AnswersPublished26 sources
Cloudflare Clairic: How Its System One Extraction Model Works
Cloudflare’s Clairic is a preview extraction model that returns typed values—not generated prose—for $0.10 per million input tokens, with no output token charge.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Cloudflare’s Clairic is a preview extraction model that returns typed values—not generated prose—for $0.10 per million input tokens, with no output token charge.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is Cloudflare’s Clairic, previewed on October 19, 2026, and how does this non-text-generating “System One” extraction model work—includ. Article summary: Clairic is Cloudflare’s preview “System One” extraction model, announced on October 19, 2026: it returns typed values rather than generating prose, and can extract, compute, estimate, and forecast from unstructured input. Topic tags: general, documentation, general web. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fak
Cloudflare’s Clairic is a preview model for turning unstructured text and images into structured values such as dates, numbers and lists. Rather than write a response, it returns fields in the types an application requests. Cloudflare reports faster and cheaper results in its demonstrations, but those examples are not enough to establish that Clairic is generally more accurate or efficient than large language models (LLMs).3
4
14
What Clairic does
Clairic is a “System One” extraction model available in preview on Workers AI as @cf/cloudflare/clairic. Cloudflare says it can extract, compute, estimate and forecast values from unstructured inputs. It charges $0.10 per million input tokens; because it does not generate output tokens, there is no output-token charge.3
7
18
A request combines a state—text and optionally up to four images—with named fields that specify the type of answer needed. Clairic returns the requested fields in one pass instead of producing a sequence of generated words.16
20
The supported field types include text spans, numbers, dates, times, datetimes, durations, lists, estimates and forecasts. Text spans can include character offsets, while a forecast can return values for up to 64 future steps.1
20
How uncertainty appears in the output
Scalar answers include a confidence score and p_null, the probability that the input does not determine the answer. The response can also include a value and a guess; when the value is missing, it can be returned as null rather than presented as a certain fact.2
9
18
Estimate and forecast fields provide a median and an 80% interval, from p10 to p90. These are uncertainty estimates from the model, not guarantees that the actual value will fall inside the interval. Cloudflare recommends using the confidence, p_null and interval information to decide what to automate and what to send for review.2
What Cloudflare’s performance examples show
In a damaged-product support example, Clairic extracts claim details, Cloudflare’s Clef model selects an option such as refund or repair, and GLM-5.3 generates the customer-facing response. Reporting on the test gives Clairic’s extraction step as 0.46 seconds and Clef’s decision step as 0.22 seconds. The complete mixed-model pipeline took 19.3 seconds, compared with 61.7 seconds for a version using GLM-5.3 at every step.4
15
The reported costs were $6,032 per million claims for the mixed pipeline and $11,740 for the GLM-5.3-only version—a 49% reduction in that example.15 These figures describe one particular workflow, not expected savings for every application.
Cloudflare also tested next-day forecasts of U.S. airport checkpoint passenger counts over 21 days. The reported average error was 3.3%, and 17 of the 21 actual counts fell within the model’s 80% interval.6
9 That is a small demonstration on one series, not evidence of broad forecasting performance across different data or conditions.
Practical limits to keep in mind
Cloudflare describes Clairic’s forecasting as basic: near-term forecasts follow recent patterns, while forecasts further out can flatten toward the average and may miss events not reflected in the input data.1
17
The documentation also draws a boundary around arithmetic. One- and two-step calculations—such as totals, differences, conversions and date arithmetic—are described as reliable. Long chains of dependent calculations and totals across many table rows are not yet reliable; Cloudflare recommends extracting the needed values with Clairic and doing the math in code.1
17
So far, the performance case rests on task-specific examples, not a broad comparative benchmark demonstrating that Clairic is more accurate, faster or cheaper for all workloads. Treat those advantages as possibilities to test against your own task, rather than universal guarantees.4
14
15
The “System One” label and Jev
Cloudflare uses “System One” to frame Clairic as a fast, purpose-built alternative to models that spend time generating a reasoned textual response.5 PPC Land reported that TypeSafe had introduced the System One idea earlier alongside a model named Jev. That reporting establishes a reported antecedent, but the available sources do not establish a direct technical lineage or substantiate a specific Kahneman-inspired explanation for Clairic’s name.
13
Cloudflare has also said it plans to release open weights, but the sources provided do not give a release date.7
12
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Cloudflare Clairic: How Its System One Extraction Model Works"?
Cloudflare’s Clairic is a preview extraction model that returns typed values—not generated prose—for $0.10 per million input tokens, with no output token charge.