Cloudflare Clef-omni Brings Audio, Video, Images and Text Into One AI Model
Cloudflare is expanding its Clef family of open-weight decision models with a multimodal addition, faster inference for Clef, and lower pricing for Clef-flash. The October 9, 2026 announcement focuses on making AI-powered decisions more flexible, responsive and affordable for developers.
AI models are becoming useful for more than reading text and classifying written information. Developers increasingly need systems that can evaluate photographs, recordings, video clips and written instructions together. Combining these inputs can be difficult when every type of media requires a separate processing step.
That is the problem Cloudflare is addressing with Clef-omni, its latest addition to the Clef family. Announced on October 9, 2026, the model accepts audio, video, images and text within a single decision-making workflow. Cloudflare also announced a lower input-token price for Clef-flash and serving optimizations that make Clef faster.
These updates target developers building classification systems, automated workflows and AI agents that need to interpret information and select appropriate actions rather than simply generate conversational responses.
One decision model can evaluate multiple kinds of input, while improvements to speed and pricing make the Clef family more practical for different workloads.
What Is Cloudflare Clef-omni?
Clef-omni is an open-weight decision model designed to evaluate different types of input and produce structured decisions. Unlike a workflow that first converts audio into text and then sends that text to another model, Clef-omni can process the supplied modalities together through a unified pipeline.
The model accepts text and images alongside audio files in WAV or MP3 format and video files in MP4 or WebM format. For video, it can consider the soundtrack alongside the visual frames, allowing evidence from what is seen and heard to contribute to the same decision.
Cloudflare built Clef-omni on the Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation, retaining its primary comprehension capabilities while removing the text-to-speech output components.
Why a Unified Multimodal Pipeline Matters
Traditional multimodal workflows may rely on several connected models. One model transcribes speech, another processes visual information, and a decision model combines the results to determine what should happen next.
Clef-omni is designed to reduce this chain of processing. It evaluates the supplied media and the available answer options together, using a structured decision process instead of generating a long natural-language response.
Several processing stages
Transcribe audio, separate video information, combine the outputs and then make a decision.
One unified decision process
Evaluate text, images, audio and video together within a single model request.
The model uses a mixture-of-experts architecture and a two-stage attention-routing approach to gather evidence for possible answers and calculate confidence scores. A built-in lexical grammar helps preserve the meaning of the available options, supporting decisions that follow a specified output structure.
Because Clef models are designed for decision-making rather than conventional long-form text generation, Clef-omni avoids the additional output-generation step used by many language-model workflows. This design is intended to support fast, schema-constrained decisions across different input types.
What Could Developers Do With Clef-omni?
Cloudflare's announcement demonstrates a workflow for reviewing an installation using a photograph, an audio recording and a video. Instead of processing each file through an independent chain of models, a developer can submit the information together and ask targeted questions.
Inspect an image
Determine whether a device's model and serial-number label is visible.
Evaluate a recording
Assess whether an operating unit sounds normal or contains unusual rattling or grinding.
Review a video
Check whether a fan is running while considering the video's visual and audio information.
The example illustrates how a decision model can evaluate several pieces of evidence against specific questions. Similar architectures can be relevant to media inspection, automated classification and workflows that require a defined answer format.
The actual reliability of any implementation will depend on the task, the quality of its inputs and how the developer validates the model's decisions.
Cloudflare Makes Clef Faster
Alongside Clef-omni, Cloudflare improved the serving infrastructure for its existing Clef model on Workers AI. The company reports that the updated deployment returns decisions up to twice as fast in its tested input-size scenarios.
The improvement comes primarily from infrastructure optimizations, including a move to SGLang for model serving. Cloudflare says it did not release new Clef model weights as part of this speed update because the changes were concentrated on how the model is served.
Source: Cloudflare's October 9, 2026 announcement. Values show the previous and updated median latency, respectively, for the specified input sizes.
