Businesses can pay for access to a premium AI model, receive a convincing answer, and still have no proof from the response itself that the advertised model handled their request. A name displayed on a screen tells them only what the provider claims.
Anthropic unveiled FlashProve on Nov. 2, an open-source research prototype that could allow customers to independently verify the model behind an AI response without revealing the model’s proprietary weights.
As more organizations access AI through outside platforms, the ability to confirm which models power those services could give customers a new way to evaluate providers and the technology they're paying for.
A 400-billion-parameter model puts FlashProve to the test
FlashProve starts with a mathematical fingerprint representing a specific AI model. After answering a request, the provider can generate a separate proof that customers check against that fingerprint without accessing the model’s weights, according to Anthropic's research.
Researchers tested the system on an integer-arithmetic version of Meta’s Llama 4 Maverick, using 4-bit weights. The mixture-of-experts model contains approximately 400 billion total parameters. Claude helped optimize the verification software, with researchers using Lean, a formal verification system, to prove specified mathematical security properties. Those guarantees do not automatically extend to the fast C++ verification software.
Each proof was approximately 19 MB and took 17 seconds to verify on a computer processor using 24 threads. Customers performing the check would not need the expensive graphics processors used to run the AI model. The main benchmark checked one generated token per request within a 1,024-token window. Researchers also tested checking every reply token and reported no measurable speed difference.
Generating those proofs was considerably more demanding. Anthropic measured roughly 70 to 95 times the GPU time of prompt processing with vLLM, a widely used inference engine, in its benchmark comparisons.
Businesses could use FlashProve to check AI services
Companies may evaluate an AI model for specific tasks before approving it for use. If a service later delivers a different version, its responses could differ from those of the system employees were cleared to use. Platforms such as Microsoft Foundry offer access to models from multiple developers, providing a potential application for independent verification.
Unauthorized changes present another scenario. Organizations could use the technology to investigate whether an approved model had been altered or replaced. External evaluators could also check whether a system reviewed during AI safety assessments matched the version later made available to customers.
Anthropic previously explored a different verification method in its Frontier Safety Roadmap. FlashProve's GitHub repository makes the newer prototype available for research but identifies it as an unmaintained research artifact. No integration into commercial Claude services has been announced.
What eWeek found: AI identity checks need more than a passing proof
FlashProve’s technical documentation shows why businesses would need to examine what a proof actually covers before treating it as evidence that a provider delivered the model purchased.
- Exact matching differs from ordinary AI generation. None of the 192 stored replies generated by conventional floating-point inference matched the integer model’s highest-scoring token at every position. Researchers scored these replies without conducting full cryptographic verification. The result shows a compatibility problem between inference methods; it does not establish that a provider substituted another model.
- More flexible checks offer narrower assurances. An alternative scoring policy allows deviations from the model’s highest-scoring tokens. Acceptance depends on the configured limits and does not establish how the text was generated. Businesses would need to inspect those limits before deciding whether the result provides meaningful assurance.
- Model identity needs an independent reference. Customers must obtain the model’s commitment ID from a trusted party that vouches for its identity. Accepting that identifier only from the service being checked would leave the customer without an independent basis for identifying the model.
- Coverage matters. A proof applies to the token positions selected for checking. Verifying part of a reply does not authenticate the remaining text, so customers would need to ensure the check covers the output they intend to trust.
For businesses managing enterprise AI governance, these distinctions could shape future provider evaluations. Buyers would need to ask which model commitment is authenticated, which outputs are covered, and what deviations the verification policy permits.
Even a successful identity check would leave accuracy, safety, and suitability for a business task to separate evaluations. FlashProve’s potential value is giving customers stronger evidence about model delivery, provided the verification conditions are clear.
More AI news: AI agents may soon spend more time practicing in simulated offices, as Plato develops training environments based on actual business workflows.