Skip to content
Hire an AI employee
AI EmployeeSuper-AgentsAgent-to-AgentBenchmarksPricingBlog
Guides 08 Choosing a platform

Cloudflare Clairic: An Extraction Model, Not an LLM

Nitish Garg · Founder & CEO, CellCog · 20 October 2026 · 6 min read · 1,293 words
At a glanceQuick answers
What is Clairic?
A Cloudflare model that pulls values out of messy input, such as an invoice total, a delivery date or a list of damages, and returns them as typed fields with confidence scores instead of free text.
How is it different from an LLM?
It reads the input once and returns values directly, with no reasoning and no generated tokens, so you pay only for input and every answer is the type you asked for or null.
What does it cost?
$0.10 per million input tokens on Workers AI, with no charge for output, per Cloudflare’s pricing page on October 20, 2026.
Data illustration on a near-white ground: a crumpled receipt and an email feed into one flat machine that outputs neat typed tags for an order number, a date and a dollar total, under the figures 0.46 seconds and 10 cents per million input tokens
Fig 0One pass from messy input to typed values. Made by CellCog's image agent, running GPT Image 2.5.

Cloudflare has released Clairic, a model that pulls typed values out of emails, receipts and photos in a single pass instead of writing an answer like an LLM. “Today, we’re announcing a preview of a new class of System One-style models: extraction models.” The launch post went up at 17:48 UTC on October 19, 2026, by Alex Reneau, Michelle Chen and Kevin Flansburg, and Clairic is live on Workers AI as @cf/cloudflare/clairic.

It is the second System One model class after decision models such as TypeSafe’s Jev and Cloudflare’s own Clef. Where those pick among options you list, Clairic returns values: a total, a date, a list of damages, an estimate with a range. This post reads Cloudflare’s launch post, model page and pricing page directly.

On this page · 8 sectionsOpen
  1. What Clairic returns
  2. What it costs
  3. Cloudflare’s pipeline test
  4. What Cloudflare says it cannot do yet
  5. Where it fits
  6. Where CellCog stands
  7. What we are watching for
  8. Sources
Key points5 · 6 min full read
  1. A document turning into neat labeled tags.
    Cloudflare released Clairic in preview on Workers AI on October 19, 2026, as the first System One extraction model, and says open weights will follow.
  2. One arrow passing straight through a block.
    Clairic answers in one parallel pass with no output tokens: you declare typed fields and get each value back with a confidence score, or null when it is not in the input.
  3. A coin beside a stack of tokens.
    It costs $0.10 per million input tokens on Workers AI, reads text plus up to four images, and has a 65,536-token context window.
  4. A short bar beside a long bar.
    In Cloudflare’s claims test, Clairic plus its decision model Clef plus GLM-5.3 finished in 19.3 seconds against 61.7 for GLM-5.3 alone, at 49% lower cost.
  5. A warning triangle over a calculator.
    Cloudflare’s own limits: multi-step arithmetic is not reliable yet, estimates need evidence in the input, and forecasting is basic.

§ 01What Clairic returns

You send a state (text, plus up to four images) and a map of named, typed fields. Clairic reads the input once and fills every field in the same pass, so asking for thirty fields costs about what three do. There is no reasoning and no output text to parse.

Animation: a customer email with five highlighted fragments passes once through a Clairic block; five field cards fill in together with order_id A-88412, delivered_on 2026-09-28, amount_paid 701.10, days_since_delivery 11, and a grey tracking_number card reading null
Fig 1A messy email goes in once; typed values come out, and a missing value comes back null
Type What comes back
Extracted value A span copied from the input, with character offsets and a confidence score
Computed value A number, date, time, duration or list read from the input or worked out from it
Estimated value A distribution inside the bounds you declare, with a median and an 80% interval
Forecasted value One estimate per future step of a series, each with its own 80% interval
Missing value null, with the probability that the value is absent and the model’s best guess
Table 1What Clairic can return (Cloudflare launch post, October 19, 2026)

Cloudflare’s worked example is a support email. Clairic adds $649.00 and $52.10 tax into an amount paid of 701.1, counts 11 days between delivery and the email, lists both kinds of damage as spans, and returns null for a tracking number the email never mentions. In a house-price example, the 80% interval for price is narrow ($370,000 to $470,000) while days on market is wide (24 to 148), which shows how much the input actually supports each answer.

A customer email on the left with the order number, dates and prices highlighted; on the right five field cards: order_id A-88412, delivered_on 2026-09-28, amount_paid 701.10, days_since_delivery 11, and tracking_number null in grey

01A support email becomes five typed fields; the missing tracking number comes back null

Two bars: Clairic plus Clef plus GLM-5.3 at 19.3 seconds and 6,032 dollars per million claims; GLM-5.3 for every step at 61.7 seconds and 11,740 dollars per million claims

02The cascade finished the claims pipeline in 19.3 seconds against 61.7

1 / 2
Fig 2Clairic's worked example and Cloudflare's pipeline test

§ 02What it costs

Model Kind Price per million input tokens
@cf/cloudflare/clairic Extraction $0.100
@cf/cloudflare/clef-omni Decision $0.150
@cf/cloudflare/clef-flash Decision $0.038
Table 2Cloudflare’s System One models on Workers AI (pricing page, read October 20, 2026)

All three charge only for input tokens. The model page lists a 65,536-token context window and vision support for Clairic.

