San FranciscoAnthropic’s FlashProve verifies an AI answer in 17 seconds.
Developers renting a model cannot otherwise tell whether a provider swapped in a cheaper one.
On Llama 4 Maverick, a cryptographic proof lets a 24-thread CPU check the output against committed weights it never sees.
Proving still costs 70 to 95 times the graphics-chip time of ordinary inference, and it needs the provider’s cooperation.
Anthropic released the code under Apache 2.0 so other researchers can reproduce it and cut the cost.
How each outlet framed it
- Zhidx
- reports Anthropic's FlashProve as a cryptographic check against model swapping, with the 70–95x GPU overhead as the cost caveat
Sources: Zhidx