AINov 4, 2026, 1:41 PM UTC·1 min read

Anthropic's FlashProve Proves Which AI Model Wrote a Response

Anthropic has released FlashProve, an open-source prototype that uses cryptographic proofs to show a response really came from the model a provider claims.

Illustration of a cryptographic proof seal attached to an AI model response on a screen. — Image: Webrazzi

Anthropic is working on an open-source system that lets users check whether an AI response was actually generated by the model a provider says it used. The prototype, called FlashProve, attaches cryptographic proofs to each answer.

The goal is to close two gaps. Users could detect whether a provider quietly swapped in a cheaper model to cut costs, and whether an attacker tampered with a response along the way.

How it works

FlashProve relies on publishing a fingerprint unique to a model without exposing its weights. The provider first creates a commitment for the model. Each response then comes with a proof that the given input and output were produced by that committed model.

Users can verify the proof without access to the weights or heavy computing power. Anthropic says verification is designed to be light enough to run on the user's own device.

The uses go beyond end users. A third-party service could prove that a software security scan was run with an advanced model, while independent evaluators and regulators could audit whether a deployed model matches the one declared.

The cost

Anthropic tested the system on Llama 4 Maverick, a model with 400 billion parameters. On a single H100 GPU, proofs can be generated at roughly 110 tokens per second. A 1,024-token request produces a proof file of about 19 MB. Compared with simply running the model, the process increases GPU time by 70 to 95 times.

That overhead shows how far the approach is from everyday deployment. The company has published the technical report, the protocol definition and working code under the Apache 2.0 license. FlashProve is not a finished product, and Anthropic currently has no plans to put the system into production.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Dialo: learn languages by speaking

Talk to an AI conversation partner; every sentence comes back corrected and explained in your language. 9 languages, A1 to C2.

ediviewer: read EDI files like documents

Open EDIFACT and X12 messages, see every segment explained, convert to JSON/XML, export PDF and Excel. Runs locally, no upload.

giftdetect: the right gift in minutes

Pick who it is for, the occasion and their personality. The search engine reads product descriptions from local shops and ranks what fits. Free, no sign-up.

Tags:anthropicflashproveopen sourcecryptographyai verificationllama 4