§ 03Cloudflare’s pipeline test

Cloudflare ran an insurance-claim workflow two ways: Clairic extracts, Clef decides, and GLM-5.3 writes the reply; or GLM-5.3 does all three steps from the same email, policy and prompts.

Animation: two lanes race on one time axis; the cascade lane of Clairic 0.46 seconds, Clef 0.22 seconds and GLM-5.3 18.6 seconds finishes at 19.3 seconds, while GLM-5.3 doing extract 16.1, decide 21.9 and reply 16.4 seconds, plus other time, finishes at 61.7 seconds
Fig 3Both pipelines on one clock: the cascade finishes in 19.3 seconds, GLM-5.3 alone in 61.7
Step Cascade GLM-5.3 for every step
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
End to end 19.3 s 61.7 s
Cost per million claims $6,032 $11,740
Table 3Claims pipeline: System One cascade vs GLM-5.3 for every step (Cloudflare launch post)
Seconds per step in Cloudflare's claims test, System One models vs GLM-5.3Bar chart of seconds per step: Clef decide 0.22, Clairic extract 0.46, GLM-5.3 extract 16.1, GLM-5.3 decide 21.9, with Clairic extract highlightedClef decide0.22Clairic extract0.46GLM-5.3 extract16.1GLM-5.3 decide21.9Seconds per step in Cloudflare's claims test, System One models vs GLM-5.3Bar chart of seconds per step: Clef decide 0.22, Clairic extract 0.46, GLM-5.3 extract 16.1, GLM-5.3 decide 21.9, with Clairic extract highlightedClef decide0.22Clairic extract0.46GLM-5.3 extract16.1GLM-5.3 decide21.9
Fig 4Seconds per step in Cloudflare's claims test, System One models vs GLM-5.3

The cascade took 69% less time and cost 49% less. GLM-5.3’s three steps add up to 54.4 seconds; the 61.7-second figure is the end-to-end time Cloudflare reported. Both pipelines got the facts and the decision right in every run, but Clairic’s repair estimate was identical each time while GLM-5.3’s moved between $250 and $220. Cloudflare also reports that Gemma 4 and Qwen failed the all-LLM version because they could not work out values like the repair cost.

§ 04What Cloudflare says it cannot do yet

The post is unusually frank about this checkpoint: “Reading is strong, but multi-step arithmetic isn’t yet.” One- and two-step sums and date math work; long chains of dependent steps and totals across many table rows do not. Its advice: “At least for now, have Clairic predict the value and let code do the math.”

Estimates need comparables or history in the input, intervals can run narrow on unfamiliar data, and forecasting is meant for general patterns, not detailed long-range work. The summary line is the one to keep: “The architecture makes every answer well formed, not necessarily correct.” Every answer carries a confidence score and a null probability so you can decide what to automate and what to check.

§ 05Where it fits

Clairic is for the repeated, structured steps that teams now send to a frontier LLM: invoice totals, delivery dates, claim amounts. The pattern Cloudflare describes is a split of labor: extraction models turn input into values, decision models turn values into choices, and an LLM turns choices into words. “We intend to release the model open-weights soon.”

§ 06Where CellCog stands

CellCog does not host Clairic; every CellCog agent tier runs on Claude Opus 5.5. We track System One models because they change what the small, repeated steps inside agent workflows cost.

§ 07What we are watching for

  • The open-weights release Cloudflare says is coming.
  • Independent tests of Clairic’s calibration on outside data.
  • Whether TypeSafe or Microsoft answer with extraction models of their own.

§ 08Sources

Frequently asked6 questions

Q1When was Clairic released?

Cloudflare published the launch post on October 19, 2026 at 17:48 UTC. The model is a preview on Workers AI as @cf/cloudflare/clairic.

Q2Is Clairic open source?

Not yet. Cloudflare says it intends to release the model with open weights after more work. Today it is a hosted endpoint only.

Q3What kinds of values can it return?

Text spans with character offsets, numbers, dates, times, durations, lists, estimates with an 80% interval, and forecasts for a series, each typed and bounded by the schema you send.

Q4What is a System One model?

A fast, single-pass model with typed outputs, named after Kahneman’s fast thinking. TypeSafe’s Jev introduced the idea for decisions; Clairic applies it to extracting values.

Q5Where does Clairic struggle?

Cloudflare says long chains of dependent arithmetic, aggregating many table rows, estimates without evidence and detailed forecasting are weak spots in this checkpoint.

Q6Does CellCog use Clairic?

No. Every CellCog agent tier runs on Claude Opus 5.5. We follow System One models because they change what it costs to run the small, repeated steps inside agent workflows.

Published 20 October 2026 All Choosing a platform →