Cloudflare cuts Clef-flash price to $0.038 and its context to 24k tokens

Cloudflare cuts Clef-flash price to $0.038 and its context to 24k tokens

9 October 2026: Cloudflare has cut the price of Clef-flash, the smallest model in its Clef family of decision models on Workers AI, from $0.09 to $0.038 per million input tokens. The hosted model's context window shrinks at the same time, from the 64,000 tokens advertised at launch to 24,000, so a workflow that sends Clef-flash longer inputs now has to move to a larger model.

Not yet independently verified. Single main source: Cloudflare's own post. The pricing page confirms the three prices but does not state context windows, and the post does not name a licence for the open weights. We will update this when it can be confirmed, and remove this note.

A cheaper Clef-flash that accepts shorter inputs

Cloudflare ties the smaller window directly to the lower price. By its own count only 0.24% of requests to Clef-flash exceed 24,000 input tokens, and it points customers with longer inputs to Clef, which keeps its 64,000-token window at $0.24 per million input tokens.

The limit applies to the hosted service only. The weights on Hugging Face are unchanged and were trained for a 256,000-token window, which still applies to anyone who runs the model themselves.

The price is also listed in Cloudflare's developer documentation, which converts it into the Neurons that Workers AI bills against: 3,455 Neurons per million input tokens for Clef-flash, 13,636 for Clef-omni and 21,818 for Clef. Every Workers account gets 10,000 Neurons a day at no charge, which at the Clef-flash rate covers about 2.9 million input tokens daily, and usage beyond that on a paid plan costs $0.011 per 1,000 Neurons.

Clef-omni adds audio and video at $0.15

The new model, Clef-omni, takes audio in wav or mp3, video in mp4 or webm, images and text in a single call, where Clef handled images and arrays of video frames. Cloudflare says decision models have been largely text-only since Jev, the model from TypeSafe whose API Clef is described as fully compatible with, and prices Clef-omni at $0.15 per million input tokens.

Built on a Qwen3-Omni backbone

Cloudflare describes Clef-omni as built on the Qwen3-Omni-30B-A3B-Instruct mixture-of-experts model. It keeps the comprehension backbone and discards the text-to-speech output components. The backbone is frozen and Cloudflare trains low-rank adapters on top of it.

Like the other Clef models, it does not write an answer. It runs one pass over the whole payload and scores the candidate values for each question, so there is no transcription or captioning step. The post gives median latencies of about 130 milliseconds for text-only decisions, 150 for images, a few hundred for audio clips and about 1.5 seconds for a 21-second video with sound. The weights are open on Hugging Face.

Where the cheapest model falls behind

The post's benchmark tables show why Cloudflare keeps three models. Clef-flash tops the BFCL tool-calling test at 98.76 against 98.47 for Clef, 98.2 for Clef-omni and 95.75 for Jev, and it scores 97.73 on the home-appliances test where Clef reaches 82.95.

Intent classification is where it loses ground. On CLINC150 with out-of-scope examples Clef-flash scores a macro-F1 of 66.77 while Clef gets 97.43 and Clef-omni 97.7, and on BANKING77 it manages 90.93 against 94.20 for Clef. Jev still leads on the When2Call test, at 80.97 against 65.58 for Clef-flash. These are Cloudflare's own figures.

Clef gets faster without new weights

Cloudflare moved serving to SGLang, and the hosted Clef speeds up without any retraining. Median latency for inputs of about 800 tokens falls from 262 to 152 milliseconds, from 616 to 305 at about 3,400 tokens and from 2,721 to 1,635 at about 16,000. The 95th-percentile figures at the same sizes fall from 438 to 351, 777 to 531 and 3,250 to 1,805 milliseconds.

Cloudflare says it submitted changes to SGLang, due to ship in version 0.5.22, and that launch commands for self-hosting are on the updated Hugging Face collection.

How Cloudflare uses Clef itself

The post lists four internal uses: the public GitHub docs repository uses Clef to detect and close spam issues, EmDash, Cloudflare's content management system, uses it to moderate plugin libraries for phishing, the data loss prevention team scans for personal data such as government IDs, and the threat intelligence team uses it to help detect malicious domains.

For teams already calling Clef-flash, the practical check is input size. Images and audio are converted to input tokens for pricing and limits, so a request that looks short can still pass 24,000 tokens.

Our AI model category lists Clef, Clef-flash and the new Clef-omni among its models.

The post says Cloudflare's developer documentation carries the most up-to-date prices.

Sources

• Cloudflare Blog: Introducing Clef-omni with full multimodality, plus a faster Clef and a cheaper Clef-flash, 9 October 2026
• Cloudflare Developers: Workers AI pricing (read 2026-10-11), 11 October 2026