![Cloudflare’s Clef-omni: 30B Model Costs $0.15/M [2026]](https://shattered.io/wp-content/uploads/2026/10/cloudflare-clef-omni-30b-decision-model-2026-gen.jpg)
Cloudflare rolled out a new member of its open-weight model lineup on October 9, 2026: Clef-omni, a multimodal decision model that reads text, JSON, images, audio, and video in one pass. The release, published through Cloudflare’s Workers AI changelog, pushes the company’s bet on cheap, structured AI inference further into territory that used to require a full chatbot-grade model. Clef-omni does not write essays or hold a conversation. It answers typed questions with probabilities, and Cloudflare says that narrow focus is exactly the point.
The announcement, confirmed by Cloudflare’s own changelog, places Clef-omni inside the Workers AI catalog under the model ID @cf/cloudflare/clef-omni. It joins two existing siblings, Clef and Clef-flash, to round out a three-model family built specifically for what Cloudflare calls decision points in automated workflows: the moments where an AI agent has to classify, route, or choose rather than generate paragraphs of prose.
What a “Decision Model” Actually Does #
Most AI headlines in 2026 are still about bigger chat models, so Clef-omni’s design is worth slowing down on. A decision model does not generate free-form text at all. Feed it a schema listing the allowed answers to a question, and it returns a probability for each option. There is no streaming output to parse, no risk of the model wandering off-topic, and no chance it invents an answer outside the list you gave it.
That constraint sounds limiting until you picture where it fits. An agent moderating uploaded video might need one answer: does this clip violate a content policy, yes or no, with a confidence score attached. A customer support pipeline might need to route a ticket into one of eight categories. A fraud-detection system might need to flag a transaction as suspicious or clean. In each case, a full chat model is overkill, slower than necessary, and prone to producing output that needs cleanup before the next system stage can use it. Clef-omni skips the cleanup step entirely because the output format is fixed from the start.
This is a different philosophy than Cloudflare’s own Workers API tooling, which focuses on wiring general-purpose endpoints into production. Clef-omni instead narrows the job the model is asked to do, which is also why Cloudflare can run it cheaply at the edge instead of routing every call to a heavyweight data-center cluster.
Inside Clef-omni’s 30B-A3B Architecture #
Clef-omni is built as a mixture-of-experts model with 30 billion total parameters, of which only 3 billion are active for any given input. Cloudflare’s own documentation describes the design as 30B-A3B, shorthand that is becoming familiar across the open-weight model world in 2026. The idea behind mixture-of-experts is straightforward: instead of running every parameter for every request, the model routes each input to a small subset of specialized sub-networks, so inference cost tracks closer to a 3B model than a 30B one.
Why the Foundation Model Choice Matters
Cloudflare did not train Clef-omni’s multimodal backbone from scratch. The model is built on top of Qwen3-Omni-30B-A3B-Instruct, the foundation model line published by Alibaba’s Qwen team. Cloudflare’s own Hugging Face model card for Clef-omni, which is the same documentation referenced in the Workers AI changelog, names Qwen3-Omni as the base. Leaning on an existing multimodal foundation model lets Cloudflare focus its own engineering on the decision-output layer, rather than rebuilding audio and video understanding from zero.
That is a pattern worth tracking across the broader decision-model trend this year. Microsoft’s Decision-1, a separate 9B model aimed at similar structured-choice tasks, is also reported to build on a Qwen-family base. Two major cloud platforms converging on the same open foundation model for their decision-model products, within months of each other, says something about how far ahead the Qwen line has pulled for multimodal groundwork.
Audio, Video, Image, and Text in a Single Pipeline #
What separates Clef-omni from its two smaller siblings is the input menu. Clef and Clef-flash handle text and structured data. Clef-omni adds images, audio, and video, and Cloudflare says all of it moves through a single forward pass rather than a chain of separate transcription and vision models bolted together.
- **Audio input:** WAV and MP3 files
- **Video input:** MP4 and WebM files
- Image input: standard image formats passed alongside text and JSON context
- Text and JSON: the same structured schema input used by Clef and Clef-flash
Running every modality through one pipeline instead of a stitched-together stack of separate models cuts out a step that has quietly become a tax on a lot of production AI systems: converting an audio or video model’s output into a format a second model can read, then feeding that into a third decision layer. Clef-omni collapses that into one call. For an agent that needs to decide whether an uploaded video clip breaks a rule, that is one API round trip instead of three.
