Cloudflare's Clairic Extracts Typed Data in 0.46 Seconds

Cloudflare's new extraction model Clairic completed a document extraction step in 0.46 seconds — 35 times faster than the 16.1 seconds GLM-5.3 needed for the same job, according to Cloudflare's own benchmark. The company released Clairic as a preview on Workers AI on October 19, and it represents a bet that not every task developers currently hand to large language models needs generated text at all.

Clairic reads text or images and returns typed values: totals, dates, IDs, durations, estimates and forecasts delivered as JSON. It answers every requested field in a single forward pass, with no reasoning and no output tokens. That's the whole trick, and it explains both the speed and the price — $0.10 per million input tokens, with output costing nothing.

The stakes here are straightforward. Developers routinely use expensive reasoning LLMs to pull invoice totals into JSON, a job that wastes most of what those models do. Cloudflare's argument, laid out in its blog post, is that extraction, decisions and prose belong to three different model classes, not one.

What exactly is Clairic?

Clairic builds on the "System One" idea from TypeSafe's Jev model, named for psychologist Daniel Kahneman's fast, intuitive mode of thinking. Jev and Cloudflare's earlier decision model Clef score options you list in advance. Clairic handles the answers you can't list ahead of time — an invoice total, a delivery date, a repair cost.

Cloudflare frames this as a three-part stack:

• Extraction models turn raw input into typed values.
• Decision models turn values into decisions.
• LLMs turn decisions into words.

The model offers a 65,536-token context window and accepts up to 4 images per request. Its parameter count is not disclosed. It runs on Cloudflare-hosted GPUs under the model ID @cf/cloudflare/clairic, with open weights promised but not yet shipped.

The takeaway: Clairic is a specialist, built to replace the extraction step developers currently bolt onto general-purpose LLMs.

How does a Clairic request work?

A request carries a state — text plus up to 4 images — and a map of named, typed fields, each of which can include plain-language instructions. Supported types include text, number, date, time, datetime, duration, list, estimate and forecast, per the Workers AI changelog.

Every scalar answer ships with p_null, the probability the value is absent, plus a confidence score. Text answers return exact spans with character offsets. Estimates and forecasts return a median and an 80% interval from p10 to p90, with forecasts covering up to 64 steps.

There's no field limit, but there's a billing catch. Fields that don't fit beside the input in one 8,192-token pass split into parallel passes, each seeing the whole document. In Cloudflare's docs example, a 61-token invoice ran in 3 passes and billed 655 input tokens — 472 of them schema tokens. Wide schemas can cost more than the document itself.

In short: the API is honest about uncertainty, but schema-heavy jobs pay a hidden tax.

What did Cloudflare's tests show?

In a support-email example, Clairic added price and tax into $701.10 and computed 11 days since delivery — neither number appears in the email. A missing tracking number returned null instead of a plausible guess. On a skewed receipt photo, it read all 7 present values, returned null for the absent discount, and logged its lowest confidence, 0.776, on the order number.

The forecasting test was more mixed. Predicting daily TSA checkpoint counts for 21 days, Clairic posted an average error of 3.3%, with 17 of 21 actual counts inside its 80% interval. Three of the four misses fell on Columbus Day weekend. Multi-step arithmetic, Cloudflare acknowledges, is unreliable.

The headline test is a damaged-product claims pipeline. Clairic extracted facts, Clef chose refund or repair, and GLM-5.3 wrote the reply, benchmarked against GLM-5.3 doing all three steps:

• The cascade finished in 19.3 seconds versus 61.7 seconds — roughly 3.2x faster.
• Extraction took 0.46 s against 16.1 s; the decision step took 0.22 s against 21.9 s.
• Cost per million claims fell 49%, from $11,740 to $6,032.
• The reply step still dominated at 18.6 of the 19.3 seconds.

Both pipelines got the facts and decision right, but consistency differed. Clairic's repair estimate was identical across runs; GLM-5.3 returned $250, $250 and $220. Cloudflare also says Gemma 4 and Qwen failed the all-LLM version outright.

The takeaway: the numbers are real, but they come entirely from Cloudflare's own pipeline — no standard accuracy benchmark has been published.

How does Clairic stack up against Clef, Jev and GLM-5.3?

Metric	Clairic	Clef	Jev	GLM-5.3
Developer	Cloudflare	Cloudflare	TypeSafe AI	Z.ai
Class	System One extraction	System One decision	System One decision	System Two LLM
Output	Typed values, spans, forecasts	Probabilities over listed options	Calibrated decisions	Free-form text
Context window	65,536 tokens	65,536 tokens	Not disclosed	1,048,576 tokens
Image input	Yes, up to 4	Yes, incl. video	Not disclosed	Not disclosed
Input price	$0.10 / M tokens	$0.24 / M tokens	$0.042 / M tokens	$1.40 / M tokens
Output price	None	None	Free	$4.40 / M tokens
Speed	0.46 s extraction	209.3 ms median	70–500 ms end to end	16.1 s extraction
Availability	Workers AI preview	Workers AI, Apache 2.0 weights	TypeSafe API, early access	Workers AI, paid plan

Clairic undercuts Clef on price but sits well above TypeSafe's Jev at $0.042 per million tokens. GLM-5.3, at $1.40 input and $4.40 output per million tokens, costs an order of magnitude more for jobs it doesn't do well.

The takeaway: Cloudflare is assembling a cheap, fast alternative stack to reasoning models — and pricing it to move developers off LLM-based extraction.

Where can developers use it?

Clairic is live as a preview, callable through the Workers AI binding with env.AI.run() or the REST API, and it works with AI Gateway on the same endpoints. Until Cloudflare ships the promised weights, there's no local option.

Sub-second extraction, calibrated uncertainty and zero output-token pricing make the economics hard to ignore for high-volume document pipelines. The open question is accuracy on harder inputs, and that won't be answered until independent benchmarks exist — or until the weights land and researchers can test them directly.