Cloudflare today released a preview of Clairic, a model that reads text and images and returns typed values - numbers, dates, durations, lists and estimates - instead of generating sentences. Hosted on Workers AI under the name @cf/cloudflare/clairic, it is the first of what the company calls "System One extraction models," and Cloudflare says it intends to publish the model with open weights. In a test run by Cloudflare, a three-step customer-claim pipeline built around Clairic cost 49% less and finished about 3.2 times faster than the same pipeline run entirely on a large language model.

In Short

Cloudflare made an AI model that does not write answers; it fills in a form with values it finds or works out from an email, a receipt or a table of numbers. That matters for anyone who uses chatbots to pull totals, dates or order numbers out of documents, because chatbots can garble a digit, invent a value or return output that breaks the software reading it. With Clairic, every answer comes back in the format requested, a missing value comes back as empty instead of guessed, and each answer carries a score saying how sure the model is.

What Cloudflare published

The announcement appeared on the Cloudflare Blog on October 19, 2026, under the title "Introducing Clairic, a new System One class of extraction models." Three Cloudflare authors are credited: Alex Reneau, Michelle Chen and Kevin Flansburg. The post is listed as a 16-minute read and carries the tags AI, Developers, Machine Learning, Research and Workers AI.

Its opening claim is broad. Clairic, according to Cloudflare, "is able to extract, compute, estimate, and forecast values from unstructured inputs - all at a significantly higher accuracy, faster latency, and cheaper cost than traditional LLMs." The post does not support "higher accuracy" with a general benchmark. Its evidence consists of worked examples, a 21-day forecasting test and a single pipeline comparison, all run by the company.

The release is a preview. "The current endpoint is an early checkpoint and we expect it to get better," the authors wrote. No date is given for the open-weights release, which the post describes only as coming "soon" and, later, "after some more work." No per-token price for Clairic appears anywhere in the post.

Where the "System One" label comes from

The vocabulary is borrowed. Cloudflare refers to Daniel Kahneman's book Thinking, Fast and Slow, which separates fast, intuitive judgment (System 1) from slow, deliberate reasoning (System 2). Under that framing, a large language model that reasons step by step is a System Two tool.

According to Cloudflare, the "System One model" idea was introduced about a month earlier by a company called TypeSafe, alongside a model named Jev, which the post describes as "purpose-fit for a specific use case, and priced accessibly given its niche domain." The post says the concept "took the industry by storm." PPC Land has not independently verified TypeSafe's announcement or that characterization; both come from Cloudflare's account.

Until today, every System One model was a decision model, according to the post. Jev and Clef - the latter hosted on Workers AI as @cf/cloudflare/clef - take an input and a list of possible answers and return a probability for each option. Clairic is different. It returns values that are "directly extracted from the input, computed from it, or predicted beyond it."

Cloudflare splits most model tasks into three output types: "a number, a choice from a fixed set, or open-ended text." Clairic covers the first and also pulls typed values such as names, dates and lists. Clef covers the second. LLMs, in Cloudflare's framing, are left with the third.

How a request works

The mechanics are what separate Clairic from a chatbot. A request consists of a "state" - the text to read plus up to four optional images - and a map of named fields. Each field gets a type and, optionally, plain-language instructions. The model answers every field in what Cloudflare calls "one parallel pass."

"There is no reasoning and no output tokens," the post states. Because the model only processes input tokens in a single forward pass, Cloudflare says there is no charge for output. That is a direct contrast with generative models, which bill for every token they write, including the reasoning tokens many current models generate before answering. According to the post, System One models "typically respond in under a second, while an LLM that reasons through the same step takes several seconds."

Nine field types are listed:

  • text returns a span of the input with its character offsets
  • number returns a value read or computed, honoring integer and decimal rules
  • date, time and datetime return calendar values in ISO form
  • duration returns a length of time in a chosen unit
  • list returns matching text, numbers or dates in input order
  • estimate returns a distribution inside declared minimum and maximum bounds, on a chosen step grid, on a linear or log scale
  • forecast returns one estimate per future step of a series

Every scalar answer includes a value called p_null, the probability that the value is not in the input, plus a confidence score. Estimates come with a median and an 80% interval running from the 10th to the 90th percentile, and all 99 percentiles are available on request. There is no cap on the number of fields. If too many fields are requested to fit beside the input in one pass, the server splits them into parallel passes that each see the whole input, then merges the results.

