Technology & AI

Anthropic publishes open-source prototype to check which AI model replied

FlashProve is a research prototype from AI firm Anthropic, aimed at developers who cannot tell if their paid AI model is the one answering.

By Standard Topic. AI-generated from the reporting listed under Coverage. Published . Updated .

This story is no longer being updated. It reflects reporting up to .

What happened

Anthropic is developing FlashProve, a system that checks which AI model produced a response, so customers of AI APIs can detect if a provider quietly routes them to a cheaper model. On Nov 4 the company released it as an open-source prototype, which works only with integer models and greedy decoding and adds about 70-95x overhead, with no deployment planned.

Why it matters

Developers using AI APIs have little way to know whether the model answering their requests is the one they are paying for. The reports describe the approach as having a costly catch: heavy computational overhead.

Timeline

  1. Nov 3, 7:46 PM ET: Anthropic was reported to be developing FlashProve to verify which AI model generated an AI response.
  2. Nov 3, 11:34 PM ET: Anthropic researchers unveiled FlashProve as a prototype, which processes 110 tokens per second on a single H100 GPU but is limited to integer models and greedy decoding.
  3. Nov 4, 2:31 PM ET: Anthropic released FlashProve as an open-source prototype at about 72x GPU overhead, with no deployment planned.

Coverage

More on this story

More from Technology & AI

All Technology & AI news · Technology & AI story archive