Cloudflare's Clef-omni scores audio and video against a decision schema in one call

A Clef-omni request is one POST to the Workers AI endpoint for `@cf/cloudflare/clef-omni` ^[10]. The JSON body holds a `state` string, base64 data URIs in `images`, `audio` and `videos` arrays, and a `questions` object in which each key carries a type and plain-language instructions ^[10]. Audio goes in as WAV or MP3, video as MP4 or WebM ^[1]. The published sample reviews one installation. It asks whether the model and serial number label is visible in the photo, whether the unit runs without rattling or grinding, and whether the fan is running in the video ^[11]. The model returns a probability for each option the schema allows ^[9].

Inside, the model makes one prefill pass over the complete payload and generates no output tokens, according to Cloudflare ^[4]. Every valid option first gathers evidence from the input. Field vectors then cross-attend across the full context to compute confidence scores ^[5]. Runtimewire, citing the model card, describes a joint schema head that scores options across questions together ^[6]. I think this is the right design for picking from a fixed menu. There is no decode loop to pay for, and teams that route on a confidence threshold get a model trained for calibration. Training freezes the Qwen3 backbone, fits LoRA adapters, and pairs label-smoothed cross-entropy with Brier score calibration ^[12].

Cloudflare's post makes the pipeline case in one sentence: "Instead of setting up cascading pipelines of models that transcribe speech-to-text, or splitting audio and image channels from video, you can just call one model to make decisions across any modality." ^[2]

Cloudflare's latency medians are about 130 milliseconds for text, about 150 for images and a few hundred for audio, and it timed a 21-second video with sound at about 1.5 seconds ^[13]. The pass covers the whole payload ^[4]. For the video figure to transfer, your media has to resemble that one clip. A ten-minute support call is a much larger prefill. Runtimewire noted that the published material does not include independent testing of accuracy or latency on customer workloads ^[14].

Workers AI lists Clef-omni at $0.15 per million input tokens and Clef at $0.24 ^[18]. The multimodal model costs 37.5% less per input token ^[23]. The scoreboard comes from Cloudflare's own evaluations ^[8]. Clef-omni posts 94.8 macro-F1 on BANKING77 and 97.7 on CLINC150+OOS, and trails Clef on several other tests ^[7]. On the invoice-processing exact-action workflow it scores 60.2, against 64.7 for Clef and 61.8 for TypeSafe's Jev ^[8]. That puts it 4.5 points behind Clef and 1.6 behind Jev ^[17].

For a model line sold on unambiguous answers, the Jev price claim is ambiguous. Cloudflare's post says the cut makes Clef-flash cheaper than Jev ^[26]. Runtimewire attributes the claim to Clef-omni and adds that the comparison depends on the applicable Jev price and workload ^[27].

On Clef-flash, the hosted price fell about 58% ^[22] while the hosted window shrank 62.5% ^[24]. Cloudflare's share of oversized requests is measured across all of its Clef-flash traffic ^[20]. It transfers to a given team only if that team's prompt lengths match Cloudflare's mix. One customer sending long contracts could make up much of that tail. Publishing the share was the right call, and so was leaving the weights alone. Self-hosters keep a 256,000-token window ^[21], more than ten times the hosted limit ^[25].