The model Cloudflare built to answer questions without writing a single sentence can now watch video and listen to audio. On October 9 the company released Clef-omni, extending its family of open decision models to handle audio in WAV and MP3 formats and video in MP4 and WebM, alongside the text and images its two earlier models already accepted.
Clef is a decision model, a close relative of TypeSafe’s Jev. It does not generate prose. It reads an input state and a set of typed questions, then returns a probability for every allowed answer. Cloudflare trained the first two models in the family, Clef and Clef-flash, and has now added a third that can process sound and motion in a single request.
The new model runs on a 30-billion-parameter mixture-of-experts backbone with 3 billion active parameters. Cloudflare prices it at $0.15 per million input tokens. Text requests return in roughly 20 milliseconds, the company said, fast enough for the routing, classification and scoring jobs the model is meant to handle.
The launch came with a price cut on the small model. Clef-flash now costs $0.038 per million input tokens, down from $0.09, which Cloudflare said makes it cheaper than Jev. To reach that number, the company shrank the hosted model’s context window to 24,000 tokens from 64,000.
The weights for Clef-omni are open, published on Hugging Face under an Apache 2.0 license, continuing the practice Cloudflare set with the first two Clef models. The models run on the company’s Workers AI platform and are compatible with the Jev API, so a developer can swap model identifiers without rewriting a call.
The product sits in a field that has crowded quickly in a matter of weeks. TypeSafe, a San Francisco startup, raised an $870 million round led by Andreessen Horowitz the same week, valuing it at $7.5 billion on the strength of Jev, which returns typed probability judgments rather than text. Microsoft shipped Decision-1, its own scoring model, priced at $0.042 per million input tokens with output free of charge. Decision models have gone from an experiment to a contested category in under a month.
The appeal is cost and speed. A model that returns a probability is cheaper to run than one that composes a paragraph, and it gives a machine a way to make the thousands of small calls that a large language model handles with expensive and often overlong replies: which document to route where, which answer to score higher, which alert to raise.
Cloudflare’s argument for the multimodal version is that many of those small calls begin with sound or pictures. A support system that has to classify a complaint embedded in a customer’s voicemail, or a review pipeline that has to flag a video against policy, now has a model that reads the media directly instead of running it through a separate transcription or captioning step first. The company frames Clef-omni as one decision model for every input type, taking all of them in a single request.
The pricing of the family reflects a careful tiering. Clef, the larger model, holds at $0.24 per million input tokens with a 64,000-token window. Clef-flash is the budget tier. Clef-omni, at $0.15, sits in between, priced for workloads that need richer input than text but less than the full model.
Cloudflare’s entry is distinguished by the distribution it already owns. The company runs one of the largest networks on the internet, and its AI Gateway and Workers AI platform let a customer drop a model into an existing edge pipeline without standing up new infrastructure. That is a different posture from a startup that must first find users.
The bet underneath the product is that a large share of AI inference will be spent deciding, not composing. If that proves right, the winners will be whoever can serve those decisions cheapest and fastest, and the price war that has already cut Clef-flash to $0.038 a million tokens suggests the field is preparing to fight on those terms.
The open-weight choice is as much a strategy as the model itself. Cloudflare has made a habit of releasing its models under permissive licenses and serving them at the edge, betting that developers will pick the model that is easiest to run rather than the one with the loudest benchmark. Clef-omni extends that bet to media, and the price cut on Clef-flash extends it to the bottom of the market, where most inference volume actually lives.
Clef-omni’s place in that race is still unproven. Multimodal decision-making is young, and the model will have to show it can hold accuracy on video and audio while keeping the low latency that made the text version attractive. For now, Cloudflare has widened the family and lowered the floor, and given developers one more reason to treat the decision model as a standard tool rather than a novelty.
