Cloudflare rolled out Clef-omni on October 9, putting a new 30-billion-parameter model on Workers AI that does something most AI launches don't: it refuses to chat. Clef-omni is a decision model — it reads text, JSON, images, audio, and video in a single request and answers with calibrated probabilities over a schema of typed questions you define, with no prose and no output parsing. It is built on a 30-billion-parameter mixture-of-experts backbone with roughly 3 billion active parameters, post-trained from Qwen3-Omni-30B-A3B-Instruct, served as @cf/cloudflare/clef-omni at $0.15 per million input tokens, with open weights published on Hugging Face.
The category is the news as much as the model. A decision model takes a "state" — the content under review — plus a schema of questions such as whether a clip violates policy, each with fixed answer options, and returns a probability for every option of every question in one forward pass. Cloudflare quotes about 20 milliseconds for text requests and under 100 milliseconds for image or audio. Compare that to the standard approach: generate a paragraph of judgment with a chat model, then scrape a verdict out of the prose. Every moderation queue, ticket router, and upload screener in production has that parsing step somewhere, and it is exactly the kind of glue that breaks at 3 AM. Decision models delete the parsing problem by contract.
The genuinely new part is the omni in Clef-omni. Judging a video with a soundtrack previously meant a chain: transcribe the audio, split and caption the frames, then feed both into a text decision model — every stage adding latency and a fresh chance to garble something. Clef-omni aligns what it sees with what it hears natively. That collapse from a pipeline to a single call is the kind of engineering win that shows up directly in latency percentiles and in the error budget you no longer spend on stitching.
Keep the evaluation picture in perspective, though. Cloudflare's own framing and independent write-ups put Clef-omni close to its sibling Clef on routing and intent tasks but noticeably behind on hard reasoning — one benchmark summary has it at 47.4 on GPQA Diamond, while it tops the CLINC150+OOS intent set at 97.7 macro-F1. This model is a judge, not a thinker. Point it at classification, routing, scoring, and guardrails — Cloudflare's stated use cases — and evaluate it on your own policy categories, not the benchmark headline. Your categories are never the benchmark's.
The surrounding pricing move deserves attention too. The same October 9 update cut Clef-flash from $0.09 to $0.038 per million input tokens — undercutting TypeSafe's Jev — and roughly doubled hosted Clef's speed via an SGLang serving upgrade, all behind one unchanged System One API. Cloudflare launched the Clef family on October 1 and shipped the omni-modal third member eight days later, which reads as a company treating decision models as a competitive battleground rather than a side project. There is a distribution logic that should worry incumbents: guardrail workloads run on every piece of user content, making them far more cost- and latency-sensitive than chatbots, and Cloudflare already sits in front of much of that traffic. For builders, the broader signal is the pattern: post-training an open-weight omni-modal base into a narrow commercial role is faster and cheaper than pretraining from scratch, and the differentiation is moving to the fine-tuning and serving layer. If your product ingests voice notes, video clips, and images alongside text, a single scoring call now beats a pipeline — test it against your own categories before you believe the brochure.