Because the model generates no tokens, Cloudflare argues it "can't produce a mistyped answer or 'hallucinate' in the same way as an LLM (though it can, of course, make mistakes)." Outputs are either a value of the requested type or null. The distinction matters. Hallucination in generative models has been traced to training that rewards guessing; research from OpenAI and Georgia Tech published in September 2025 argued that benchmarks penalize models for admitting uncertainty, and proposed confidence thresholds as a fix. Clairic's p_null and confidence fields are an architectural version of that idea, exposing the uncertainty instead of hiding it.

The customer-email example

Cloudflare's first worked example is a short customer complaint. The customer writes that order A-88412 "arrived on Sept 28, 2026 with a cracked screen and a dented corner," that they paid "$649.00 plus $52.10 tax," and that "It's now Oct 9, 2026."

Six fields were requested: order ID, delivery date, total paid including tax, days since delivery, a list of damage types and a tracking number. The results, according to Cloudflare:

  • The order ID came back as "A-88412", with character offsets 10 to 17 and a confidence of 0.999957
  • The delivery date came back as 2026-09-28
  • The amount paid came back as 701.1, a figure that appears nowhere in the email; Clairic added the price and the tax in the same pass
  • Days since delivery came back as 11, computed from two dates in the text
  • The damage list returned "cracked screen" and "dented corner" as exact spans
  • The tracking number came back null, with p_null of 1

That last field shows the design at work. The model's internal guess for the tracking number was "$52.10" - the tax amount - with a confidence of 0.416. Instead of returning that wrong value as an answer, the system returned null and kept the guess visible in a separate field. A generative model asked for JSON containing a tracking number may produce a plausible-looking string, and the error would only surface downstream.

Cloudflare frames the alternative bluntly. Extraction is "probably the most common job handed to LLMs today," the post says, and its failure modes are "hard to catch": invalid JSON, a digit misread during generation, and no way to trace an excerpt back to the source. "Most people build extraction pipelines to combat this, with multiple layers of parsing, schema validation, retries and spot checks," according to Cloudflare. Because every field is answered in the same pass, "asking for thirty fields costs about the same as asking for three."

Receipts and images

A second example used what Cloudflare calls "a real-life photo of a receipt," where "the text is not perfect and the angle of the image doesn't help its legibility." Clairic returned an order number (77245, confidence 0.776), a purchase datetime (July 1, 2026 at 13:34, confidence 0.831), a quantity of 2, a unit price of $7.99, a subtotal of $15.98, sales tax of zero and a tip of $2.39. The discount field came back null, with p_null of 0.9994.

The confidence spread is worth noting. The two lowest-confidence answers - order number and timestamp - are exactly the fields most likely to be smudged on a photographed receipt, and those scores fell below 0.85 while the price fields stayed above 0.999. That is the behavior a confidence score is meant to have, though Cloudflare published only this single image result. The image capability puts Clairic in the territory of OCR, the technology that turns scanned text into machine-readable characters, while adding typed extraction on top.

Estimates and forecasts

Estimating values

For values not stated in the input but supported by evidence, Clairic returns a distribution. Cloudflare's example fed the model five comparable home sales and the details of a subject property: 1,620 square feet, three bedrooms, built in 2001, in a neighborhood called Elm Park. The estimate field was bounded between $50,000 and $5,000,000 on a log scale with a $1,000 step.

The answer: a value of $418,000, a median of $415,000, and an 80% interval of $370,000 to $470,000.

The post argues that LLM estimates come back as "a single number with no sense of how sure the model is," so downstream software "can't tell a confident $180 from an unconfident $180." A decision model could approximate an estimate by choosing among price buckets, Cloudflare notes, but buckets "are coarse, every bucket adds input tokens, and each new category adds more sources of confusion."

One detail in this section does not line up. The text says "price is narrow ($370,000-$470,000), days on market is wide (24-148)." The code example requests only a sale price field, and the output shown contains no days-on-market answer. The 24-to-148-day range appears in the post without the request or output that produced it.

The post also points at a use with obvious relevance to marketers. "Because fields are unlimited and answered in parallel, you can score many items in one call (one estimate per candidate, lead, or listing) and sort them in code," according to Cloudflare.

Forecasting TSA passenger counts

The forecasting test used public data: the daily number of people screened at Transportation Security Administration checkpoints at U.S. airports. Each morning, Cloudflare gave Clairic the real counts so far, starting from July 16, 2026, and asked for the next day's figure. It did this for 21 consecutive days, with data accessed on October 15, 2026.

