Anthropic Releases FlashProve for Verifying AI Model Outputs

AI · First reported 3 Nov, 19:54 · Updated 20:34

Image: Wikimedia Commons · Ysherazi · CC BY-SA 4.0

Anthropic’s open-source FlashProve research prototype attaches cryptographic proofs to AI responses so an external verifier can check which committed model computation produced them. The report describes its performance and limits, while a brief NewsBytes item notes its purpose but provides few technical details.

FlashProve is an open-source research prototype designed to let an external verifier check that a response came from a particular committed model computation.
Anthropic’s report says one 80 GB H100 generates proofs for Llama 4 Maverick at about 110 tokens per second; a 1,024-token request produces a proof of roughly 19 MB.
The prototype has constraints including support for up to 1,024 tokens, greedy decoding, and proof generation after responses; Anthropic describes the repository as not maintained.

Summary written by Lux AI from 2 articles

Key facts

Project: FlashProve, an open-source research prototype from Anthropic.
Reported model and hardware: Llama 4 Maverick on one 80 GB H100.
Reported proving throughput: About 110 tokens per second.
Reported proof size: Roughly 19 MB for a 1,024-token request.
Reported proving overhead: About 70 to 95 times the GPU time of native generation with vLLM, excluding the original inference work.
Prototype constraints: Supports up to 1,024 tokens and greedy decoding; proofs are generated after inference.