Clef-omni’s Price Tag: $0.15 per Million Input Tokens #
Cloudflare priced Clef-omni at $0.15 per million input tokens, a figure confirmed in the company’s own Workers AI pricing documentation tied to the October 9 release. That is the same announcement window in which Cloudflare also cut the price of Clef-flash, the smallest model in the family. Ground News, which aggregated coverage of the launch, put Clef-flash’s new price at $0.038, down from its prior rate, though the exact billing unit was not spelled out in the portion of that coverage circulated so far.
Cloudflare’s changelog title for the release, “Clef-omni adds audio and video input, Clef-flash is now cheaper, and Clef is faster,” bundles three separate updates into one announcement: a brand-new multimodal model, a price cut on the cheapest existing model, and a speed improvement on the mid-tier model. That is an unusually busy single changelog entry, and it signals Cloudflare is iterating on the Clef family on a tight cadence rather than treating it as a one-off side project.
The Clef Family, Side by Side #
Here is what Cloudflare has confirmed about each model in the family as of the October 9 announcement. Several specifics about Clef and Clef-flash, including their exact parameter counts, were not detailed in the material Cloudflare has published so far, so those cells are marked accordingly rather than estimated.
| Model | Role in the Family | Modalities | Price (input) | What Changed Oct. 9 |
|---|---|---|---|---|
| Clef-omni | Multimodal decision model | Text, JSON, images, audio (WAV/MP3), video (MP4/WebM) | $0.15 per million tokens | New model, first launch |
| Clef-flash | Lightweight decision model | Text and JSON | $0.038 (unit not fully detailed) | Price cut from prior rate |
| Clef | Standard decision model | Text and JSON | Not detailed in current release | Inference speed increased |
The pattern across the row is consistent with how Cloudflare has positioned Workers AI generally: cheap by default, with the most expensive tier reserved for the model doing the most work. Clef-omni carries the multimodal premium, Clef-flash undercuts everything for simple text-only decisions, and the standard Clef sits in between.
Output Format: Probabilities, Not Prose #
The defining trait of every model in the Clef line is what comes back from a call. Clef-omni returns a probability for each allowed option in a schema the developer supplies, for each typed question asked. It does not produce free-form text, and Cloudflare’s documentation is explicit that this removes the need for output parsing on the application side.
That design choice has a direct cost benefit tied to it. A chat model asked to make a yes-or-no decision still generates a full sequence of tokens, hedges, and often a paragraph explaining its reasoning before landing on an answer buried in the text. A decision model skips straight to the number. For an agentic system running thousands of decisions per minute, the difference between parsing a sentence and reading a probability is a difference in both latency and the amount of code needed to keep that pipeline from breaking when the model phrases its answer differently than expected.
Integration With Jev, SystemOne, and AI Gateway #
Cloudflare’s documentation lists Clef-omni as compatible with Jev and SystemOne, and states that the model also works with AI Gateway, Cloudflare’s layer for managing, logging, and controlling calls across multiple AI providers. Jev, the decision/agent-model product from TypeSafe AI, has drawn its own attention this year after reaching a $7.5 billion valuation in a matter of weeks, so compatibility between Clef-omni and Jev’s API shape is a meaningful interoperability signal rather than a footnote.
Routing Clef-omni calls through AI Gateway also means teams already using Cloudflare’s gateway to manage costs and logging across OpenAI, Anthropic, or other providers can slot a decision model into the same monitoring pipeline without standing up separate infrastructure. That is consistent with Cloudflare’s broader strategy of making Workers AI the default landing point for teams that are already inside the Cloudflare ecosystem for DNS, CDN, or edge compute.
Apache-2.0 Licensing and What It Buys Developers #
Clef-omni ships under an Apache-2.0 license, the same permissive terms Cloudflare has used for the rest of the Clef family. That license lets companies self-host the weights, fine-tune the model, or redistribute it commercially without the royalty or usage restrictions attached to some competing proprietary decision-model offerings.