For October 14, Clairic forecast a median of 2,419,283 travelers, with an 80% interval of 2,270,774 to 2,556,769. TSA actually screened 2,419,474 people that day. On PPC Land's calculation, that miss was 191 people, or less than 0.01%.

A single day is not the test, though. Across all 21 days, Cloudflare says the forecasts were off by 3.3% on average, and 17 of the 21 actual counts landed inside the 80% interval - about 81%, close to what a calibrated 80% interval would be expected to produce. The model picked up the weekly rhythm from the numbers alone: "quieter Tuesdays and Saturdays, busier Thursdays, Fridays, and Sundays."

The misses cluster. "Three of the four days outside the interval fall on the Columbus Day weekend, when more people flew than the previous weeks suggested," according to the post. Clairic "forecasts from the patterns in the series but doesn't account for events the series doesn't show, like holidays." For anyone forecasting advertising traffic, where Black Friday, Prime Day and other events drive the biggest swings, that limitation is significant.

Cloudflare is careful here. Clairic "is not claiming to replace a dedicated time-series model, but it is a little faster and cheaper than your regular LLM," the post says. The headline claim of "significantly" faster and cheaper performance is not repeated for forecasting.

The pipeline benchmark

The most detailed numbers come from a three-stage pipeline built on the damage-claim email. In Cloudflare's version, Clairic extracts the values and estimates a repair cost; Clef chooses between a full refund and a free repair based on a policy (refunds within 30 days, repair preferred when it costs less than 60% of the price); and GLM-5.3, a generative model, writes a reply under 60 words. The comparison pipeline used GLM-5.3 for all three steps, "from the same email, policy and prompts."

Cloudflare's reported results:

Clairic, Clef, GLM-5.3 GLM-5.3 for every step
End-to-end latency 19.3 s 61.7 s
Extract Clairic: 0.46 s GLM-5.3: 16.1 s
Decide Clef: 0.22 s GLM-5.3: 21.9 s
Reply GLM-5.3: 18.6 s GLM-5.3: 16.4 s
Input tokens 701 ($184 per million claims) 390 ($546 per million claims)
Output tokens 1,329 ($5,848 per million claims) 2,544 ($11,194 per million claims)
Cost per million claims $6,032 $11,740

The cascade was 69% faster, or about 3.2 times quicker, and cost 49% less, according to Cloudflare. Both percentages check out against the table. The extraction and decision steps together took 0.68 seconds, against 38 seconds for GLM-5.3 on the same two steps. Almost all of the cascade's remaining time - 18.6 of 19.3 seconds - was the LLM writing the reply.

Several details in the table deserve attention. The three step times in the all-LLM column add to 60.4 seconds, not the 61.7 seconds shown as the total; the cascade column adds up exactly. The post does not explain the 1.3-second gap. The cascade also used more input tokens (701) than the all-LLM version (390), yet its input cost was lower. On PPC Land's arithmetic, that implies about $0.26 per million input tokens for the cascade against about $1.40 for GLM-5.3, but the post does not break out what Clairic or Clef charge per token, so the blended figure cannot be separated. Output tokens drive most of the cost in both columns, which is where a model that writes nothing has its advantage.

Consistency was the other finding. "Both pipelines got the facts and the decision right in every run," Cloudflare says, "but Clairic's repair estimate was identical each time, while GLM-5.3's changed on every run ($250, $250 and $220, with different ranges each time)." That wording implies three runs. The post does not say how many runs were performed in total.

Cloudflare also tested two other models. "Under the all-LLM pipeline both Gemma 4 and Qwen failed at this task, as they were unable to deduce values like repair_cost from the inputs," the post states. No version numbers, run counts or failure rates are given for either model, and the conclusion the post draws - "showing how LLMs are not great at data extraction tasks" - rests on that one task.

Prompt injection and auditability

The pipeline design has a security argument attached. Clef "decides from a five-line summary of typed values instead of a free-form email, which can also be susceptible to prompt injection," according to Cloudflare. The final LLM "never sees the email, only the decision and the predictions it needs, so there are fewer numbers for it to misuse." In the all-LLM version, "every value in the reply has to be checked by hand."

That separation means raw customer text never reaches the component that decides what to do. A malicious instruction hidden in an email would have to survive conversion into a typed number, date or text span before influencing anything downstream.

