Published
Updated
Clef-omni scores a 21-second video in about 1.5 seconds, at $0.15 per million input tokens
This topic was created 21 days ago, and the information it contains may have evolved or changed since then.
Cloudflare's October 9, 2026 blog prices Clef-omni at $0.15 per million input tokens. The same post cuts hosted Clef-flash from $0.09 to $0.038 and from a 64k window to 24k. A 21-second video with sound scores in about 1.5 seconds. On that post's ten-benchmark table, Clef-omni's BANKING77 macro-F1 is 94.8. The weights are Apache 2.0, from Qwen3-Omni-30B-A3B-Instruct. We did not call the API.
On October 9, 2026 at 23:11 UTC, @Cloudflare posted Clef-omni. Michelle Chen’s blog is the write-up. The page’s schema marks it published at 18:27 UTC that day. The post has no attached still. The October 1 models stay on Clef.
Clef-omni takes text, images, audio, and video in one request. The blog’s curl posts to @cf/cloudflare/clef-omni with images, audio, and videos arrays. Audio is wav or mp3. Video is mp4 or webm. The sample asks three noul questions. We did not send that call.
The backbone the blog names is Qwen3-Omni-30B-A3B-Instruct. It says the model keeps that comprehension stack and drops the text-to-speech output. Training is the Clef recipe: the backbone stays frozen, LoRA adapters are trained, and the loss is label-smoothed cross-entropy plus a Brier term. The Hugging Face card, opened October 10, lists license apache-2.0 and that base model. The safetensors file counts 35,259,818,545 BF16 parameters. The repo was created October 9. We did not run the weights.
The same blog reprints prices. Clef-flash moves from $0.09 to $0.038 per million input tokens. Clef stays at $0.24. Clef-omni launches at $0.15. Hosted Clef-flash context moves from the advertised 64k to 24k. The blog says the weight files on Hugging Face are unchanged and were trained for 256k, and that 0.24% of requests exceed 24k. Clef’s hosted window stays 64k. The $0.09 figure remains the October 1 card reading on the earlier page. We did not open the docs URL the blog names as the live price list.
Clef-omni’s own times in that post are about 130 ms median for text, about 150 ms for an image, a few hundred milliseconds for audio, and about 1.5 seconds for a 21-second video with sound.
A second table on the same page times the hosted Clef model, the 27B one, after a server change. The blog says the weights did not change and names SGLang, including pull 42721 for SGLang 0.5.22. We did not open the pull.
| Input size | Before, median / p95 | After, median / p95 | Median speedup |
|---|---|---|---|
| ~800 tokens | 262 / 438 ms | 152 / 351 ms | 1.7× |
| ~3,400 tokens | 616 / 777 ms | 305 / 531 ms | 2.0× |
| ~16,000 tokens | 2,721 / 3,250 ms | 1,635 / 1,805 ms | 1.7× |
Those three rows are not the October 1 medians of 209.3 ms and 38.8 ms.
The benchmark grid adds a Clef-omni column to the October 1 ten-row table. The Clef, Clef-flash, and Jev cells match that earlier print. Clef-omni leads three of the ten: BANKING77 macro-F1 94.8, CLINC150+OOS macro-F1 97.7, and Amazon ESCI macro-F1 57.8. Jev still leads When2Call at 80.97, where Clef-omni is 63.3, and BRIGHT at 47.52, where Clef-omni is 42.0. Home appliances is the wide gap again: Clef-flash 97.73, Clef 82.95, Clef-omni 69.3, Jev 52.27.
| Benchmark | Clef-omni | Clef | Clef-flash | Jev |
|---|---|---|---|---|
| BFCL, case exact | 98.2 | 98.47 | 98.76 | 95.75 |
| ToolRet, nDCG@10 | 66.6 | 69.19 | 66.43 | 65.28 |
| API-Bank, accuracy | 92.7 | 91.93 | 93.11 | 88.19 |
| Home appliances, case exact | 69.3 | 82.95 | 97.73 | 52.27 |
| When2Call, accuracy | 63.3 | 72.37 | 65.58 | 80.97 |
| BANKING77, macro-F1 | 94.8 | 94.20 | 90.93 | 79.74 |
| CLINC150+OOS, macro-F1 | 97.7 | 97.43 | 66.77 | 89.27 |
| BRIGHT, nDCG@10 | 42.0 | 45.91 | 39.26 | 47.52 |
| Amazon ESCI, macro-F1 | 57.8 | 57.48 | 57.39 | 55.21 |
| PhishNChips, accuracy | 73.2 | 79.60 | 75.05 | 62.55 |
BANKING77 here is macro-F1. Laya’s author table and Charly Poly’s encoder bake-off use other Banking77 setups. PhishNChips 73.2 is this column. The October 1 card’s RAGTruth F1 row is a different print and is not in this table.
The workflow table is TypeSafe’s eval suite, a second setup. Invoice exact actions: Clef-omni 60.2, Clef 64.7, Clef-flash 57.1, Jev 61.8. Invoice primary action: 82.0, 86.2, 73.3, 83.1. Customer service exact: 71.6, 76.3, 77.0, 76.0. Security exact: 61.7, 62.9, 61.7, 61.7. Agent trace primary action: 65.8, 68.5, 69.8, 71.6. Jev leads the agent-trace row. Clef leads both invoice rows.
