Skip to content
TR

FlashProve: Anthropic's open system to prove which AI model answered

Anthropic has published FlashProve, an open-source research system that lets users cryptographically check which AI model produced an answer.

By Cultvia Editorial2 min read
Technology image
Illustrative image · Photo: Albert Stoynov / Unsplash

When you ask an AI service a question, you cannot easily check which model actually wrote the reply. On 2 November 2026 Anthropic published research on provable inference and a system called FlashProve that tries to close that gap. TechCrunch reported on the work on 3 November.

What is FlashProve?

FlashProve is an open-source system for what researchers call provable inference. A model provider commits to a specific model and then attaches a cryptographic proof to each response, showing the answer was produced by that model. Anthropic describes the proof type as a zkSNARK, chosen because it hides the model's weights while staying efficient to check.

The post gives one measured example. For a large open model, Llama 4 Maverick with 400 billion parameters, a single H100 GPU generates proofs at around 110 tokens per second, with a proof of roughly 19 MB per 1,024-token request. Anthropic says that costs about 70 to 95 times the GPU time of running the model itself, and that earlier open-source systems are orders of magnitude slower.

Why would anyone need to prove a model's identity?

Anthropic lists two risks. A routing service could secretly substitute a cheaper model to cut costs, or an attacker could tamper with the token stream. TechCrunch adds a third use: checking third-party claims, such as a developer saying a code library was scanned by a frontier security model.

To work, labs would publish a "fingerprint" of each model they release. The fingerprints do not reveal weights, so closed commercial models could take part. Anthropic also says verification is designed to be light enough to run on a local device.

What has actually been released?

The company says it is releasing five artifacts, including the blog post, a technical report, a protocol specification and executable code with a prover and two verifiers. It also provides auditable security statements written in Lean, a formal verification language. The code is available under an open-source licence.

Okur için: what does this mean for you?

Nothing changes in the apps you use today. Anthropic describes the post as intermediate progress rather than a finished research agenda, and says there are no plans to deploy the system at Anthropic or any other lab. Treat FlashProve as a technical building block that other researchers are invited to improve. For more AI coverage, see the Pulse hub.

FAQ

Does FlashProve make AI answers more accurate? No. It only helps show which model produced an answer, not whether the answer is right.

Can I use it now? The code and specification are public, but the system is a research release, not a consumer feature.

Frequently asked questions

Does FlashProve make AI answers more accurate?

No. It is designed to show which model produced an answer, not whether that answer is correct.

Is FlashProve available in consumer apps?

No. Anthropic published code and a specification as research and says it has no plans to deploy it at any lab yet.

Cultvia Editorial

Editorial desk

Stories signed by the Cultvia editorial desk follow our editorial policy; every story lists its sources at the end and errors are fixed under our corrections policy.

#ai#Anthropic#cryptography

Last updated:

What changed?
  1. : First published

Sources

  1. Progress on provable inference for large language models — Anthropic, 2 November 2026
  2. Anthropic is working on a system to prove AI models' identity — TechCrunch, 4 November 2026
Was this helpful?
Report an error

Spotted wrong information, a calculation error or a technical problem? Tell us briefly and our editors will review and fix the page.

We store the page address and browser details with your report. Your email is only used to reply, is never shared and is deleted once the report is closed.

Up next

Close-up of a pair of headphones next to a laptopAIWhat is Claude Meetings? Claude can now join your video calls13 November 2026 · 2 min read
More from Pulse