Cloudflare released Clairic on October 20, 2026. The preview introduces a System One extraction model on Workers AI that reads text or images and returns typed values instead of generated text. It targets a familiar developer task: pulling totals, dates, tracking IDs, and other fields into JSON without asking an LLM to improvise its way through the paperwork.
Clairic answers every requested field in one forward pass, with no reasoning and no output tokens. Cloudflare says its tests produced sub-second extraction, including a 0.46-second extraction step compared with 16.1 seconds for GLM-5.3 in the company’s claims pipeline.
The model accepts a state containing text and up to four images, plus a map of named, typed fields. Its 65,536-token context window gives those inputs room to breathe, although the parameter count remains undisclosed and there is no local option until the weights ship. Clairic runs on Cloudflare-hosted GPUs through Workers AI.
A schema for answers that do not fit a list
Clairic supports text, number, date, time, datetime, duration, list, estimate, and forecast fields. That covers answers developers cannot simply list in advance, such as an invoice total, a date, or a repair cost—precisely the sort of work that has encouraged people to point general-purpose language models at every document in sight.
Every scalar answer includes p_null, the probability that the value is absent, along with a confidence score. Text answers return exact spans with character offsets, while estimates and forecasts return a median and an 80% interval from p10 to p90; forecasts can cover up to 64 steps.
Fields that do not fit beside the input in one 8,192-token pass split into parallel passes. In a support-email example, Clairic added price and tax to reach $701.10 and computed 11 days since delivery. It returned null for a missing tracking number instead of manufacturing a plausible guess, which remains a surprisingly ambitious feature in document automation.
On a skewed receipt photo, Clairic read all seven present values and returned null for the absent discount. Its lowest confidence score was 0.776, attached to the order number. These examples show the model’s intended behavior: extract what exists, mark what does not, and expose uncertainty instead of hiding it behind fluent prose.
Faster claims processing, with limits
The main test uses a damaged-product claims pipeline. Clairic extracted the facts, Clef chose refund or repair, and GLM-5.3 wrote the reply. The cascade finished in 19.3 seconds versus 61.7 seconds for the comparison pipeline—about 3.2 times faster.
The speed came from separating jobs that are often handed to one model. Clairic’s extraction step took 0.46 seconds against 16.1 seconds for GLM-5.3, while the decision step took 0.22 seconds against 21.9 seconds. Both pipelines got the facts and the decision right, but Clairic’s repair estimate stayed identical across runs; GLM-5.3 returned $250, $250, and $220.
Cloudflare says the all-LLM version failed with Gemma 4 and Qwen. The comparison frames Clairic as a System One extraction model, alongside Clef and TypeSafe’s Jev as decision models, while GLM-5.3 serves as the reasoning model. Jev’s name comes from Kahneman’s fast, intuitive thinking, and Clairic builds on that System One idea.
The same design also handles forecasts, though the results are less tidy. Clairic predicted daily TSA checkpoint counts for 21 days with an average error of 3.3%; 17 of 21 actual counts landed inside the 80% interval, while four fell outside it. Three of those four misses occurred during Columbus Day weekend, and multi-step arithmetic remains unreliable.
No standard accuracy benchmark has been published yet, so the claims-pipeline results carry more weight than a neutral scoreboard would. Clairic is a Cloudflare model for extracting, computing, estimating, and forecasting typed values from unstructured inputs—not a replacement for every reasoning model wearing a business-casual API.
Pricing is $0.10 per million input tokens with no output charge. In the claims test, that design cut cost per million claims by 49%, from $11,740 to $6,032. The preview’s real proposition is less glamorous than a general AI assistant: use a focused model for structured work, reserve reasoning for decisions, and stop paying a language model to write tokens nobody needs.
Based on
