Cloudflare Clairic: API Price, Extraction and Preview Limits
Cloudflare's Clairic preview extracts typed fields at $0.10 per million input tokens. See its API, uncertainty signals, limitations and safe workflow fit.
Cloudflare introduced Clairic on October 19, 2026, as a preview extraction model on Workers AI. Instead of asking a general-purpose language model to generate a JSON answer, developers supply an input state and named typed fields, and Clairic returns values such as invoice dates, amounts, lists or estimates. The useful question is not whether it can replace every LLM. It is whether a repetitive business workflow needs reliable field shapes, clear missing-value handling and a human review path.
What launched, and what is available
The hosted model identifier is @cf/cloudflare/clairic. Cloudflare documents use through a Workers AI binding or its REST API; AI Gateway can sit in front of those endpoints. The model is explicitly in preview. Cloudflare says open weights are planned, but has not committed to a release date in its October 19 announcement. Do not design a local deployment around an assumed launch timeline.
Clairic at launch
Where an extraction model fits
Choose the task before the model
| Need | Better starting point |
|---|---|
| Read an order ID, amount or date from unstructured input | Clairic extraction field with typed output |
| Select approval, review or rejection from fixed options | Decision model such as Cloudflare Clef |
| Explain a discrepancy or write a customer response | Generative language model |
| Run repeatable multi-system operations with retries and approvals | Workflow automation around the model outputs |
A useful design is to separate interpretation from execution. Clairic can turn a support email into a typed order ID, date and amount. A workflow then validates those fields against the source system, routes uncertain values to a person and triggers only approved CRM or finance actions. This is a more testable architecture than giving a free-form model permission to both infer data and execute irreversible changes in one step.
Price is simple; total workflow cost is not
Cloudflare lists $0.10 per million input tokens and no output-token charge. That is a model-unit price, not a guaranteed cost per processed document: actual input size, retries, review time, surrounding API usage and the rest of the workflow still matter. Cloudflare describes speed and accuracy advantages over traditional LLMs, but those are vendor claims. Teams should compare accuracy and end-to-end cost on their own representative documents before migrating production traffic.
Preview limitations that change the rollout decision
Typed output guarantees a valid field shape, not a correct answer. Cloudflare says short arithmetic chains work better than complex calculations across many table rows; its forecasts are basic and do not anticipate unseen events such as holidays. A missing scalar value can return null, while confidence and p_null help identify uncertain cases. Estimates and forecasts include intervals, but those intervals are not promises of accuracy on unfamiliar data. High-impact decisions still require independent validation or review.
- Choose one narrow document type, such as support emails or supplier invoices, and define only the fields your downstream system actually needs.
- Use a labeled sample containing missing fields, ambiguous dates, image inputs and arithmetic edge cases to measure field-level errors.
- Check null, confidence and source spans where provided; send uncertain or consequential cases to a human review queue.
- Keep calculations that require many dependent steps, database lookups and irreversible updates in deterministic code or approved workflows.
- Track per-document accuracy, review rate and total processing cost before expanding beyond the preview.
The practical decision
For teams already automating sales operations, customer support or document intake, Clairic is worth a controlled pilot when the output is a small set of structured fields. It is not a replacement for a general writing assistant, an approval policy or a full forecasting system. Start with low-risk extraction and explicit verification; keep the rest of the business process observable and reversible.
Questions people ask
Can I call Cloudflare Clairic from an existing app without hosting model weights?
Yes. As of October 19, 2026, Clairic is hosted in preview on Cloudflare Workers AI under @cf/cloudflare/clairic. Cloudflare documents both a Workers binding and a REST API, with AI Gateway available for those endpoints. Open weights are planned, but the announcement gives no release date. Check the current model documentation before treating preview access as a long-term production commitment.
What should an automation do when Clairic cannot find an invoice field?
Treat the missing result as an unresolved input, not a zero or an invented value. Cloudflare documents null for scalar fields when the input does not determine the answer, together with confidence and p_null metadata. A safe workflow can request the missing information or send the record to review. For payment or customer-account changes, validate against the system of record before taking action.
Can Clairic calculate complicated invoice totals or forecast seasonal demand reliably?
Not without additional checks. Cloudflare says one- and two-step arithmetic is more reliable than long chains or large-table aggregation; it recommends doing complex math in code after extracting inputs. Its preview forecasting follows recent patterns but does not anticipate unseen events such as holidays. For seasonal demand or consequential financial calculations, use dedicated validation or forecasting methods and test on your own data.
When should I use Clairic instead of Clef or a generative LLM?
Use Clairic when the output is a typed value such as an amount, date, duration or estimate. Cloudflare describes Clef as a decision model for choosing among predefined options, while generative models produce open-ended content. For example, extract a customer order number with Clairic, route its case to a predefined category with a decision model, and draft an explanatory reply with an LLM if needed.
Clairic offers a low-priced, typed extraction layer for structured business workflows, but preview limitations and model uncertainty still matter. Pilot it on real inputs, validate important fields and keep execution controls outside the model.
Explore related Make Better resources
- AI Workflow Automation Templates— For repeatable downstream workflows
- What Should You Automate First?— Selecting a low-risk automation pilot
Sources & useful resources
- Cloudflare: Introducing Clairic (Oct 19)— Official model announcement and examples
- Cloudflare Workers AI: Clairic changelog— Preview, API access, typed fields, price and explicit limitations
- Cloudflare: Extraction models guide— Official field behavior, null, confidence, use-case comparison and safeguards
- Cloudflare: Clairic model reference— Official context window, input images, identifier and model pricing