---
id: "2026-11-02-anthropic-flashprove-provable-inference"
url: "https://postcutoff.com/e/2026-11-02-anthropic-flashprove-provable-inference/"
as_of: "2026-11-09T21:13:00+01:00"
date: "2026-11-02"
date_precision: day
category: research
importance: 3
confidence: high
status: [Confirmed]
sources: 4
editor: human editor
human_review: null
version: "2026-11-04"
---

As of: 2026-11-09 21:13 CET.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/e/2026-11-02-anthropic-flashprove-provable-inference/

# Anthropic open-sources FlashProve, a zero-knowledge prototype that proves a response came from committed model weights

On Nov 2, 2026 (X announcement Nov 3) Anthropic published FlashProve, an Apache-2.0 zkSNARK prototype that lets a verifier check that a model's output came from a specific committed set of weights without seeing them. Tested on Llama 4 Maverick, proving costs about 70–95x the GPU time of normal inference.

## Key facts
- Model tested: Llama 4 Maverick (400B MoE, ~17B active), third-party 4-bit conversion
- Proving ~110 tokens/s on one H100; proof ~19 MB per 1,024-token request; verification ~17 s on a 24-thread CPU
- Overhead: roughly 70–95x the GPU time of serving the model on vLLM
- Protocol properties (completeness, soundness, zero knowledge) proven in Lean; the fast C++ verifier is not formally verified
- Limits: integer model under greedy decoding only, 1,024-token context, architecture not hidden, side channels out of scope
- Anthropic's X post: 462k views (fxtwitter, Nov 4)

## What happened

Anthropic's research page describes FlashProve as "an open-source performant prototype for provable inference." The provider publishes a
commitment to its weights once; afterwards it produces proofs for batches of requests, and anyone holding the commitment can check them
without the weights. Anthropic's X post names two uses: making it harder to hijack a chat, and letting safety evaluators confirm that the
model they tested is the one served to customers.

## Why it matters

Evaluators and governments have had no way to check that the deployed model is the audited one. The prototype is far from production cost
(70–95x GPU time, short contexts, greedy decoding only), and Anthropic presents it as research, not a product.

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 186 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 125 days after its cutoff
- Gemini 4 Argon (training cutoff March 2026): 216 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 155 days after its cutoff

## Sources

1. [Anthropic: Provable inference (FlashProve)](https://www.anthropic.com/research/provable-inference) (anthropic.com, official)
2. [FlashProve code (GitHub, Apache 2.0)](https://github.com/anthropics/FlashProve) (github.com, code)
3. [Anthropic on X (Nov 3)](https://x.com/AnthropicAI/status/2117710942826123294) (x.com, official)
4. [TechCrunch: Anthropic is working on a system to prove AI models' identity](https://techcrunch.com/2026/11/03/anthropic-is-working-on-a-system-to-prove-ai-models-identity/) (techcrunch.com, press)

## Changes

- 2026-11-04 (filed): Created

## Related

- Archived: 1 post, listed in https://postcutoff.com/e/2026-11-02-anthropic-flashprove-provable-inference/index.json