Anthropic launches FlashProve, a cryptographic system to verify the identity of AI models

On November 3, 2026, Anthropic announced the launch of FlashProve, an open-source system designed to provide "provable inference" through cryptographic proofs. Developed by the company’s Frontier Red Team in partnership with researchers from EPFL, StarkWare, MIT, RESI, and Succinct, the prototype allows any user to verify that a response generated by artificial intelligence actually came from the language model it claims to have produced. The technology addresses the "bait-and-switch" problem in the API market, where providers could charge for a top-tier model while delivering results from a cheaper one.

How validation works without exposing trade secrets

The technical core of FlashProve is based on zero-knowledge proofs, specifically zkSNARKs technology. Historically, proving that a text was generated by a specific neural network required direct access to the model's "weights"—the billions of mathematical parameters that dictate its behavior. Since these weights represent the most valuable intellectual property for labs like Anthropic and OpenAI, they are kept secret in closed-source models.

FlashProve circumvents this limitation by requiring AI labs to publish only a cryptographic "fingerprint" of each released model. When a user or system receives an output, they also receive a mathematical proof tied to this fingerprint. This proof confirms irrefutably that the text was generated by that specific set of weights, without ever revealing them to the public.

In practical tests conducted by Anthropic, the team used Meta's Llama 4 Maverick model. The generation of the cryptographic proof achieved a performance of approximately 110 tokens per second using a single Nvidia H100 GPU. This number refers to the throughput of proof generation on the already completed text, rather than the speed of generating new AI responses.

The generated verification file is approximately 19 MB. Each proof checks a single output token against a context window of up to 1,024 tokens. Once generated, the 19 MB file is light enough to be verified by a standard CPU in a matter of seconds, allowing for local validation on end-user devices.

Context and background of provable inference

The journey to FlashProve began with Anthropic’s own public safety commitments. In its "Frontier Safety Roadmap," the company had set September 30, 2026, as a self-imposed deadline for developing a provable inference prototype. The central goal was to mitigate threats to the AI supply chain.

According to the company's original safety document, there is a fear that sophisticated attackers could infiltrate systems and modify models after training, whether for corporate sabotage or to force models to serve malicious objectives. The ability to reliably and provably "sign" a model's outputs drastically reduces this threat.

Although the September deadline was slightly missed, Anthropic published the research post detailing the progress on November 2, 2026, followed by the official press release on November 3.

Concrete data on the FlashProve project

• Official announcement date: November 3, 2026.
• Proof generation performance: ~110 tokens per second.
• Test hardware: Nvidia H100 GPU.
• Verification file size: ~19 MB.
• Context window validated per proof: Up to 1,024 tokens.
• Model validated in tests: Llama 4 Maverick (Meta).
• Academic and corporate partners: EPFL, StarkWare, MIT, RESI, and Succinct.

Impact on the sector, users, and the Latin American market

The introduction of FlashProve affects different layers of the technology and artificial intelligence ecosystem in distinct, practical ways.

For AI companies and developers: The system creates a barrier against "model substitution" fraud. With the proliferation of AI routing services (APIs that promise to choose the best model for each prompt), developers run the risk of paying premium fees for frontier models (like Claude Opus or GPT-4) and receiving responses from smaller, cheaper open-source models, aiming to maximize the intermediary's profit. FlashProve eliminates this possibility by providing a mathematical guarantee of origin.

For regulated sectors and information security: Audits of critical processes gain a vital tool. If a developer claims that a sensitive software library was scanned for vulnerabilities by a top-tier cybersecurity model, or if a startup claims that its business plan was validated by a specific financial AI, these claims can now be cryptographically verified by third parties.

Impact on the Latin American market: In Latin America, where currency fluctuation makes the dollar cost of AI APIs a significant bottleneck for startups, the use of lower-cost routers and API aggregators is highly popular. Implementing standards like FlashProve would allow companies in the region to use these aggregation services securely, ensuring they are not victims of "silent degradation" in quality when outsourcing access to AI models.

What is still unknown

Despite the technical documentation and open-source code made available on GitHub, the editorial briefing points to significant gaps regarding the commercial viability of the system.

First, commercial scalability remains uncertain. FlashProve is a prototype, and it is unclear how the architecture would handle the massive scale traffic of global consumer products. Second, the operational costs associated with generating the proofs are unknown; dedicating expensive H100 GPU processing cycles just to generate cryptographic proofs at 110 tokens per second could inflate the final cost of processing, passing a financial burden on to the user.

Finally, industry adoption is an unknown. So far, there is no indication that giant competitors like OpenAI, Google, or Meta itself (whose model was used in the tests) intend to adopt this standard. Anthropic itself has categorically stated that it has no current plans to implement FlashProve in its commercial services.

What to expect next

Anthropic has made it clear that FlashProve represents an intermediate step, not a completed research agenda. With the technical specifications and working code available under an open-source license on GitHub, the next practical step is decentralized analysis.

The company expects independent developers, academic cryptography researchers, and other AI security teams to use the repository to optimize current performance bottlenecks. In the long term, the technology could serve as the foundation for a future industry-standard protocol for AI identity verification, required by government regulators for compliance in high-risk sectors.

Frequently asked questions

1. Does FlashProve expose the secret data (weights) of commercial AI models? No. The system uses zero-knowledge proofs (zkSNARKs) and only requires the publication of a cryptographic "fingerprint" of the model, allowing mathematical verification of origin without revealing internal weights or the company's intellectual property.

2. Can I use FlashProve today on Claude or ChatGPT? No. FlashProve is strictly an open-source research prototype. Anthropic has confirmed that, at the moment, there are no plans to deploy it commercially in its own services or those of other AI labs.

3. What is needed to verify if a response is true using this system? Final verification requires the approximately 19 MB proof file generated by the AI server. This file is light enough for the user to process and validate authenticity locally using the CPU of a standard computer or device in just a few seconds.