Cloudflare Clairic is an AI model that reads messy input like an email or a receipt photo and returns clean, typed values without ever writing a sentence. Cloudflare released it as a preview on Workers AI on 19 October 2026. It also says Clairic is the first in a new “System One” class of extraction models: fast, single-pass and cheap.
Key takeaways
- Clairic pulls typed values (text spans, numbers, dates, durations, lists, estimates and forecasts) from text plus up to four images.
- It answers in one pass and generates no output tokens. You pay only for input, at $0.10 per million input tokens.
- Every answer comes with a
confidencescore and ap_null, the probability that the value isn’t in the input at all. - In Cloudflare’s own demo, a Clairic-based pipeline ran about 3.2x faster and 49% cheaper than an all-LLM pipeline. Nobody has independently reproduced that yet.
- Open weights are promised “soon” but haven’t been released.
What is Cloudflare Clairic, exactly?
Think of it as a very fast clerk. You hand it a document and a form with named, typed fields, such as “order_id: text”, “delivery_date: date” and “total: number”. It fills in the form. If a field has no answer in the document, it returns null and doesn’t guess.
The name “System One” comes from Daniel Kahneman’s book Thinking, Fast and Slow. System One is quick, intuitive thinking. System Two is slow and deliberate. In Cloudflare’s launch post, chat LLMs like GPT, Claude and Gemini sit in the System Two group because they produce answers one token at a time and often reason first. Clairic skips all of that and returns every field in one forward pass.
Clairic has a sibling. Earlier in October 2026, Cloudflare released Clef, an open-weight System One decision model, and Clef-flash. Clef returns probabilities for a fixed set of answers such as “refund” or “escalate”. It is API-compatible with Typesafe AI’s Jev, the model that popularised the decision-model idea.
What can it pull out of a document?
The Workers AI changelog lists nine field types. They fall into four jobs:
- Extract:
textfields return an exact span of the input, with character offsets so your app can highlight the source. - Compute:
number,date,time,datetime,durationandlistfields can be read directly or worked out, such as adding price and tax or counting the days since delivery. - Estimate: an
estimatefield returns a median and an 80% interval (p10 to p90) within bounds you set. - Forecast: a
forecastfield returns one estimate per future step, up to 64 steps.
Here’s a concrete example from Cloudflare. Given a customer complaint email, Clairic pulled out the order ID, delivery date and list of damaged items. It also computed the $701.10 total and the 11 days since delivery. The email had no tracking number, so that field came back null. In another test it estimated a home’s sale price at a median of $415,000, with an 80% range of $370,000 to $470,000.
According to the model docs, the context window is 65,536 tokens. Images can be PNG, JPEG or WebP, up to 4 MiB each. If your schema is too large for one 8,192-token pass, the server splits it into parallel passes and merges the results.
How does Clairic compare with an all-LLM pipeline?
Cloudflare’s main demo is a damaged-product claim. One version chains three models: Clairic extracts the facts, Clef decides what to do, and the GLM-5.3 LLM writes the reply to the customer. The other version uses GLM-5.3 for all three steps. These are Cloudflare’s own figures:
| Step | Clairic + Clef + GLM-5.3 | GLM-5.3 only |
|---|---|---|
| Extraction | 0.46 s | 16.1 s |
| Decision | 0.22 s | 21.9 s |
| Written reply | 18.6 s | 16.4 s |
| End-to-end | 19.3 s | 61.7 s |
| Cost per million claims | $6,032 | $11,740 |
Most of the speed-up comes from the first two steps. The LLM is still the slow part, and it still writes the reply. There’s also a consistency gain. Across repeated runs, Clairic gave the same repair estimate every time, while GLM-5.3 answered $250, $250 and then $220.
The same pattern shows up in agent design. Our guide to production-grade AI agents argues for small, predictable components over one model doing everything.
Can it really forecast?
Somewhat. Cloudflare had Clairic forecast US TSA checkpoint passenger counts every day for 21 days. The forecasts averaged 3.3% error, and 17 of the 21 actual counts fell inside the 80% interval. On one day the median forecast was 2,419,283 and the real count was 2,419,474.
The misses clustered around the Columbus Day weekend, because the model doesn’t account for holidays. Cloudflare says plainly that Clairic is not a replacement for a time-series model trained on your own data, and a DEV Community write-up makes the same caveat.
Where does it fall short?
Cloudflare lists the limits itself, which is refreshing:
- Arithmetic: one- or two-step sums are reliable. Long chains and totals across many rows are not. Extract the inputs and do the maths in code.
- Estimates need evidence in the input. Without it, the intervals get wide.
- Calibration: intervals can be too narrow on data the model hasn’t seen before.
- Correctness: in Cloudflare’s words, every answer is “well formed, not necessarily correct.”
That last point matters most. A guaranteed schema stops your parser from crashing, but the value can still be wrong. My view: the confidence and p_null scores are the real feature here, because they let you send uncertain cases to a person instead of trusting every answer.
What happens next
Clairic is still in preview. You can call it as @cf/cloudflare/clairic from a Worker with env.AI.run() or through the REST API, optionally routed through AI Gateway. Cloudflare says open weights are coming “soon” but hasn’t given a date or a parameter count.
Watch for three things: independent benchmarks of the 3.2x and 49% claims, the open-weight release, and whether other labs start shipping their own System One extractors. Price pressure on small, task-specific models is already strong, as the $0.15 MiniMax M3.1 Flash shows.
FAQ
How much does Cloudflare Clairic cost?
Cloudflare’s changelog lists $0.10 per million input tokens, with no output token charge, because Clairic doesn’t generate tokens.
Is Clairic an LLM?
Not in the usual sense. It doesn’t write text. It returns typed values with confidence scores in a single pass, and a field it can’t answer comes back null.
Can I run Clairic on my own servers?
Not yet. As of 19 October 2026 it is hosted only on Workers AI. Cloudflare plans to release open weights but hasn’t set a date.
If you have a pile of invoices, support emails or receipts that someone types up by hand, try Clairic on 50 of them, then compare its confidence scores with the errors you find before you automate anything.
1 comment