Illustration, AI-generated from the story text. Tap a panel to read the story through it.
Anthropic's Frontier Red Team published FlashProve, dated Monday, under an license. A provider commits to a fingerprint of its , then produces a that those weights generated a given reply, and anyone can check it without seeing the weights.
The use cases are catching a router that quietly swaps in a different model and checking that only approved models ran.
On Llama 4 Maverick, a 400-billion-parameter model at 4-bit, one H100 proves about 110 a second with proofs of about 18.78 MB, and a verifies one in about 17 seconds. That costs roughly 70 to 95 times the time of plain inference.
It covers , integer weights and 1,024-token windows, so sampled production traffic at long context is out of scope. At that price it is an auditing tool for a slice of requests.
Anthropic releases FlashProve, a prototype that proves which model wrote a response
Anthropic's Frontier Red Team published FlashProve, dated Monday, under an Apache 2.0 license. A provider commits to a fingerprint of its weights, then produces a zero-knowledge proof that those weights generated a given reply, and anyone can check it without seeing the weights. The use cases are catching a router that quietly swaps in a different model and checking that only approved models ran.
On Llama 4 Maverick, a 400-billion-parameter model at 4-bit, one H100 proves about 110 tokens a second with proofs of about 18.78 MB, and a CPU verifies one in about 17 seconds. That costs roughly 70 to 95 times the GPU time of plain . It covers greedy decoding, integer weights and 1,024-token windows, so sampled production traffic at long context is out of scope. At that price it is an auditing tool for a slice of requests. SourcesAA
Sources
- AProvable inference: FlashProve (Anthropic)anthropic.com
- Aanthropics/FlashProve (GitHub)github.com
This story ran in the Afternoon Brief of November 3rd 2026.