Cloudflare released Clef-omni on 9 October 2026, an open-weight decision model that takes text, images, audio and video in one call. Like the rest of the Clef family, it returns a probability per allowed answer instead of writing text, so a pipeline no longer needs a transcription or captioning step first. In the same post Cloudflare cut Clef-flash to $0.038 per million input tokens, below TypeSafe’s Jev.
Clef-omni sorts messages well and trails on agent tasks
Clef-omni is Alibaba’s Qwen3-Omni-30B-A3B with the speech output removed and a decision layer trained on top. Only 3 of its 30 billion parameters work on each token, which keeps it fast. Adding audio and video cost some accuracy on agent work, while message sorting held up.
| What is measured | Clef-omni | Alternative |
|---|---|---|
| BANKING77 (sorting bank support messages) | 94.8 | Clef: 94.20; Jev: 79.74 |
| CLINC150+OOS (intents, incl. off-topic requests) | 97.7 | Clef: 97.43; Jev: 89.27 |
| When2Call (knowing when to call a tool) | 63.3 | Clef: 72.37; Jev: 80.97 |
| Home appliances (exact tool calls) | 69.3 | Clef-flash: 97.73 |
| Invoice workflow, primary action | 82.0 | Clef: 86.2; Jev: 83.1 |
| Input price per million tokens | $0.15 | Clef: $0.24; Clef-flash: $0.038 |
Each score here is Cloudflare’s own. The invoice row comes from TypeSafe’s workflow evaluations, as reported by Cloudflare. The post does not say whether Cloudflare reran Jev itself, and it publishes no audio or video benchmark, so the new capability is the one claim without a score.
Scoring a minute of 480p video costs about $0.0014
Cloudflare counts about 8,600 tokens per minute for 480p video, plus 780 for sound. By our calculation that comes to about $0.0014 per minute, rising to about $0.0024 at full resolution. Median latency is about 130 milliseconds for text and 1.5 seconds for a 21-second clip.
- Model ID: @cf/cloudflare/clef-omni on Workers AI
- Weights: Cloudflare/clef-omni on Hugging Face, Apache 2.0
- Self-hosting: one H200-class GPU, about 64 GB in bfloat16
- Limits: 64,000 tokens, 2 videos of 60 seconds, 4 audio clips of 5 minutes
- Catch: hosted Clef-flash context drops from 64,000 to 24,000 tokens
Clef-omni is worth a pilot for teams that now transcribe or caption media only to make a decision about it: flagging uploaded videos, sorting recorded support calls, checking product photos. Run one existing stream through it and compare cost and accuracy with your current two-step setup. Do not move text-only agent decisions to it, because Clef and Clef-flash score higher there. Cloudflare says nothing about where these requests are processed, so check that before sending customer recordings.
For more information, visit the official announcement of Clef-omni on the Cloudflare blog.
