Cloudflare unveils Clairic, an AI model for pulling data out of text and images
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
Cloudflare has opened a preview of Clairic, a model that pulls amounts, dates, names and lists from text and images on Workers AI. Instead of writing word by word, it answers in one pass with each value in the requested type and how sure it is, and the company plans to release its weights. Its weak point, for now, is multi-step calculations.
On Monday, October 19, Cloudflare introduced Clairic, an artificial intelligence model built to pull specific data out of text and images, such as amounts, dates, names or lists. It is available in preview on Workers AI, the company's service for running AI models on its network, and Cloudflare says it wants to publish it with open weights soon, meaning anyone will be able to download it.
Cloudflare places it among what it calls “System One” models, a category that, according to the company, TypeSafe launched a month ago with its Jev model, drawing on Daniel Kahneman's book Thinking, Fast and Slow: fast, precise models that answer in a single pass, as opposed to “System Two” models, such as language models, which reason step by step. Until now these models, like Jev or Cloudflare's own Clef, were decision models: they choose between several options and give each one a probability. Clairic is the first extraction model.
Clairic does not write its answers word by word like a language model. It reads the input (the text and up to four images) in one pass and gives back each piece of data in the type asked for (a number, a date, a duration, a list…) together with a confidence score and the probability that the value is missing, or a null value if it cannot find it. It also works out figures from the text, such as adding up a price and its tax, estimates values that do not appear but for which there are clues, with an 80% interval, and predicts the next values in a series. According to Cloudflare, users pay only for what it reads, since it produces no output text, and it usually replies in less than a second. Because it does not generate text, it does not “hallucinate” the way a language model does, although it can get things wrong.
In a test run by Cloudflare itself on a complaint email, the chain made up of Clairic, which pulled out the data, Clef, which decided what to do, and the GLM-5.3 language model, which wrote the reply, took 19.3 seconds, compared with 61.7 seconds with GLM-5.3 doing everything, and cost 49% less. With GLM-5.3, both setups got the data and the decision right every time, but Clairic's estimate of the repair cost was the same in each run, while GLM-5.3's changed from one attempt to the next. According to Cloudflare, Gemma 4 and Qwen, doing the whole job on their own, could not complete the task because they failed to work out values such as that cost.
Cloudflare warns that this is an early version it expects to improve: it reads well, but calculations with several steps do not come out right yet, and it recommends leaving the arithmetic to code; its estimates need clues in the input, its intervals can come out too narrow with unusual data, and its forecasts are basic and do not replace a time-series model. It can be used from a Worker or through its REST API. The announcement gives neither the price nor a date for the open weights.
Why it matters · analysis and opinion
Until now, pulling data out of an email, an invoice or a form with AI almost always meant asking a language model, which is slow, charges for every word it writes and sometimes makes up the result. Clairic proposes splitting the work: a model that only reads and returns values with its level of confidence, and the large model only for the part that requires writing. For anyone automating administrative tasks, such as complaints, orders or bookkeeping, the idea is appealing because every step is typed and can be checked. But there is no need to rush: this is a preview, the performance figures come from Cloudflare itself and it is not yet known when the open weights will arrive. The sensible move is to test it with real data and let code do the sums.
Official source: Cloudflare · Written with the help of AI: how we make the news
