Cloudflare cuts Clef-flash’s price and hosted input limit

Hosted Clef-flash accepts shorter requests after the price cut, but its downloadable weights are unchanged. Clef-omni adds audio and video input.

Hosted Clef-flash accepts shorter requests after the price cut, but its downloadable weights are unchanged. Clef-omni adds audio and video input.
Cloudflare announced a cut to Clef-flash’s hosted input price on 9 October, but reduced its context window—the amount of input it can process in one request—from 64,000 to 24,000 tokens. Tokens are the units used to measure model input. The company also released Clef-omni, which accepts audio and video alongside text and images.
The Register reported on 1 October that Cloudflare had released Clef and the smaller, faster Clef-flash. These decision models answer questions with predefined options, such as yes or no, a multiple-choice selection or a ranking. As The Decoder explained on 2 October, they return probabilities for those options rather than long text responses, giving developers structured results that software can act on.
Earlier Clef models supported text, images and arrays of video frames, while TypeSafe’s Jev decision model was text-only. Cloudflare says Clef-omni accepts audio files in WAV or MP3 format and video files in MP4 or WebM format, as well as text and images. It says developers can make decisions across those inputs in one API call, without separately transcribing speech or splitting a video’s audio and images.
Cloudflare says it also released Clef-omni’s weights—the learned settings that make the model work—on Hugging Face, allowing developers to run it on their own infrastructure.
Cloudflare describes the smaller hosted Clef-flash context window as a trade-off required for the lower price. It says only 0.24% of requests in its usage data exceeded 24,000 input tokens, and cites that finding as the reason for the change. For longer hosted requests, it recommends Clef, which retains a 64,000-token context window and its existing price.
| Model | Price per million input tokens | Price status |
|---|---|---|
| Clef-flash | $0.038 | Reduced from $0.09 |
| Clef | $0.24 | Unchanged |
| Clef-omni | $0.15 | Launch price |
| Jev | $0.042 | Reported on 1 October |
Clef’s unchanged rate is nearly six times Jev’s price reported on 1 October. Jev allows up to 64,000 tokens across a request, but limits the supplied state plus the longest individual question to 32,000. Clef’s API—the software interface used to send requests—is fully compatible with Jev’s. The Register reported that developers could replace Jev with Clef locally or through Cloudflare’s hosting without rebuilding that interface.
The hosted Clef-flash limit does not change its downloadable weights. Cloudflare says those weights are untouched and were trained to support a 256,000-token context window for self-hosting. Running the models locally avoids Cloudflare’s hosted token charge, but requires substantial graphics processor, or GPU, memory.
Michelle Chen, identified in The Register’s 1 October report as Cloudflare’s AI Platform group product manager, gave the following minimum memory requirements. They assume one request running at a time and a 64,000-token context window, not the larger window Cloudflare now describes for self-hosted Clef-flash.
| Model | GPU memory |
|---|---|
| Clef-flash | 41 GB |
| Clef | 85 GB |
The Register reported that Clef and Clef-flash weights were downloadable from Hugging Face under Apache-2.0 licence terms. Downloadable weights do not make the training data public: although Cloudflare called the release “open source,” Chen confirmed that the training datasets were not publicly available.
Cloudflare says changes to the infrastructure serving Clef on Workers AI, its model-hosting service, reduced response times. It released no new Clef weights because the changes were not materially to the weights or model architecture. The following median times—the midpoint of measured response times—are Cloudflare’s figures, not independently verified results.
| Approximate input size | Before | After |
|---|---|---|
| 800 tokens | 262 milliseconds | 152 milliseconds |
| 3,400 tokens | 616 milliseconds | 305 milliseconds |
| 16,000 tokens | 2,721 milliseconds | 1,635 milliseconds |
Cloudflare also says its GPU network helps reduce network delays. An earlier company example, reported by The Decoder on 2 October, compared its threat intelligence team’s use of Clef to classify a website with its fastest general-purpose language model. The timings covered fetching, rendering and classifying that site; they were one company-reported example, not an independent performance test.
| Model | Time for the full process |
|---|---|
| Clef | 2.2 seconds |
| Cloudflare’s fastest general-purpose language model | 4.7 seconds |
For website classification, answer options can include whether a site is an online store or a phishing site. Cloudflare’s training aims to make assigned probabilities match how often answers are correct, but a high probability does not guarantee a correct classification.
The Register’s 1 October report described Cloudflare’s tests using the Jev Decision Index, a benchmark available on Hugging Face. Cloudflare reported higher accuracy but slightly slower responses for Clef than the other open decision models tested, while Clef-flash matched most of their accuracy and ran much faster. At the time of that report, Cloudflare’s scores had not been reproduced for inclusion in the official Decision Index ranking.
The Decoder reported on 2 October that Cloudflare was rolling out a service to fine-tune Clef—train it further for customers’ tasks—using reinforcement learning, a training method based on feedback. Engineers working directly with customers would initially handle the training, with a self-service platform planned for later.
Under that planned workflow, customers would build datasets from requests logged through Cloudflare’s AI Gateway, then evaluate the requests in containers, isolated computing environments used as a training sandbox. A new training component would let them deploy customized models on Workers AI.
For self-hosting, Cloudflare says commands for running Clef with SGLang, software used to serve models, are already available in its Hugging Face repository and SGLang’s cookbooks. Its 9 October announcement says changes supporting Clef are due in SGLang version 0.5.22.

Transluce reports heavy traffic and attack-bearing requests against US and Canadian government websites. The findings do not establish a successful breach, and US officials found no service impact.
3 min

Its 2026 Digital Defense Report describes faster vulnerability discovery and malware development, while inadequate testing prevents rapid deployment of fixes and humans still direct most campaigns.
2 min

Its tools connect SI assistants to services such as iMessage and WhatsApp. Open-source software still accounts for 98% of use, TechCrunch reports.