# Anthropic open-sources FlashProve, a zero-knowledge prototype that proves which model wrote a response

Canonical: https://theagenticweekly.com/news/anthropic-provable-inference · Evidence: Reported · First seen: 2026-11-03 06:40 UTC · Source published: not stated by the source · Brief published: 2026-11-03 07:00 UTC · Updated: 2026-11-03 07:00 UTC · Author: Agentic Weekly

> On November 2, 2026 Anthropic published FlashProve, which it calls "an open-source performant prototype for provable inference". On Llama 4 Maverick (400B parameters, 4-bit) it reports about 110 tokens per second of proving on a single H100, with proofs of about 19 MB per 1,024-token request.

## What changed

- Anthropic released FlashProve on November 2, 2026: a prover, a fast C++ verifier, an executable Lean verifier, a technical report and Lean proofs, all under the Apache 2.0 license.
- A provider publishes a commitment to its weights once (about 39 minutes per model), then attaches proofs that the committed model produced the exact output tokens, without revealing the weights.
- On Llama 4 Maverick in 4-bit, the post reports 223 seconds to prove a batch of 24 requests on one H100 (110 tokens per second) and 18.78 MB per proof.
- Checking a proof needs no GPU: the post reports 17 seconds on 24 CPU threads, after downloading a commitment of about 6.5 MB.
- Proving costs about 72 times the per-GPU time vLLM needs for the same requests (66 to 77 across runs), and about 94 times at vLLM's highest tested load.
- It is a research prototype: it covers mixture-of-experts models and integer greedy decoding only, and the fast C++ code carries no formal guarantee.

## In the source's words

> "Today, we release FlashProve: an open-source performant prototype for provable inference"
> (Anthropic research post, https://www.anthropic.com/research/provable-inference)

> "We release our code, Lean proofs, and documentation under the Apache 2.0 license."
> (Anthropic research post, https://www.anthropic.com/research/provable-inference)

> "This release is a research prototype"
> (Anthropic research post, Limitations, https://www.anthropic.com/research/provable-inference)

## What we have not verified

- We did not run the prover or verifier; all speeds and proof sizes are Anthropic's reported measurements.
- The post says the Lean protocol, the paper and the fast code are not guaranteed to match each other; we did not audit them.
- The GitHub repository describes itself as not maintained; we did not check whether any provider plans to ship proofs in production.

## Timeline

- Source published: not stated by the source
- First seen by our watcher: 2026-11-03 06:40 UTC (via Anthropic (news, research, engineering))
- Brief published: 2026-11-03 07:00 UTC

## Sources

- [Anthropic: Progress on provable inference for large language models](https://www.anthropic.com/research/provable-inference) (primary source) (fetched 2026-11-03)
- [anthropics/FlashProve on GitHub](https://github.com/anthropics/FlashProve) (primary source) (fetched 2026-11-03)

## Related

- [All news briefs](https://theagenticweekly.com/news)
- [This week's releases](https://theagenticweekly.com/this-week)
- [New AI models this week](https://theagenticweekly.com/models/this-week)

## How to cite

Agentic Weekly, "Anthropic open-sources FlashProve, a zero-knowledge prototype that proves which model wrote a response", published 2026-11-03 07:00 UTC, https://theagenticweekly.com/news/anthropic-provable-inference