Cloudflare previews Clairic, a typed extraction model on Workers AI

The jobs in the post are the ones that show up in document pipelines: an invoice total, a delivery date, how many days late a shipment is, what a repair will cost, what a series does next. Cloudflare says a decision model such as its own Clef, or TypeSafe's Jev, needs the answers listed in advance, and that many of these values cannot be listed that way. Clairic is the model it built for that gap[1][2].

Cloudflare says the design follows TypeSafe's System One models, which it places about a month before this post, and it uses Daniel Kahneman's fast-and-slow framing the way TypeSafe did. In Cloudflare's telling, Clairic and Clef answer in one forward pass and do not generate tokens. A general language model is the slow step: it writes text, and you pay for those output tokens[1].

The first paragraph also says Clairic does this "at a significantly higher accuracy, faster latency, and cheaper cost than traditional LLMs"[1]. The post does not attach a general benchmark to that sentence. The measurements it does publish are one internal pipeline comparison and a 21-day forecast on a public series, both run by Cloudflare.

A request is a state and a map of typed fields. The state is the text to read. You can also send up to four images, embedded as PNG, JPEG or WebP, at most 4 MiB each and 8 MiB decoded in total. Remote image URLs are not accepted. Text answers are spans of the state, with character offsets[3].

The model page lists a 65,536-token context window, vision support, and unit pricing of $0.10 per million input tokens[3]. The changelog says there is no output-token charge, because the model generates none[2]. You call it with env.AI.run() from a Worker, or with the Workers AI REST API. Cloudflare also says AI Gateway can sit in front of those endpoints[2].

The field types in the changelog are text, number, date, time, datetime, duration, list, estimate and forecast. A number honours integer and decimals. A list comes back in input order. An estimate is a distribution inside the min and max you set, on your step, on a linear or log scale, with a median and an 80% interval from p10 to p90. Set options.include_quantiles and you also get all 99 percentiles. A forecast is one of those estimates per future step, up to 64 steps[2].

Every scalar answer includes p_null, the probability that the input does not determine the value, and a confidence score. When the model does not find an answer it returns null rather than a filled-in string. The blog's support-email example shows a missing tracking number with p_null of 1 and value: null, while guess is a string that does appear in the email[1]. The model page's invoice example does the same for a purchase-order number that is not on the invoice[3].

Cloudflare says there is no cap on how many fields you can ask for. Fields that do not fit beside the input in one 8,192-token pass are split across parallel passes, each of which still sees the whole input[2][3]. The docs are plain that a typed result is predictable, not therefore correct. A constraint can force a date into November. The payment can still happen in December[4].

These are Cloudflare's runs. The post does not link an independent evaluation, and this story did not call the API.

On a damaged-product email, Cloudflare's cascade is Clairic for the values, Clef for the decision, and GLM-5.3 (@cf/zai-org/glm-5.3) for a short reply. The comparison is the same three steps done entirely by GLM-5.3, with GLM-5.3 also writing the reply in the cascade so the last step matches[1].

Cloudflare's table puts the cascade at 19.3 seconds end to end and the all-GLM pipeline at 61.7 seconds. The step times are Clairic 0.46 seconds to extract, Clef 0.22 seconds to decide, and GLM-5.3 18.6 seconds to reply, against 16.1, 21.9 and 16.4 seconds for GLM-5.3 on those three steps. Input tokens are 701 for the cascade, which Cloudflare costs at $184 per million claims, against 390 tokens and $546 for the all-GLM run. Output tokens are 1,329 ($5,848 per million claims) against 2,544 ($11,194). The totals are $6,032 and $11,740 per million claims. Those two token-cost lines add up to the totals. Cloudflare calls that 69% less time, about 3.2 times faster, and 49% cheaper, which matches the headline times and the two totals[1].

The cascade's three step times also add to 19.3 seconds. The three GLM step times add to 54.4 seconds, 7.3 seconds short of the 61.7 second end-to-end number. The post does not say what that gap is.

Cloudflare says both pipelines got the facts and the decision right in every run, that Clairic's repair estimate stayed the same, and that GLM-5.3's estimate moved ($250, $250 and $220, with a different range each time). It does not say how many runs that was, beyond those three GLM estimates. It also says Gemma 4 and Qwen, used as the all-LLM pipeline, could not produce a repair cost. It does not say how they were prompted or how many times they were tried[1].

The forecast example is the daily number of people screened at US airport checkpoints, which Cloudflare attributes to Transportation Security Administration checkpoint travel numbers accessed on 15 October 2026. Each morning it asked for the next day, for 21 days. For Wednesday 14 October it reports a median of 2,419,283 against an actual of 2,419,474. Across the 21 days it reports an average error of 3.3%, with 17 of the 21 actual counts inside the 80% interval. It says three of the four misses fell on the Columbus Day weekend, and that the model follows the pattern in the series and does not account for events the series does not show, such as holidays[1]. This story did not re-download that TSA series or rerun the forecast.

A separate illustration estimates a sale price from five comparable sales: a point value of $418,000, a median of $415,000, and an 80% interval from $370,000 to $470,000. The prose says a days-on-market estimate in the same example ran from 24 to 148 days. The JSON printed with the example contains only the price field[1]. It is a worked example, not a scored test.

The blog and the changelog describe the endpoint as an early checkpoint[1][2].

One- and two-step calculations, such as totals, differences, conversions and date arithmetic, are what Cloudflare calls reliable. Longer chains of dependent steps, and aggregations across many rows of a table, are not. Its advice is to have Clairic read the inputs and let code do that maths. Estimates work best when the input contains comparables, history or a description. Without that, the interval is wide. A next-step forecast can follow a recent pattern. Forecasts several steps out flatten toward the average, and the model does not stand in for a time-series model trained on your own data. Intervals are trained to be calibrated and, Cloudflare says, are calibrated on most of its held-out sets, but can run narrow on unfamiliar data. It does not publish those sets[1][2].

The architecture point Cloudflare keeps repeating is the one that matters for anyone wiring this into a payment or a refund: every answer is a valid value of the type you asked for, or null, and that does not make it the right value. Use confidence, p_null and the interval to decide what to automate and what to send to a person[2][4]. The model terms say Cloudflare offers it as is and as available, makes no warranty, and may stop providing it without a refund, credit or other compensation[3].

Most of the calls in a back office are not essays. They are "read this and return these fields." Cloudflare is hosting a model that will not emit free text for that job, bills only the input at $0.10 per million tokens, and attaches a null probability to each scalar answer. That is a concrete product difference from asking a general model for JSON, if the endpoint behaves the way the docs describe.

Whether it is more accurate is still Cloudflare's claim. The comparison is one internal demo, the forecast is 21 days of one public series, and the launch materials give no parameter count, no training description and no third-party evaluation. The open-weight release is a plan. Until those weights exist, the thing you can actually run is a preview endpoint Cloudflare can withdraw.