Clef-omni adds audio and video to structured decision scoring

Cloudflare’s new Clef-omni accepts audio and video alongside text and images, then scores allowed answers within a supplied schema. It uses a multimodal model backbone without generating a conversational response. This can combine media classification into one request instead of requiring an application to assemble separate transcription, captioning and decision stages.

Cloudflare · Oct 9, 2026

Why it matters

Teams already evaluating decision models have a concrete new input capability to test on their own labelled examples.

More context

Try the Workers AI endpoint with an account and token, or review the published weights for self-hosting requirements. This follows the earlier Clef release. The same announcement lowers hosted Clef-flash pricing but cuts its context window from 64k to 24k; larger requests need review.

Keep in mind

The hosted pricing tradeoff does not change the downloadable Clef-flash weights. Reported latency and accuracy are Cloudflare’s results, and the comparison table does not show one model winning every task. We have not evaluated multimodal scoring or confidence calibration.

Original read