For a company the size of Cloudflare, giving away the weights under a permissive license while also charging for hosted inference through Workers AI is not a contradiction. It is the same model AWS, Google, and Meta have used with their own open releases: open weights build developer goodwill and ecosystem lock-in, while the hosted, managed version is where the recurring revenue sits. Clef-omni’s Apache-2.0 license puts it on the same competitive footing as other open decision models racing to be the default choice for agent builders who don’t want to depend on a closed API they can’t inspect or self-host if pricing changes.
Competitive Landscape: Decision Models Are Having a Moment #
Clef-omni is not landing in an empty field. 2026 has turned into the year smaller, cheaper decision and routing models started shipping in parallel across several companies, all chasing the same problem: agentic systems need fast, structured answers far more often than they need another paragraph of generated text.
| Model | Company | Size | Category | Notable Detail |
|---|---|---|---|---|
| Clef-omni | Cloudflare | 30B total, 3B active | Multimodal decision model | $0.15 per million input tokens, built on Qwen3-Omni |
| Decision-1 | Microsoft | 9B | Decision model | Claimed 35x speed improvement over comparable models |
| Drex 1.5 | Nace.AI | 9B | Decision model | Scored 58.08 on its benchmark suite at launch |
| Large 4 (“Le Chonk”) | Mistral | 1T total, 52B active | General-purpose open-weight model | Positioned as a broader foundation model, not decision-specific |
| Jev | TypeSafe AI | Not publicly detailed | Decision/agent model | Reached $7.5 billion valuation within weeks of launch |
Open-Weight Decision Models vs. General Foundation Models
The table above splits into two camps. Clef-omni, Decision-1, and Drex 1.5 are purpose-built for narrow structured decisions, all clustered in the 9B-to-30B range where inference stays cheap. Mistral’s Large 4, at 1 trillion total parameters with 52 billion active, is a different kind of product entirely, a broad foundation model meant to compete on general capability rather than decision-making speed. The fact that three separate companies, Cloudflare, Microsoft, and Nace.AI, have each shipped a dedicated decision model within the same few months suggests the market has converged on the idea that agent pipelines need a cheap, specialized layer that sits apart from the flagship chat model entirely.
Historical Context: Cloudflare’s Multi-Year Push Into Edge AI #
Workers AI did not start as a decision-model platform. Cloudflare built it as a general hosted-inference product, letting developers call a range of open models directly from Workers without managing GPU infrastructure themselves. The Clef family represents a narrower, more recent branch of that strategy, one aimed specifically at the agent economy that has grown up around tool-calling and multi-step automated workflows over the past two years.
That trajectory mirrors Cloudflare’s pattern elsewhere in its platform. The company has repeatedly taken a general infrastructure capability, like compute at the edge or object storage, and later layered a specialized, cheaper product on top once a specific use case proved out. The Clef family, and Clef-omni in particular, reads as the AI-inference version of that same playbook: start broad with Workers AI, then carve out a narrow, high-volume slice of the market, decision inference for agents, and price it aggressively enough that switching away becomes a hard sell.
Market Impact: What Clef-omni Means for the Edge AI Race #
The timing of this release matters as much as the model itself. Agentic AI systems, the kind that chain together multiple model calls to complete a task without constant human input, have become the dominant framing for AI product announcements across the industry in 2026. Every one of those agents needs to make small, fast decisions constantly: should this step run, does this output pass a check, which branch of logic applies next. Clef-omni is a direct bet that a meaningful share of those decisions will involve multimodal input, not just text, as agents start handling video moderation, audio transcription review, and image classification as routine sub-tasks rather than edge cases.
For Cloudflare specifically, Clef-omni strengthens the case for keeping an entire agent pipeline inside Workers AI and AI Gateway rather than splitting inference calls across multiple vendors. A company already paying Cloudflare for edge compute and routing has one more reason not to stand up a separate multimodal classification service elsewhere. That is a meaningful competitive lever against AWS, Google Cloud, and Azure, all of which offer their own hosted inference but have not, as of this announcement, shipped a directly comparable dedicated multimodal decision model at a comparable price point.
