Cloudflare Cuts Clef-flash Price from $0.09 to $0.038 per Million Tokens and Adds Clef-omni
Cloudflare has updated Clef, a family of open models that don't write text but choose an answer from predefined options, such as "spam or not." Clef-flash dropped from $0.09 to $0.038 per million input tokens (a token is a chunk of text that serves as the billing unit), and the new Clef-omni accepts text, photos, audio, and video in a single request. Standard Clef is now 1.7–2x faster. For moderation, this means one call instead of a chain of models.
AI-processed from Cloudflare AI Blog; edited by Hamidun News
Imagine you need to automatically filter spam in issues, catch phishing in plugins, or search files for passport data. You don't need a chatbot that writes long answers for that. A model that picks an option from a ready-made list ("spam / not spam", "contains personal data / doesn't") and says how confident it is will do. Cloudflare builds exactly these "solver models," and the family is called Clef. This week the company released Clef-omni, which understands not only text and images but also audio and video. The cheapest version, Clef-flash, dropped from $0.09 to $0.038 per million input tokens (a token is a chunk of text; requests are billed by token count). Standard Clef is now 1.7–2x faster.
What's new in Clef-omni
Clef doesn't answer in words. You send a description of the situation and a list of questions, and the model answers each one and returns a confidence score.
An example from Cloudflare's blog: checking the installation of an air-conditioning unit. You send a photo of the device, a recording of it running, and a video of the fan, then ask: is the serial number plate visible, is there any rattling or grinding, is the fan spinning? Previously you would have had to transcribe the sound separately and extract frames from the video. Now everything goes into one request. WAV and MP3 are supported for audio, MP4 and WebM for video.
Underneath is Qwen3-Omni-30B-A3B-Instruct, a mixture-of-experts model (only a subset of "experts" activates per response). Cloudflare took the part of the model that "understands" the input and removed the speech synthesis block. Since Clef-omni generates nothing, it doesn't need intermediate captions for the files: it makes one pass over the entire request and immediately scores all allowed options. Training works the same way as for Clef: the base is frozen, small LoRA adapters are fine-tuned on top, and the model's confidence is calibrated with the Brier score (it penalizes confidently wrong answers).
Claimed speed: text decisions take about 130 ms at the median, images about 150 ms, and audio several hundred milliseconds. A 21-second video with sound is processed in roughly 1.5 seconds. The weights are published on Hugging Face.
Cheaper and faster: what changed in Clef-flash and Clef
Clef-flash now costs $0.038 per million input tokens instead of $0.09. Cloudflare says this is cheaper than Jev from TypeSafe, a competing model. Standard Clef stays at $0.24, and Clef-omni starts at $0.15.
The discount has a cost. Clef-flash, as hosted in Workers AI, now accepts up to 24 thousand tokens instead of the 64 thousand previously stated. According to the company's statistics, the limit is exceeded by only 0.24% of requests. If your documents are longer, switch to Clef with the 64k window. The weights on Hugging Face haven't changed: they were trained for context up to 256k, and self-hosting won't impose the 24k limit.
Standard Clef got faster without changing its weights: the optimizations were made on the server side, and serving was moved to SGLang. Cloudflare submitted a pull request adding Clef support to that project (it will ship in SGLang 0.5.22). Median response times:
- about 800 tokens: from 262 to 152 ms (1.7x);
- about 3,400 tokens: from 616 to 305 ms (2x);
- about 16,000 tokens: from 2,721 to 1,635 ms (1.7x).
What the tests show and where the pitfalls are
The benchmarks in the blog should be read with caution: these are the company's own numbers. Clef-omni handles function calling well (BFCL test: 98.2% exact match) and user intent detection: CLINC150 at 97.7 and BANKING77 at 94.8 on macro-F1 (which averages quality across all classes). On the API-use test (API-Bank), it scored 92.7%, outperforming standard Clef (91.93%).
But there is no universal win. On When2Call, where the model decides whether a tool needs to be called, Clef-omni scores 63.3% against 72.37% for Clef and 80.97% for Jev. On home appliances, it scores 69.3% exact match, while Clef-flash scores 97.73%. On the PhishNChips phishing test, it scores 73.2% against 79.6% for Clef. In TypeSafe's operational scenarios, such as invoice processing, standard Clef is also ahead: 64.7 versus 60.2 on exact actions.
This makes sense: multimodality makes a model more general, but not more accurate on every narrow task. If you only have text, Clef or Clef-flash look like the more sensible choice.
What to do in practice
Cloudflare itself uses Clef to automatically close spam in its documentation repository, to check EmDash CMS plugins for phishing, to find personal data such as government IDs, and to detect malicious domains. The API is compatible with Jev and available through AI Gateway, so trying it out only requires changing the model identifier.
A sensible approach: Clef-flash for high-volume text checks under 24k tokens, Clef for long documents and maximum accuracy, and Clef-omni when the stream includes voice and video. Before launching, run the models on your own sample: the home appliance and phishing tests show that the gap between models can be very large on a specific task.
- Nov 21, 2026AURA 2.0 Video Model Starts Talking on Its Own: Response Audio Arrives in 344 ms
- Nov 21, 2026Anthropic raised a record $81.3B in its IPO, but the founders kept control
- Nov 21, 2026ONNX Runtime vanishes from GitHub: Microsoft hides repositories after package hack
- Nov 21, 2026Microsoft Lets Businesses Fine-Tune MAI Models on Their Own Data in Azure
Key Questions Answered
What's new in Clef-omni?
Clef doesn't answer in words. You send a description of the situation and a list of questions, and the model answers each one and returns a confidence score. An example from Cloudflare's blog: checking the installation of an air-conditioning unit. You send a photo of the device, a recording of it running, and a video of the fan, then ask: is the serial number plate visible, is there any rattling or grinding, is the fan spinning?
What are the key details on Cheaper and faster: what changed in Clef-flash and Clef?
Clef-flash now costs $0.038 per million input tokens instead of $0.09. Cloudflare says this is cheaper than Jev from TypeSafe, a competing model. Standard Clef stays at $0.24, and Clef-omni starts at $0.15. The discount has a cost. Clef-flash, as hosted in Workers AI, now accepts up to 24 thousand tokens instead of the 64 thousand previously stated. According to the company's statistics, the limit is exceeded by only 0.24% of requests. If your documents are longer, switch to Clef with the 64k window.
What the tests show and where the pitfalls are?
The benchmarks in the blog should be read with caution: these are the company's own numbers. Clef-omni handles function calling well (BFCL test: 98.2% exact match) and user intent detection: CLINC150 at 97.7 and BANKING77 at 94.8 on macro-F1 (which averages quality across all classes). On the API-use test (API-Bank), it scored 92.7%, outperforming standard Clef (91.93%). But there is no universal win.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.