Anthropic FlashProve Aims to Prove Which AI Made an Answer
Anthropic FlashProve is an open-source tool under development that could expose AI companies secretly using Claude to generate responses while claiming the answers came from their own models.
The system, called FlashProve, uses cryptographic technology to verify which AI model produced a response. Although Anthropic has not explicitly linked the project to Chinese AI companies, it could help address practices the company has previously accused Chinese competitors of using.
FlashProve verifies which model generated an output; it does not determine whether another model was trained on Claude’s responses.
The allegations behind the tool
In September 2026, Anthropic accused Moonshot AI of secretly routing Kimi users’ requests through Claude and presenting the responses as Kimi’s own. According to Anthropic’s September 2026 report, Moonshot secretly forwarded nearly 300,000 Kimi customer requests to Claude over ten days, primarily using its Opus models.
Anthropic alleged that Moonshot accessed Claude through 5,380 fraudulent accounts, mostly appearing to operate from Singapore and Japan. It also accused the company of saving conversations and extracting Claude’s reasoning data to improve its own models.
Earlier, in February, Anthropic accused DeepSeek, Moonshot and MiniMax of conducting large-scale model-distillation campaigns involving more than 16 million exchanges through approximately 24,000 fraudulent accounts. Model distillation involves using responses from a more capable AI system to train another model. Although the technique has legitimate applications, Anthropic says unauthorised extraction of Claude’s capabilities violates its terms.
In September, Anthropic also accused Alibaba, Zhipu and Xiaomi of similar activities, and alleged that DeepSeek had secretly forwarded some customer requests to Claude.
These remain Anthropic’s allegations rather than independently established findings. The companies named have not been shown, in the reporting of the tool, to have accepted the claims. Readers should treat the allegations as claims by one competitor about others, not as proven facts.
How Anthropic FlashProve works
Anthropic introduced FlashProve on November 2, 2026, as a system for provable inference, allowing users to mathematically verify which AI model generated a response.
At present, users generally have to trust a company’s claim about which model produced an answer, even though a provider could secretly substitute a cheaper or more powerful model.
FlashProve aims to prevent such misrepresentation through cryptographic verification. To use the system, AI companies would publish a unique cryptographic fingerprint associated with each model. When the model generates a response, the provider can produce mathematical proof linking that output to the published fingerprint. Users can then verify whether the response came from the advertised model. If a different model generated the response, the verification should fail.
Why companies would not need to reveal their models
One major advantage of FlashProve is that companies can verify their models without revealing their internal weights. Model weights are the numerical values that determine how an AI system processes information and generates responses. Companies such as Anthropic and OpenAI generally keep the weights of their leading models private.
FlashProve allows providers to publish cryptographic fingerprints without exposing that proprietary information. The system is also designed to support lightweight verification on local devices, although generating the proofs requires considerably more computing power. Participation would require AI providers to publish the necessary fingerprints and generate proofs.
What FlashProve could check beyond chatbots
Beyond identifying which model generated a response, FlashProve could help verify claims made by companies using AI for professional tasks. FlashProve could provide mathematical evidence confirming which model performed a claimed task, and help researchers verify the models used during benchmark testing.
Proving a model’s identity does not guarantee that its answers are accurate. Users will still need to judge the quality of the output itself.
The cost problem
Anthropic has demonstrated FlashProve using Meta’s Llama 4 Maverick, which contains 400 billion total parameters. Using four-bit quantised weights and a single Nvidia H100 GPU, the system generated proofs at approximately 110 tokens per second.
For a request containing 1,024 tokens, the proof was approximately 19MB and could be verified on a CPU within seconds. However, proof generation required approximately 70 to 95 times more GPU computation than comparable inference using vLLM.
That additional computing cost remains a major obstacle to widespread adoption. A system that costs 70 to 95 times more to run will not be used for every chat message. It is more likely to be used for high-value checks, such as benchmark results, regulated work or disputed claims, where the cost of proof is worth paying.
Related coverage: Tencent unveiled a low-cost Hy4 AI model; Anthropic warned AI may soon build itself; Apple AI smart glasses are tipped for late 2027; and the PTA registered BIGO LIVE and Likee in Pakistan.
Frequently asked questions
What is FlashProve?
FlashProve is Anthropic’s open-source system for provable inference. It uses cryptographic proofs to verify which AI model produced a response.
Does FlashProve prove a model was trained on Claude?
No. It verifies which model generated an output. It does not independently prove unauthorised model distillation or identify the training data used to build competing models.
Why would a company use FlashProve?
It allows a company to prove it used the advertised model without revealing its internal model weights, which are proprietary.
What is the main obstacle to adoption?
Cost. Proof generation required approximately 70 to 95 times more GPU computation than comparable inference in Anthropic’s demonstration.
Is FlashProve available in Claude now?
No. Anthropic has not announced plans to deploy it in Claude or any commercial service. It is released for researchers to study, test and improve.
