Decided on a Friday evening, trained over the weekend, launched on Thursday. That is how Cloudflare tells the story of how Clef, its first decision models, came together in its new post. A week after the launch come three changes, all in one post.
Fact
The facts: on 9 October 2026 Cloudflare released Clef-omni, a decision model in the Clef family that takes audio (wav or mp3) and video (mp4 or webm) in addition to text and images, in a single request. The weights are open on Hugging Face. The base is Qwen3-Omni-30B-A3B-Instruct (mixture-of-experts): the backbone is frozen, and low-rank adapters (LoRA) are trained on top, as with Clef. According to Cloudflare, a text decision comes back in about 130 milliseconds at the median, an image in about 150, and a 21-second video with sound is scored in about 1.5 seconds. Clef-omni costs $0.15 per million input tokens. Clef-flash gets cheaper, from $0.09 to $0.038 per million input tokens, which in Cloudflare's words makes it cheaper than Jev; its hosted version, however, now takes a context of up to 24k tokens instead of the advertised 64k. According to Cloudflare, only 0.24% of requests exceed 24k input tokens. The Clef-flash weights on Hugging Face are unchanged and support up to 256k when self-hosted. Clef stays at $0.24 and 64k and is faster: the median latency of the hosted version drops 1.7 to 2 times, and one of the optimisations is the move to SGLang; there are no new weights. Support for Clef will land in SGLang 0.5.22 (PR #42721).
The two changes you will feel are in the input, and they go in opposite directions: Clef-omni takes more kinds of input, and Clef-flash takes shorter input.
Until now a decision on a call recording has gone through two models: one turns the sound into text, the other decides on the text. Clef-omni does both in one call, without transcribing the sound into text. Cloudflare's example is an installation check: a photo of the unit, a recording of it running and a video of the fan, plus three yes or no questions - whether the serial number label is visible, whether the unit sounds smooth and whether the fan is running.
Now the other direction. Clef-flash falls from $0.09 to $0.038 per million input tokens. Cloudflare calls it cheaper than Jev, and the arithmetic holds: $0.038 against $0.042, the price TypeSafe announced for Jev in September. The discount, however, is paid for in context. The hosted Clef-flash now reads up to 24k input tokens instead of the 64k Cloudflare advertised a week ago.
One model now listens and watches. The other reads less.
If you are among those 0.24% of requests over 24k, Cloudflare itself advises you to switch to Clef, which stays at 64k and at $0.24. The two 0.24s are at least easy to remember. There is a third way too: the Clef-flash weights on Hugging Face are untouched and hold up to 256k when self-hosted. The ceiling applies only to the hosted version.
If you send call recordings to a model that turns them into text only to decide something about them, Clef-omni deserves a test. If you send long contracts to Clef-flash, the new price is not for you: above 24k the hosted version is no longer enough.
Official primary sources
Cloudflare - Introducing Clef-omni with full multimodality, plus a faster Clef and a cheaper Clef-flash, 09.10.2026TypeSafe - Introducing System One models and Jev (Jev pricing)The visual is generated code-art, no third-party images.