What the Hugging Face Model Card Confirms #
Cloudflare published Clef-omni’s documentation directly on Hugging Face, where the model card lists the same core details as the Workers AI changelog: a 30B-A3B mixture-of-experts architecture built on Qwen3-Omni-30B-A3B-Instruct, released under Apache-2.0, and designed to return per-option probabilities rather than generated text. Publishing the full model card on Hugging Face, in addition to the Workers AI listing, gives developers who want to self-host or inspect the weights a path to do so outside Cloudflare’s own infrastructure, consistent with the Apache-2.0 terms.
That dual presence, hosted on Workers AI for convenience and published on Hugging Face for portability, is becoming the standard release pattern for open-weight models from infrastructure companies. It lets Cloudflare court both camps at once: developers who just want an API endpoint, and developers who want to run the weights themselves and only use Cloudflare’s gateway for logging and governance.
Predictions: Where Clef and Decision Models Go Next #
A few trends look likely to play out over the next several months, based on the pace of this release and the broader pattern across the industry this year.
- More modalities will keep arriving in small increments. Clef started as text-only, Clef-flash and Clef kept that scope, and Clef-omni is the first to add audio and video. Expect Cloudflare to keep expanding modality support across the family rather than launching an entirely new model line each time.
- Price competition between decision models will intensify. With Microsoft’s Decision-1 and Nace.AI’s Drex 1.5 both targeting the same narrow use case, and Cloudflare already cutting Clef-flash’s price in the same announcement as Clef-omni’s launch, a race to the bottom on cost-per-decision looks likely through the rest of 2026.
- AI Gateway compatibility will become a selling point for every new Cloudflare model. Clef-omni’s explicit interoperability with Jev and SystemOne suggests Cloudflare is prioritizing ecosystem compatibility over exclusivity, a strategy likely to extend to future releases.
- Enterprises will start treating decision models as a distinct line item from chat models in their AI budgets. As the cost difference between a probability-only call and a full generative call becomes clearer, expect procurement teams to start separately tracking spend on decision inference versus conversational inference.
- Open-weight licensing will keep being the default for this category. Every decision model covered in this piece, from Clef-omni to Decision-1 to Drex 1.5, has shipped with permissive or openly published weights, suggesting the category has settled on openness as a competitive norm rather than an exception.
What Remains Unconfirmed #
Not every detail about this release is public yet. Cloudflare has not published benchmark scores for Clef-omni against competing decision models, nor has it detailed Clef’s exact new price following the speed improvement mentioned in the same changelog entry. Context window length, latency figures, and hardware requirements for self-hosting the Hugging Face weights were also not specified in the material published as of this writing. Readers tracking the Clef family closely should expect Cloudflare to fill in some of those gaps in follow-up documentation rather than treat the October 9 announcement as the final word on the model’s full specifications.
Frequently Asked Questions #
What is Cloudflare Clef-omni?
Clef-omni is a multimodal decision model from Cloudflare, released October 9, 2026, that accepts text, JSON, images, audio, and video, and returns a probability for each allowed answer in a schema rather than generating free-form text.
How much does Clef-omni cost?
Cloudflare prices Clef-omni at $0.15 per million input tokens, available through Workers AI under the model ID @cf/cloudflare/clef-omni.
What foundation model is Clef-omni built on?
Clef-omni is built on Qwen3-Omni-30B-A3B-Instruct, the multimodal foundation model from Alibaba’s Qwen team, adapted by Cloudflare into a 30B-parameter mixture-of-experts decision model with 3B active parameters.
What file formats does Clef-omni support?
Clef-omni accepts WAV and MP3 audio files, and MP4 and WebM video files, alongside standard images and text/JSON input.
Is Clef-omni open source?
Yes. Clef-omni is released under the Apache-2.0 license, and its model card is published on Hugging Face in addition to being hosted on Cloudflare Workers AI.
How is Clef-omni different from Clef and Clef-flash?
Clef and Clef-flash handle text and JSON only. Clef-omni is the first model in the family to add image, audio, and video input, processing all modalities in a single pipeline.
Does Clef-omni work with other Cloudflare AI products?
Cloudflare’s documentation states Clef-omni is compatible with Jev and SystemOne, and that it works with AI Gateway for logging and governance across AI providers.
Can Clef-omni replace a chatbot model?
No. Clef-omni is not designed for open-ended conversation. It is built specifically for structured decision points, such as classification or routing tasks, where the answer must come from a fixed set of options rather than generated prose.