The figures show that the improvement varies with input size. In Cloudflare's reported tests, the approximately 3,400-token input experienced the largest median speedup, reaching 2.0×. These measurements describe the tested serving configuration rather than a guarantee that every request will be twice as fast.
Clef-flash Gets a Major Price Reduction
Cloudflare has also reduced the input-token price of Clef-flash, making the model a more affordable option for developers running classification and decision workflows at scale.
The new price is $0.038 per million input tokens, down from $0.09. Clef remains at $0.24 per million input tokens, while Clef-omni launched at $0.15 per million input tokens, according to the announcement.
| Model | Input price | Hosted context |
|---|---|---|
| Clef-flash | $0.038/M | 24K tokens |
| Clef | $0.24/M | 64K tokens |
| Clef-omni | $0.15/M | 64K tokens |
M = one million input tokens. Prices are those reported in the October 9 announcement; check Cloudflare's current documentation before deploying a production workflow.
The lower hosted price for Clef-flash comes with a smaller context window: 24K tokens instead of the previously advertised 64K. Cloudflare says the self-hosted weights remain unchanged and were trained to support a 256K context window. Developers who need a larger hosted context can consider Clef instead.
Choosing a model should therefore involve more than comparing the price per million tokens. Input length, required modalities, latency and the shape of the desired output all affect which option is appropriate.
How Cloudflare Uses the Clef Model Family
Cloudflare describes Clef as a way to bring classification and detection capabilities directly into application workflows. The company highlights several internal use cases that demonstrate how decision models can support operational tasks.
Spam issue detection
Cloudflare says its public GitHub documentation repository uses Clef to detect and close spam issues.
Phishing-related plugin moderation
The EmDash content management system uses Clef to moderate plugin libraries for phishing.
Potential personal information detection
Cloudflare's data loss prevention team scans for potentially sensitive information, including government identification numbers.
Malicious domain detection
The company's threat intelligence team uses Clef to help identify malicious domains.
These examples concern the broader Clef model family and should not be interpreted as evidence that every listed workflow specifically uses Clef-omni. They illustrate the type of classification and decision tasks Cloudflare wants developers to build with its models.
How Developers Can Try Clef-omni
Clef-omni is available through Cloudflare Workers AI. The model also has open weights published on Hugging Face, giving developers a route to explore self-hosting options.
Cloudflare states that Clef-omni follows the same System One API as Clef and Clef-flash and works with AI Gateway. Developers can start by reviewing the model documentation and adapting an API request to include the relevant media inputs and structured questions.
Before integrating a model into a live application, developers should check current pricing, input-format requirements, context limits and output behavior in the official documentation. Testing against representative examples is also essential for understanding how the model performs on a specific task.
What This Cloudflare AI Update Means
- Clef-omni expands input support. It accepts text, images, audio and video in one decision-making workflow.
- Clef gets faster. Cloudflare reports median speedups of 1.7× to 2.0× across the tested input sizes.
- Clef-flash becomes cheaper. Its hosted input price falls from $0.09 to $0.038 per million input tokens.
- Lower cost has a trade-off. The hosted Clef-flash context window is reduced to 24K tokens.
- Decision models target specific tasks. Classification, detection and structured action selection are central to the Clef family.
- Model selection still matters. Developers need to balance input type, context requirements, price and performance.
Cloudflare Is Making AI Decisions More Accessible to Developers
The Clef update is about three connected improvements: broader input capabilities, faster inference and lower entry-level pricing. Clef-omni brings audio and video into a model family that already supported text and images, while the existing Clef models receive their own performance and cost improvements.
For developers, the practical question is which model best matches the task. Clef-omni is relevant when a decision depends on several types of media. Clef-flash offers a lower-cost option when its smaller hosted context window is sufficient, while Clef retains a larger hosted context window.
These releases do not eliminate the need for evaluation, validation or sensible model selection. They do, however, give developers more options for building decision-focused AI applications without automatically assembling a separate model pipeline for every input type.
Source: Cloudflare Blog — October 9, 2026. Pricing, performance measurements and product details reflect the announcement and may change.