# Cloudflare releases Clef-omni with audio and video input, cuts Clef-flash price and speeds up Clef

published: 2026-10-12T13:41:57.485Z
topic: infrastructure
source: https://blog.cloudflare.com/clef-faster-cheaper-multimodal/
url: https://aiagentnews.io/stories/cloudflare-releases-clef-omni-with-audio-and-video-input-cuts-clef-flash-price-a

Cloudflare has released Clef-omni, an open-weight decision model that accepts audio and video alongside text and images. It also lowered the hosted price of Clef-flash and made the hosted Clef model faster.

Clef-omni takes audio in WAV or MP3 format and video in MP4 or WebM format, in addition to text and images. Earlier Clef models handled images and video frame arrays but not audio. Cloudflare says developers can now make decisions across all these modalities in one model call, without chaining speech-to-text or splitting video into separate streams. The model is built on the Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation and its weights are on Hugging Face.

Cloudflare cut the hosted price of Clef-flash to $0.038 per million input tokens from $0.09, saying this makes it cheaper than Jev. The hosted context window for Clef-flash drops to 24k from 64k. The Hugging Face weights still support a 256k context window for self-hosting, and Cloudflare reports that 0.24% of requests exceed 24k input tokens.

Cloudflare made the hosted Clef model faster through serving-layer changes on Workers AI, without new weights. Median latency fell for inputs of about 800, 3,400 and 16,000 tokens, with speedups between 1.7x and 2.0x. Part of the change is a move to SGLang for serving, and the pull request for Clef support is slated for SGLang 0.5.22. Clef stays at $0.24 per million input tokens, and Clef-omni launched at $0.15.

## Why it matters if you build agents

Agent builders can send audio and video directly to one decision call instead of building transcription or frame-splitting steps first. Clef-flash is cheaper now, but its hosted 24k window means longer inputs should go to Clef, which keeps 64k.

Clef is described as fully Jev-API compatible and available through AI Gateway, so trying it should mainly require changing the model ID.