Anthropic publishes a prototype tying an answer to fixed weights
single source· 1 articles · confidence: low · first seen 2026-11-03 22:08 UTC
What this means for you
Nothing to act on yet: this is a prototype, not a service or an API. No endpoint, price or date exists, and the throughput figure covers one model on one H100 — not a general claim that any model's output can be proved at that speed.
Anthropic has released FlashProve, an Apache 2.0 prototype that uses a zkSNARK — a cryptographic proof that a computation was carried out correctly, checkable without redoing it or seeing the hidden inputs — to show that a model's reply came from weights committed to in advance, without revealing those weights. In the published configuration it proves replies from Llama 4 Maverick at 110 tokens per second on a single H100, with proofs of 19 MB. MarkTechPost reports the figures as Anthropic's own; no hosted service, price or availability date was given.
Models in this story
Key facts
- ·FlashProve is released under the Apache 2.0 licence as a zkSNARK prototype. source
- ·The prototype proves replies from Llama 4 Maverick at 110 tokens per second on a single H100. source
- ·Proofs in that configuration are 19 MB. source
- ·The stated purpose is proving a reply came from committed weights without revealing them. source
- ·MarkTechPost published its report on 3 November 2026. source
What the sources say
- MarkTechPost — Brief report giving the licence, one model, one GPU, the throughput and the proof size.
Sources
The original reporting. Follow these — they did the work.