Anthropic releases FlashProve, open code that proves which model wrote a reply

TL;DR

• Anthropic released FlashProve, an Apache 2.0 research prototype that proves a reply came from a committed model without revealing its weights.
• On Llama 4 Maverick, one H100 proves about 110 tokens a second, and a CPU checks each 18.78 MB proof in 17 seconds.
• It handles contexts of up to 1,024 tokens, costs about 72 times vLLM's GPU time, and TechCrunch reports no lab plans to deploy it.