What Cloudflare says does not work yet

The post closes with a candid list of limits. Each is attributed to Cloudflare:

  • Arithmetic: "Reading is strong, but multi-step arithmetic isn't yet." One- and two-step calculations - totals, differences, conversions, date arithmetic - "seem reliable," but longer chains and aggregation across many table rows are hard "for any one-pass model without a scratchpad." The authors recommend having Clairic predict the value and letting code do the math.
  • Evidence: estimates work best from comparables, history or descriptive text. Asking for a value without evidence "may be limited in specificity."
  • Calibration: "Calibration is measured, not guaranteed." Intervals are trained to be calibrated and are on most held-out sets, but "they can run narrow on unfamiliar data."
  • Forecasting: "Forecasting is basic." The model is intended for "non-critical, general pattern matching, not detailed long-horizon forecasting."
  • Correctness: "The architecture makes every answer well formed, not necessarily correct."

The last line undercuts the opening claim of "significantly higher accuracy" in a useful way. A typed output removes a class of formatting errors. It does not remove wrong values.

Why this matters for marketing operations

Advertising and marketing teams are among the heaviest users of LLMs for structured work: pulling figures from invoices and insertion orders, reading performance reports, classifying leads, reconciling receipts. The trouble with that work has been documented. A study by NP Digital covered in February found 47.1% of marketers encounter AI inaccuracies several times a week, and 36.5% said incorrect AI-generated content had already been published. More than 70% reported spending hours each week fact-checking AI output.

The industry has also been wary of letting generative models touch money. In January, executives told PPC Land that the ad industry won't hand LLMs "the keys" to spend budgets, citing the mismatch between probabilistic LLM outputs and deterministic auction logic. Clairic does not resolve that tension, but it changes its shape. A system that returns a typed number with a confidence score and a null for missing data is easier to gate with rules than one that returns prose.

Cost is the other pressure. OpenRouter data from August showed agents using roughly five times the tokens per task that human users do, while token prices had started rising after years of decline. Every call that moves from a generative model to an input-only model removes output tokens from the bill entirely. The inference economics are the core of Cloudflare's pitch, and the benchmark shows output tokens making up 97% of the cascade's cost even after extraction and decisions were moved off the LLM.

Clairic also extends a run of Cloudflare AI products that PPC Land has tracked through 2026. In June, Cloudflare added dollar-denominated spend limits to AI Gateway. In early October it opened a web search API for AI agents with providers priced from $0.25 to $7 per 1,000 requests, and set AI Search billing to start on November 1 at $0.75 per million ingestion tokens. Most of PPC Land's Cloudflare coverage, though, concerns its role as gatekeeper for AI crawlers - most recently its plan, described on October 16, under which Verified bots face trust scores for ignoring site rules. Clairic sits on the other side of that business: a model for developers building on Cloudflare, not a control for publishers.

The open-weights promise is also notable. Berlin search engine Ecosia dropped Mistral for hosted open-weight models earlier this month, citing cost. A small, task-specific extraction model with published weights could run outside Cloudflare's infrastructure entirely, which would reduce its value as a Workers AI lock-in tool but widen its reach. Until the weights appear, Clairic is available only on Cloudflare's hosted endpoint.

What remains unanswered

Several questions are left open by the post. There is no published price per token for Clairic. There is no accuracy benchmark beyond worked examples. The pipeline comparison used one claim type, one competing model in the head-to-head, and an unstated number of runs. TypeSafe and Jev, the inspiration for the category, are described only in Cloudflare's words. And the open-weights release has no date.

The model's name, according to the post, combines "clerical," "Clair" (from "Clair de Lune" and clairvoyance) and "[Cl]oudflare's [AI R]egression model." Cloudflare invites developers to call it from a Worker with env.AI.run() or from anywhere through its REST API.

Timeline

Summary

Who: Cloudflare, through a blog post by Alex Reneau, Michelle Chen and Kevin Flansburg. The model is aimed at developers building on Workers AI.

What: A preview of Clairic, a "System One extraction model" that returns typed values - text spans, numbers, dates, durations, lists, estimates and forecasts - with confidence scores and null answers, in a single forward pass with no output tokens. In Cloudflare's own test, a claims pipeline using Clairic, Clef and GLM-5.3 cost 49% less and ran about 3.2 times faster than GLM-5.3 alone.

When: Today, October 19, 2026. Open weights are promised without a date.

Where: On Cloudflare's Workers AI platform as @cf/cloudflare/clairic, callable from a Worker or through the REST API.

Why: Cloudflare argues that many structured-output calls sent to frontier LLMs ask the same kinds of questions - totals, dates, delays, costs - and that a model built only to return values is faster, cheaper and easier to audit. The company acknowledges that multi-step arithmetic, calibration on unfamiliar data and forecasting remain weak, and that well-formed answers are not necessarily correct ones.