Cloudflare Clairic Preview: The Model That Extracts Data Without Generating a Single Token

On October 19, 2026, Cloudflare released something that doesn't quite fit the word "AI model" the way we've been using it. Clairic — now in preview on the company's Workers AI platform — is the first "System One extraction model": a model that never generates text at all. You feed it a document, declare the fields you want back as typed values, and it returns them in a single forward pass. No tokens stream out. No prose to parse. And because nothing is generated, there is no output token charge: $0.10 per million input tokens is the whole bill.

Quick answer

• What: Clairic (@cf/cloudflare/clairic) is Cloudflare's preview extraction model on Workers AI. Instead of writing text, it returns typed values — text spans, numbers, dates, lists, estimates, forecasts — pulled from your input.
• Cost: $0.10 per million input tokens; zero output charge, because it generates zero tokens. Preview pricing, so it can change.
• Speed: Cloudflare says answers typically land in under a second (vendor claim — not independently verified).
• Catch: It's an early checkpoint. Arithmetic beyond one or two steps is unreliable, forecasts flatten out, and the weights aren't public yet — hosted endpoint only.
• Verdict: Not a chatbot replacement. It's the part of your pipeline that currently makes an LLM read an invoice and format the total as JSON — done by something built for that job.

What "System One extraction model" actually means

The name comes from Daniel Kahneman's Thinking, Fast and Slow: System 1 is fast, intuitive pattern-matching; System 2 is slow, deliberate reasoning. Cloudflare's framing is that LLMs are System 2 machines, and there's a class of work — reading the invoice total, computing how many days late a delivery is, estimating a repair cost — where deliberation is wasted budget.

The category didn't start with Cloudflare. TypeSafe's Jev, a "decision model" that picks among options you define up front and returns probabilities, kicked off the System One wave about a month ago (September 2026). Cloudflare answered with its own decision model, Clef — press coverage dates its launch to October 1, 2026 — and has since added Clef-omni for audio and video. Clairic extends the idea to jobs where you can't list the options in advance. Nobody can pre-list every possible invoice total, delivery date, or repair cost. So Clairic answers with values, not choices: extracted, computed, estimated, or forecast.

How a Clairic request works

The request shape is different from anything you're used to. You send a state — the text to read, plus up to four optional images — and a map of typed fields. Each field names the value you want, declares a type, and can carry plain-language instructions.

Cloudflare's own walkthrough uses a customer support email: "Hi, order A-88412 arrived on Sept 28, 2026 with a cracked screen and a dented corner. I paid $649.00 plus $52.10 tax..." Ask for amount_paid as a number with two decimals and Clairic returns 701.1 — it added the price and tax in the same pass even though neither the sum nor "total" appears in the email. Ask for days_since_delivery as a duration and it works out 11 days from two dates. Ask for a field that isn't in the email at all — the tracking number — and instead of a plausible-looking guess, you get null, with a guess field showing what it would otherwise have picked.

That last detail is the one that matters most to me. Hallucination is the tax on every LLM extraction pipeline today. Clairic doesn't eliminate it — Cloudflare says plainly that it can still make mistakes — but the failure mode is different: a typed value of the declared type, or null with a p_null probability, plus a confidence score. That's machine-reviewable. A chatty LLM answering "Based on my reading of the document, the total appears to be..." is not.

The field types

• text — an exact span of the input, returned with character offsets, so you can highlight the source.
• number — read or computed, honoring integer and decimals constraints.
• date, time, datetime — real calendar values in ISO form.
• duration — a length of time in the unit you choose.
• list — every matching value in input order.
• estimate — a distribution inside your min/max bounds on your step grid, returned as a median with an 80% interval (p10–p90).
• forecast — one estimate per future step of a series, up to 64 steps ahead.

There's no cap on field count: fields that don't fit beside the input in one pass are split across parallel passes that each see the whole input, then merged. Cloudflare's note that "asking for thirty fields costs about the same as asking for three" is the honest pitch — per-field LLM calls are where extraction budgets quietly die.

Pricing: $0.10 per million input tokens, and that's it

The official changelog pricing section is exactly one line: Clairic costs $0.10 per million input tokens. Because it generates no tokens, there is no output token charge.

Put that against the market. In our Claude Haiku 5.5 pricing breakdown, Anthropic's cheapest tier is $0.10 per million input tokens plus $0.50 per million output tokens (for requests under 100K tokens). For a task like pulling six fields from a 3,000-token document — roughly one invoice — Clairic costs about $0.0003. An LLM at Haiku 5.5's cheapest tier would charge the input plus whatever it writes back. The gap widens the more fields you ask for, because Clairic folds them into the same pass while an LLM burns output tokens per answer.

The fine print: it's preview pricing on an early checkpoint, and Cloudflare makes no promise it stays there. Open weights are "intended" — Cloudflare's word — for release "soon," with no date attached. Today the hosted endpoint is the only way in.

