blog.cloudflare.com web signal

Cloudflare Ships Multimodal Clef-omni, Cuts Flash to $0.038/M

By Published Updated
5 sources tracking this story
Multimodal Inference AI Infrastructure ai-business

TL;DR

  • Clef-Flash's new $0.038/M price sits below TypeSafe Jev, giving developers the first cheaper hosted alternative in the decision-model tier.
  • Cloudflare built Clef-Omni on Qwen3-Omni-30B open weights, revealing a productization strategy over proprietary model training.
  • Only 0.24% of production requests exceeded 24k tokens, giving Cloudflare data cover to shrink the context window and fund the price cut.

Cloudflare dropped the input price of Clef-flash from $0.09 to $0.038 per million tokens and launched a multimodal sibling, Clef-omni, that scores a 21-second video clip with sound in about 1.5 seconds for $0.15 per million. The announcement, posted October 9 by Michelle Chen on the Cloudflare blog, bills itself as a sequel: "Following last week's release of Clef and Clef-flash, Cloudflare's open-weight decision models, we decided to bring forth more gifts."

Clef-omni accepts audio, video, image and text input in a single call, built on a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation with what the post calls a two-stage attention routing approach. Text-only decisions return in roughly 130ms median latency, images in about 150ms.

The hosted Clef-flash context window shrinks from 64k to 24k tokens, which Cloudflare justifies with one line of usage data: "From our usage data, we see that only 0.24% of requests exceed 24k input tokens." Teams past that ceiling can self-host; the weights are public.

The second gift is a serving stack. Cloudflare moved Clef to SGLang and reports median speedups of 1.7x to 2.0x from infrastructure changes alone, with the integration landing as PR #42721 in SGLang 0.5.22. It arrives the same week we logged Tsinghua's TokenRouter hitting 2x to 64x decode speedups for token-level routing, one of 94 inference stories we've tracked in the last 90 days.

What others are reporting

Coverage cluster as of 24h after publish

  1. Cloudflare Developer Docs Read →

    First-party technical changelog with BANKING77 and CLINC150+OOS benchmark tables, per-input-size latency numbers, and base64 media ingestion code examples.

    Clef-omni reads every modality directly in one request.
  2. RuntimeWire Read →

    Frames the release as competitive positioning against Jev, critically noting the context-window shrinkage and the need for independent accuracy validation.

    Clef-omni is designed to return structured decisions from those inputs in one API call, avoiding separate transcription and media-processing steps.
  3. AIBase News Read →

    Identifies Qwen3-Omni-30B as the base architecture and frames the dual announcement as a tiered pricing play targeting different developer cost segments.

    Clef-omni natively supports audio and video formats such as wav, mp3, mp4, and webm, eliminating manual preprocessing steps.
  4. lavx.hu Read →

    Focuses on the competitive pricing victory over Jev and quotes the pipeline-collapse framing: audio synced to visual frames in a unified sequence.

    Clef-omni collapses that stack. Media maps directly into a unified sequence, with audio synced to visual frames.

Originally reported by blog.cloudflare.com

Read the original article →

Original headline: Cloudflare Launches Clef-Omni Multimodal, Cuts Clef-Flash Pricing 58% to $0.038/M