The blog names four internal uses and does not print a score for them: spam issues on the docs repo, phishing checks in EmDash, PII scans for data loss prevention, and malicious-domain checks. Those sentences are not rows in either table.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
@Cloudflare, October 9, 2026, 23:11 UTC, status 2108696873510850758. The text says Clef-omni processes audio, video, images, and text in one pipeline, that Clef-flash pricing was lowered, and that Clef inference is up to 2.0 times faster. The link is the blog below. The post has no attached still.
Blog https://blog.cloudflare.com/clef-faster-cheaper-multimodal/, byline Michelle Chen. Schema datePublished 2026-10-09T18:27:34.129Z, dateModified 2026-10-09T22:14:16.400Z. Prices printed there: Clef-flash was $0.09 per million input tokens and is now $0.038. Clef stays $0.24. Clef-omni is $0.15. Hosted Clef-flash context is 24k, where the earlier advertisement was 64k. The blog says the Hugging Face weights are unchanged and were trained for a 256k window, and that 0.24% of requests exceed 24k. Clef stays at 64k. We did not open the developer docs page the blog points at for a second price reading.
Clef-omni latency in that post: text about 130 ms at the median, images about 150 ms, audio a few hundred milliseconds, a 21-second video with sound about 1.5 seconds. The Clef serving table, same post, is a different measurement. About 800 tokens: median and p95 from 262 / 438 ms to 152 / 351 ms, 1.7 times. About 3,400 tokens: 616 / 777 to 305 / 531, 2.0 times. About 16,000 tokens: 2,721 / 3,250 to 1,635 / 1,805, 1.7 times. The post says those changes are in the server, names SGLang, and names pull 42721 for release 0.5.22. We did not open that pull.
Ten-benchmark table, Clef-omni, Clef, Clef-flash, Jev. BANKING77 macro-F1 94.8, 94.20, 90.93, 79.74. CLINC150+OOS 97.7, 97.43, 66.77, 89.27. Amazon ESCI 57.8, 57.48, 57.39, 55.21. The Clef, Clef-flash, and Jev cells match the October 1 blog print. When2Call is 63.3, 72.37, 65.58, 80.97. BRIGHT is 42.0, 45.91, 39.26, 47.52. Home appliances is 69.3, 82.95, 97.73, 52.27.
Workflow table, same post, labeled TypeSafe evals. Invoice exact actions 60.2, 64.7, 57.1, 61.8. Invoice primary action 82.0, 86.2, 73.3, 83.1. Customer service exact 71.6, 76.3, 77.0, 76.0. Security exact 61.7, 62.9, 61.7, 61.7. Agent trace primary 65.8, 68.5, 69.8, 71.6.
Hugging Face API for Cloudflare/clef-omni, opened October 10. cardData.license is apache-2.0. base_model is Qwen/Qwen3-Omni-30B-A3B-Instruct. safetensors.parameters.BF16 is 35259818545. createdAt is 2026-10-09T04:10:22.000Z. We did not download the weights or call @cf/cloudflare/clef-omni.
Compare
The October 1 median row, 209.3 ms and 38.8 ms against a printed Jev cell of 524.1 ms, stays on the Clef page. The October 9 Clef table is three input sizes on the hosted 27B model. The October 9 Clef-omni times, about 130 ms for text and about 1.5 seconds for a 21-second video, are a third print.
BANKING77 macro-F1 94.8 is this blog's ten-row table. Laya's 0.425 against a published Jev 0.870, and Charly Poly's encoder bake-off, are other Banking77 setups. Fastino's October 3 chart, Clef 61.2 and Clef-flash 57.1, is a Decision Index composite. This post does not print that composite for Clef-omni.
The October 1 cards listed Clef-flash at $0.09 and a 65,536-token window. The October 9 blog lists $0.038 and a hosted 24k window, and says the weights file was trained for 256k. Jev's $0.042 price on this desk is TypeSafe's models page, not a cell in Cloudflare's table.
Terms
- Clef-omni
- Cloudflare's October 9, 2026 decision model, hosted as @cf/cloudflare/clef-omni at $0.15 per million input tokens. Apache 2.0 weights from Qwen3-Omni-30B-A3B-Instruct. The blog's BANKING77 cell is macro-F1 94.8.
- Clef-flash price, October 9
- The October 9 blog lists hosted Clef-flash at $0.038 per million input tokens and a 24k context window. The October 1 card listed $0.09 and 65,536 tokens. The blog says the Hugging Face weights were trained for 256k and were not replaced.
- Clef serving times, October 9
- Median and p95 on the hosted Clef model after the October 9 server change. About 800 tokens moves from 262 / 438 ms to 152 / 351 ms. About 3,400 tokens moves from 616 / 777 ms to 305 / 531 ms. About 16,000 tokens moves from 2,721 / 3,250 ms to 1,635 / 1,805 ms. This is not the October 1 median of 209.3 ms.