JA EN
shiichan

Meet Clef-omni: audio and video decisions in one request, and Clef-flash gets cheaper

Cloudflare Changelog ★★★★ View Markdown
#cloudflare#ai#serverless

Hi, it's Shii-chan! The Clef family on Workers AI got a big update.

Cloudflare Changelog developers.cloudflare.com

What was announced?

Cloudflare's Changelog announced three Workers AI changes:

  • Clef-omni (@cf/cloudflare/clef-omni) is new
  • Clef-flash got a price cut and is now cheaper than Jev
  • Clef got faster

Clef is an open-weight family of decision models. They don't generate text; they score every allowed answer in a single pass.

The story so far

To make a decision about a voice recording or a video, you used to chain models: transcribe the speech, split the audio and visual tracks, then pass the results to a text decision model.

What changes

Clef-omni reads text, images, audio (WAV or MP3), and video (MP4 or WebM) directly in one request. A video's soundtrack is aligned with its frames, so the model can reason over what is seen and heard at the same time.

Dive Deep

Clef-omni specs:

  • A 30B-parameter mixture-of-experts (MoE) backbone with 3B active parameters
  • Latency: about 20 ms for text, under 100 ms for image or audio, and about 300 ms for a 21-second video clip with sound
  • Pass media as base64 data URLs in the images, audio, and videos fields
  • 64K context window, $0.150 per million input tokens
  • Image, audio, and video inputs are converted to input tokens and billed that way
  • Same System One API as Clef and Clef-flash, and works with AI Gateway. Change the model ID and set the model selector to clef-omni to switch

Changes to the other models:

  • Clef-flash: price cut from $0.090 to $0.038 per million input tokens. The hosted context window shrank from 64K to 24K (only 0.24% of requests exceed 24K). Use Clef for longer inputs. The Hugging Face weights are unchanged and support up to 256K
  • Clef: up to 2x faster, with unchanged weights. At about 3,400 tokens, median latency went from 616 ms to 305 ms and p95 from 777 ms to 531 ms. Part of the gain comes from moving to SGLang, and Clef support is coming in SGLang 0.5.22

Pricing and context side by side:

  • clef-flash: $0.038 / 24K
  • clef: $0.240 / 64K
  • clef-omni: $0.150 / 64K

On benchmarks, Clef-omni leads the family on BANKING77 (94.8), CLINC150+OOS (97.7), and Amazon ESCI (57.8).

Wrap-up

  • Clef-omni decides over text, images, audio, and video in one request
  • Clef-flash drops to $0.038, with a 24K hosted context
  • Clef is up to 2x faster
  • Switching is easy since it uses the same API as the other Clef models

It's a great fit for developers who want audio and video classification without building a pipeline.

Related posts