News reporting
Cloudflare Clef-omni brings audio and video to agent decisions
The new model scores permitted choices from mixed media. Its documented context limits and uneven benchmark results matter before those choices trigger actions.
By Agent Report Editorial
Editorial review · · 3 min read
Source-based reportingWritten and source-checked with Codex from Cloudflare’s October 9 announcement and model documentation, checked October 11, 2026. Vendor benchmarks were not independently replicated; examples below are editorial analysis. Grok supplied the separately attributed AI commentary on October 11.Updated
Cloudflare released Clef-omni on October 9, adding audio and video to its decision-model family. The model scores a supplied set of answers from mixed-media evidence. For an agent builder, it offers a narrower job than generating an entire response: choose a branch in a workflow.
That distinction becomes useful when a support request arrives as a message, a photograph and a short recording. An application might ask which team should receive it, whether more evidence is needed and whether it can safely wait for normal business hours. These are examples of how we would evaluate the model, not deployments we have tested.
The interface gives builders something specific to check
The Workers AI model documentation describes a state plus typed questions, with a probability returned for each allowed option. The hosted context window is 64,000 tokens, and a request can contain up to four audio clips and two videos.
One operational detail deserves attention: media and questions consume context first. If they exceed the window, the request fails; otherwise, the textual state is truncated to fit the space left over.
Our assessment is that this creates a predictable test case. A long clip could leave less room for the customer's written explanation. The application should check what it sends and whether essential facts survive preparation, rather than assuming a valid response means the full case was considered.
More input types do not mean better results on every task
Cloudflare's launch table gives Clef-omni an invoice-processing exact-action score of 60.2, compared with 64.7 for Clef. Its customer-service exact-action score is 71.6, against 77.0 for Clef-flash. These are vendor-reported benchmark results, not our measurements.
The editorial conclusion is narrower than a model ranking: if the task is already text-only, the new modality support is not sufficient reason to switch. A team needs to show that its extra input changes the decision in a useful way.
For a customer-service trial, we would compare the same cases with and without the recording. Review incorrect routing, requests that should have gone to a person, and cases where adding media worsened an otherwise correct answer. Count the cost of preparing those inputs too.
Decide what happens when the evidence is poor
A permitted answer can still be wrong. An unclear photograph, unrelated attachment or incomplete recording can fit the request format while failing to answer the question.
Keep a route to human review and test it with deliberately incomplete cases. Choose the action threshold from the consequences of an error: suggesting a queue and authorizing a customer refund deserve different treatment. Returned probabilities should be evaluated against actual outcomes before becoming permission to act.
For each decision, retain enough of the input reference, allowed choices, model version, scores and final action to explain a mistake. Apply the same access and retention rules that protect the underlying customer material.
Clef-omni makes the decision step easier to isolate. The buyer's job is to establish whether that step improves the completed workflow. Our guide to measuring agent ROI keeps that comparison tied to useful work and review costs.