Anthropic Develops FlashProve Tool to Verify AI Model Outputs

Anthropic is developing an open-source verification system designed to let users confirm which artificial intelligence model actually generated a given response, a capability that could reshape how transparency is enforced across the rapidly growing AI industry. The tool, named FlashProve, was unveiled on November 2, 2026, and uses cryptographic techniques to provide what the company calls provable inference, allowing mathematical verification rather than relying on a provider's assurances.

At present, users largely have to take AI companies at their word when those firms claim a particular model produced an answer. FlashProve aims to close that trust gap by cryptographically attesting to the identity of the model behind any given output, giving developers, businesses and regulators a way to independently check claims about model provenance.

The announcement follows a series of allegations by Anthropic regarding competitors. The company has said it reported cases in which Chinese AI laboratories allegedly extracted Claude's capabilities to train their own systems, and that some firms routed customer requests through Claude while presenting the answers as their own. Anthropic has not explicitly tied FlashProve to any specific company, and the allegations against competitors have not been independently established.

According to Anthropic's September 2026 report, Moonshot AI allegedly forwarded nearly 300,000 Kimi customer requests to Claude over a ten-day period, primarily using its Opus models, through what the company described as thousands of fraudulent accounts appearing to operate from Singapore and Japan. Anthropic also alleged that conversations were saved and Claude's reasoning data extracted. Earlier, in February, the company accused DeepSeek, Moonshot and MiniMax of conducting large-scale model-distillation campaigns involving more than 16 million exchanges through roughly 24,000 fraudulent accounts, and in September it made similar allegations against Alibaba, Zhipu and Xiaomi.

Anthropic has drawn an important distinction about what the technology can and cannot do. FlashProve is built to verify which model generated an output, not to determine whether a competing model was trained using Claude's responses. Model distillation, the practice of using a stronger system's outputs to train another model, has legitimate applications, but Anthropic maintains that unauthorized extraction of Claude's capabilities violates its terms of service.

The open-source nature of the project could matter for the wider industry, since independent verification tools may become a baseline expectation for AI providers competing on trust. For Pakistani developers and businesses building on third-party AI models, such verification could eventually offer a clearer way to confirm which system is actually powering the services they pay for, though the tool's adoption and practical rollout remain to be seen.