Cloudflare Clef-omni: open decision model for audio and video

Not every AI call needs a text answer. Often an application needs a decision: is this spam, which team should take this ticket, how urgent is it. That is the job of a young class of models, and this week it got two pushes: OpenAI's Decisions API, in beta since 6 October, and on 9 October Clef-omni from Cloudflare (announcement), an open-weight decision model that also listens to audio and watches video.

What a decision model is

A decision model does not write. You hand it a situation and a set of questions, and it returns typed answers with probabilities:

• Yes or no: how likely a condition is true, for example "is there visible damage in this photo?"
• A choice from a fixed list, for example the department that should handle a message.
• A score on ordered levels, for example severity.

Because no text is generated, the answers come back fast. Cloudflare's Clef models score every allowed option in one pass over the input, and OpenAI says its Decisions API returns typed answers about ten times faster than the Responses API.

Clef-omni: text, images, audio and video

• Input: text, images, audio (wav or mp3) and video (mp4 or webm) in one call. The first Clef took images and video frames, but no sound.
• Base: Qwen3-Omni-30B-A3B-Instruct, a mixture of experts. Cloudflare keeps its understanding part, leaves out the speech output and adds a small head that scores all options of all questions at once.
• Speed, by Cloudflare's numbers: about 130 ms at the median for text, 150 ms for images, a few hundred milliseconds for audio clips and about 1.5 seconds for a 21-second video with sound.
• Price on Workers AI: $0.15 per million input tokens.
• Open: Apache 2.0, with weights on Hugging Face.

In Cloudflare's own tables, Clef-omni trails Clef on several text tasks, so its case is media, not text. The rest of the family changed too: Clef-flash dropped from $0.09 to $0.038 per million input tokens, but its hosted context shrank from 64k to 24k tokens (the weights stay trained for 256k). Clef stays at $0.24 and got about 1.7 to 2 times faster at the median on Workers AI. Clef speaks the API of Jev, the decision model from TypeSafe, so switching means changing the model ID.

OpenAI's Decisions API

According to the guide, the Decisions API is in public beta, with gpt-6-luna as the only model and its own endpoint, POST /v1/decisions. It takes text, images or both and answers with a probability, a choice or a score. OpenAI expects general availability "in the coming weeks". For anything that has to generate an object in your own schema, OpenAI points to Structured Outputs, and to function calling for tool calls.

Our take

Decision models are the unglamorous part of AI that ends up in every pipeline: spam filters, ticket routing, moderation, triage. That OpenAI and Cloudflare both push them in the same week says a lot about where the volume is. The difference is control: Clef's weights are open under Apache 2.0 and run on your own hardware, while the Decisions API is a hosted service with one model for now. Cloudflare's own examples show the kind of work it means: closing spam issues in its docs repository, checking plugin libraries for phishing and scanning data for personal information such as government IDs.

In AyCode

AyCode does not route your prompts through a decision model: you pick the agent and the model for each chat (docs). Second opinion asks another provider to review the task and recorded changes from one turn or the whole chat, with prioritized findings, file, line and a verdict (docs). Large diffs may be shortened; the answer is review help, not a guaranteed judgment. More on the agents in AyCode.