FlashProve aims to sign AI model responses

Anthropic presents FlashProve, an open-source prototype that seeks to prove which model produced a response. The tool relies on cryptography but remains limited to research uses.

Anthropic is working on FlashProve, an open-source system designed to verify which AI model produced a given response, according to a research publication dated November 3, 2026. The idea matters for companies as well as regulators: to prevent a provider from quietly replacing one model with another, and to be able to prove that an output indeed comes from the correct machine. All this without exposing the model’s weights, those internal parameters that run the engine.

A cryptographic proof to identify the correct model

FlashProve is based on what Anthropic calls a “provable inference”: a mathematical proof that a given result indeed comes from a specific model. The principle is reminiscent of an email signature, except here the signature accompanies an AI response. According to the researchers, the system must resist impersonation attempts, including if the model provider tries to disguise its outputs.

In its announcement, Anthropic cites several possible uses. A company could, for example, demonstrate that an advanced cybersecurity model scanned software to detect vulnerabilities. Another could prove that a business plan was validated by a specific model. The mechanism also addresses a more down-to-earth problem: preventing an actor from passing off a cheaper model as a more powerful one, to reduce costs or inflate performance. The kind of little arrangement audits don’t like much.

What the prototype can and cannot do yet

To operate, FlashProve requires labs to publish a “fingerprint” of each model, then match it to generated outputs. These fingerprints do not reveal the model’s weights, so not its proprietary core, which theoretically allows application to commercial models as well. Anthropic also says it designed the system to remain lightweight, with verification possible on a local device.

The numbers give an idea of the gain compared to early attempts of this kind. According to the technical report, FlashProve can generate proofs at about 110 tokens per second on a single H100 GPU for Meta’s Llama 4 Maverick, a 400-billion-parameter model. Each proof weighs about 19 megabytes and verifies in 17 seconds on a standard processor. The overhead remains high, around 70 to 95 times normal inference, but Anthropic considers the gap significantly lower than comparable open-source systems already published, which took more than 14 hours for 1,000 tokens and produced a 93.6-gigabyte proof.

Why Anthropic is pushing this project now

The company links FlashProve to its Frontier Safety Roadmap, which targeted a prototype inference verification by September 2026. Anthropic says it reached this goal with another project based on inference traffic signing and tracking computation on a trusted server. FlashProve takes a different path: purely cryptographic, without trusted hardware. It’s more elegant on paper but also more demanding in practice.

The prototype remains very incomplete, however. Anthropic specifies it only covers contexts of 1,024 tokens, well below common context windows. It hides weights but not the model architecture. Most importantly, the proof works when the model chooses the next most probable token, not when it generates more random responses, as in most real conversations. In other words, the tool mainly knows how to show its strength demonstration, not yet keep up in production.

An as much governance tool as a technical tool

The intended uses go beyond mere scientific curiosity. Anthropic mentions regulators able to verify that a provider runs only authorized models, or companies wanting to trace more precisely who actually reviewed code or a document. The system could also help detect tampering on servers, when an attacker or a more advanced model slips into the chain unnoticed.

For now, FlashProve remains a research prototype. Anthropic publishes the technical specification, code, and documentation under an open-source license on GitHub but announces no internal or external deployment. The company also specifies that this publication does not close a finalized program: “This research post is meant to share intermediate progress, and not a completed research agenda,” write the researchers. The future will depend mainly on one simple yet tough point: making sure that verifying an AI response does not cost more than the response itself.

Key points

  • Anthropic published FlashProve on November 3, 2026.
  • The system verifies which model produced a response.
  • The prototype remains limited to 1,024 tokens.
  • Llama 4 Maverick is used for testing, with 400 billion parameters.
  • A proof weighs about 19 megabytes.
  • Verification takes 17 seconds on a standard CPU.

By the numbers

  • 110 tokens/second — proof generation on an H100 GPU, according to Anthropic, 2026.
  • 19 megabytes — size of a proof produced by FlashProve, according to the technical report.
  • 17 seconds — verification time on a standard processor.
  • 70 to 95 times — estimated overhead compared to normal inference.
  • 93.6 gigabytes — size of the proof from the previous open-source system cited by Anthropic.

To read