Cloudflare launches Clef-omni and cuts Clef-flash pricing—with a lower hosted context limit

Cloudflare has announced Clef-omni, a decision model that accepts text, images, audio and video, alongside lower hosted Clef-flash pricing and serving improvements for Clef. The company says Clef-omni is available through its API and that its open weights have been released on Hugging Face. Cloudflare announcement

Multimodal decisions, not generated responses

Clef-omni accepts WAV/MP3 audio and MP4/WebM video alongside text and images. Cloudflare says it uses the comprehension backbone of Qwen3-Omni-30B-A3B-Instruct, without the text-to-speech components, and scores schema-defined options rather than generating response tokens. Its announced launch price is $0.15 per million input tokens. Architecture and pricing

For workflows whose required output is a decision, this offers a potential alternative to maintaining a separate transcription-and-classification pipeline. Broader input support, however, does not imply better results on every task: Cloudflare’s invoice-processing benchmark reports an exact-actions score of 60.2 for Clef-omni versus 64.7 for Clef. Engineers should evaluate the models against their own workflows rather than choose solely on modality support. Cloudflare’s benchmark results

Cheaper Flash, smaller hosted context

Cloudflare’s announced Clef-flash price falls from $0.09 to $0.038 per million input tokens, a reduction of approximately 58%. Its hosted context limit is now 24k tokens, down from the previously advertised 64k. Cloudflare says the downloadable weights are unchanged and were trained to support a 256k context window; that training context is distinct from the hosted service’s limit. Clef remains at $0.24 per million input tokens with a 64k hosted context window. Pricing and context limits

The practical migration check is request size. Teams using Clef-flash should identify requests that exceed the new 24k limit and plan to reduce their inputs or use another endpoint. The lower price should not be evaluated independently of the reduced hosted capacity.

Faster hosted Clef, planned upstream SGLang support

Cloudflare reports that Clef’s median latency for requests with roughly 3,400 input tokens fell from 616 ms to 305 ms on Workers AI. These are vendor-reported measurements of hosted serving improvements, not independent benchmarks or results from new model weights. Cloudflare says no new Clef weights were released for these optimizations. Serving performance

The company attributes part of the improvement to moving serving to SGLang. In the announcement, Cloudflare identifies SGLang 0.5.22 as the planned release for the upstream Clef integration—a separate milestone from the improvements it reports for its hosted endpoint. SGLang integration plans