Clef Omni vs the Clef family and other decision models

Clef Omni is the fully multimodal member of Cloudflare's open-source Clef decision model family, served on Workers AI. It is built on a 30B mixture-of-experts backbone with 3B active parameters (Qwen3-Omni-30B-A3B). Decision models do not write text. You send a state (text, JSON, an image, audio or video) and a schema of typed questions, and one forward pass returns a probability for every allowed option.

Use them for classification, routing, scoring and guardrails, where you need a typed answer and its confidence. Cloudflare reports about 20 ms for text requests, under 100 ms for image or audio input, and about 300 ms for a 21-second video clip with sound.

Pricing: you pay for input only, because there is no generated output.

• Clef Flash (9B) at $0.038 / 1M, the cheapest of the family
• Clef Omni (30B-A3B MoE) at $0.15 / 1M, cheaper than dense Clef and the only one that takes audio and video
• Clef (27B dense) at $0.24 / 1M
• TypeSafe Jev 1.13 ($0.042) and Respan Span-01 ($0.02) are single-purpose alternatives. Jev returns a typed choice, and Span-01 scores behaviours in a conversation.
• GPT-6 Luna Decisions is OpenAI's closed option at $0.10 in / $0.50 out.

Caveats: Every Clef model has a 64K context. No vendor has published a cross-vendor accuracy benchmark for decision models yet, so check calibration on your own labels before you switch.

Introducing Clef-omni (Cloudflare blog) · Workers AI docs