Cloudflare open-sources Clef-omni, a decision model that reads text, images, audio and video in one call, and cuts Clef-flash to $0.038 per million tokens

Why it matters

【Key points】

Cloudflare released Clef-omni on Oct 9, the third Clef decision model. Decision models don't generate text; they pick among given options with a probability for each. Clef-omni takes text, images, audio and video in one call.

It is built on Alibaba's open Qwen3-Omni-30B-A3B with a frozen backbone and LoRA adapters. Median latency is about 130 ms for text and 1.5 s for a 21-second video. Weights are on Hugging Face; hosted price is $0.15 per million input tokens.

Clef-flash drops from $0.09 to $0.038, which Cloudflare says undercuts Jev, but hosted context shrinks from 64k to 24k. Clef gets 1.7-2x faster. Clef and Clef-flash launched Oct 1 under Apache 2.0 on Qwen bases.

【Verifiable】

Blog: https://blog.cloudflare.com/clef-faster-cheaper-multimodal/

Original Clef launch (Oct 1): https://blog.cloudflare.com/clef-decision-models/

Benchmarks are self-reported; Cloudflare's table shows Clef-omni trailing Clef on several text tasks (home appliances 69.3 vs 82.95). X: 36-48 posts per hour since launch, about 250 in 24 hours.

【Our read】

1. Decision models hit a price war within a month: OpenAI Decisions API $0.10, Microsoft-Decision-1 $0.042, Clef-flash $0.038. Competition shifts to latency, input types and self-hosting.

2. Clef, Clef-omni and Microsoft-Decision-1 are all built on Qwen, so Chinese open weights now sit inside US cloud products.

3. Extra modalities cost text accuracy. Use Clef for text-only work; Clef-omni when you must judge audio or video directly.