to let users check that an AI model actually ran the way it claims to have run. The company's announcement describes FlashProve as a tool for "provable inference," and says it has been benchmarked on Llama 4 Maverick, a mixture-of-experts model with 400 billion parameters.

Open-source release · FlashProve

A library that checks whether an AI model ran as claimed

FlashProve is built for "provable inference," producing evidence that a claimed model and computation match what was actually executed. It has been benchmarked on Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model.

400B

Parameters in Llama 4 Maverick, the model used for the benchmark

MoE

Mixture-of-experts architecture of the benchmarked model

Open

Released as open-source software, so its method can be examined directly

How the verification idea works

Verification closes the gap between what a provider says ran and what actually ran.

Step 1

Claimed model

Provider states which model and computation ran

→

Step 2

Execution

Inference runs on infrastructure the user may not control

→

Step 3

Evidence

FlashProve produces proof of correct execution

→

Step 4

Match check

Claim and executed computation are confirmed to match

Why it matters

Addresses the trust gap when models are served by third parties
Lets researchers, competitors and enterprises examine the method directly
Could make verifiable inference a shared standard, not a proprietary feature
Relevant to regulated industries needing auditable evidence of AI behavior

Still unverified

Proof method and the specific checks performed
Performance overhead of generating proofs
Independent evaluation of the Llama 4 Maverick benchmark results
Scope of what FlashProve can and cannot prove, and its licensing terms

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.