Summary
On October 9, 2026, Cloudflare expanded its Clef decision-model family on Workers AI with Clef-omni, which processes audio, video, images, and text in one pipeline. Cloudflare also made Clef inference up to 2.0x faster and cut Clef-flash's input price from $0.09 to $0.038 per million tokens.
What changed
Cloudflare launched Clef-omni, built on Qwen3-Omni-30B-A3B-Instruct with the speech-output components removed, a fully multimodal decision model handling audio, video, images, and text, priced at $0.15 per million input tokens with median latency around 130ms for text and 150ms for images. It made Clef inference 1.7x to 2.0x faster (for example 616ms to 305ms at about 3,400 tokens) by moving serving to SGLang, and cut Clef-flash's input price from $0.09 to $0.038 per million tokens while lowering its hosted context window from 64k to 24k tokens (open weights remain 256k for self-hosting).
Why it matters
Decision models return calibrated probabilities over fixed answer sets instead of free-form text, giving agents fast, deterministic routing and classification at control points. Full multimodality lets those decisions act directly on audio, image, and video, and the concrete price and latency cuts, with Clef-flash now undercutting rival Jev, push task-specialized decision models as a cheaper alternative to calling a general LLM for every classification or routing step.
Evidence excerpt
Clef-omni "takes in audio and video input alongside text and image," is built on Qwen3-Omni-30B-A3B-Instruct, and launched at $0.15/M input tokens (~130ms text, ~150ms image). Clef is 1.7-2.0x faster after moving serving to SGLang; Clef-flash's price dropped from $0.09 to $0.038/M and its hosted context from 64k to 24k tokens (open weights stay 256k).