Pros and cons

What Clairic gets right

• Grounded answers with receipts. Text comes back as a span with character offsets. You can point at the source.
• Machine-readable failure modes. null, p_null, and confidence scores are designed for routing to human review, not for vibes.
• Typed by construction. Every answer is a valid value of its declared type. No schema validation layer, no retries on malformed JSON.
• Honest about nulls. A missing field returns null instead of an invented value — the single most dangerous failure in document pipelines.
• Input-only billing. $0.10/M tokens with no output charge undercuts any LLM for high-field-count extraction.

Where it falls short

• Preview means preview. Cloudflare calls the endpoint "an early checkpoint" and expects it to improve — meaning it can also regress.
• Math is shallow. One- and two-step arithmetic (totals, differences, date math) is described as reliable. Long chains and aggregations across many table rows are not. Cloudflare's advice: extract the inputs with Clairic, do the math in code.
• Forecasts are basic. Next-step forecasts follow recent patterns; several steps out they flatten toward the average. It doesn't anticipate events the series doesn't show (holidays, shocks). It doesn't replace a time-series model trained on your data.
• Estimates need evidence. Feed it comparables, history, or descriptive text and the 80% interval is useful. Without evidence, it's wide — and Cloudflare admits intervals can run narrow on unfamiliar data.
• Hosted only, for now. No weights to download, no self-hosting, no offline use. If your documents can't leave your network, wait.

Clairic vs Clef vs a traditional LLM

Cloudflare's own taxonomy splits model tasks into three: numbers (extract), choices (decide), and open-ended text (generate). Here's how the options stack up, with API access included:

The honest way to think about it — and this is Cloudflare's own framing — is as a pipeline: extraction models turn raw input into typed values, decision models turn values into decisions, LLMs turn decisions into words. Clairic isn't competing with Claude or GPT; it's competing with the regex soup and the brittle "LLM-as-JSON-printer" pattern that teams have been duct-taping together for two years.

The October context

Clairic landed in the busiest model week of October. On October 19 alone, Inception launched Mercury 3 (diffusion-based, 1,068 tokens/sec, $0.10/$0.35 per million) and MiniMax launched M3.1 Flash (1M context, $0.15/$0.60) — both fighting the speed-and-cost war for the same agent workloads Clairic targets from a different direction. The trend is unmistakable: the industry has stopped asking "how smart can the model be" and started asking "how cheap and fast can the boring parts get." Clairic is that question answered with a new architecture rather than a smaller LLM.

Who should try it now — and who should wait

Try it now if you run document pipelines on Cloudflare Workers already, process invoices/receipts/support tickets at volume, or have an LLM extraction step that occasionally invents numbers. The preview is live and the API is two lines of code.

Wait if your documents contain regulated data you can't send to a preview endpoint, if you need multi-step computation (do that in code anyway), or if you need self-hosted weights — check back when "soon" becomes a release date. And date-heavy workflows should set reference_date or add a "Today:" line to the state, per Cloudflare's own docs, or watch year-less dates drift.

FAQ

What is Cloudflare Clairic?

Clairic is a preview AI model on Cloudflare Workers AI, announced October 19, 2026. It's the first "System One extraction model": instead of generating text, it returns typed values (text spans, numbers, dates, durations, lists, estimates, forecasts) from unstructured text and images in a single forward pass.

How much does Clairic cost?

$0.10 per million input tokens, with no output token charge — because the model generates no tokens. That's the preview price on the Workers AI hosted endpoint, per Cloudflare's official changelog.

Is Clairic a large language model?

No. LLMs generate text one token at a time. Clairic reads your input in one pass and decodes each answer under the rules of its declared type. It can't write prose, answer open-ended questions, or reason through problems — and it isn't trying to. It's built for the jobs where an LLM's flexibility is wasted overhead.

Can Clairic read images?

Yes — each request's state can include up to four images alongside the text. Cloudflare's launch demo extracts order number, totals, tax, and tip from an angled photo of a receipt, returning each as a typed value with confidence scores.

When will Clairic's weights be open-sourced?

Cloudflare says it "intends to release the model with open weights soon," but no date is given. Until then, the Workers AI endpoint is the only access route.

Key takeaways

• Cloudflare's Clairic (preview, launched October 19, 2026 on Workers AI) is the first extraction-model release: it answers with typed values instead of generated text.
• Pricing is $0.10 per million input tokens with zero output charge — a pricing structure no LLM in the comparison table matches.
• Answers come with character-offset spans, p_null probabilities, and confidence scores, making nulls and uncertainty machine-reviewable.
• It's an early checkpoint: shallow arithmetic, flattening forecasts, and hosted-only access are the real limitations.
• Clairic, Clef, and LLMs are a pipeline, not rivals — extract, decide, then write.

Published October 20, 2026. Clairic is less than a day old in preview, the pricing is introductory by nature, and the open-weights release has no date — treat everything above as a snapshot of a fast-moving launch, not a long-term contract.