Clairic prices input tokens at $0.10 per million and bills nothing for output, because it produces none. It is a single forward pass through a neural network: feed it text and a list of typed fields, and it returns a date, a dollar amount, a probability distribution, or a forecast, without generating a single word of explanation. Cloudflare put the model into preview on Workers AI as @cf/cloudflare/clairic on October 19, with open weights promised to follow, as described in Cloudflare’s launch post.
That distinction, no output tokens, matters because output tokens are what large language models spend most of their latency and most of a customer’s bill generating. An LLM asked to pull an invoice total out of an email reasons its way to an answer one token at a time, at anywhere from $0.20 to $10 per million input tokens with output priced roughly five times higher, according to figures TypeSafe published when it opened this category a month earlier. Clairic instead answers every requested field in one parallel pass across the whole input. Cloudflare calls this an extraction model, a new branch of what it and TypeSafe both label System One models, a name borrowed from Daniel Kahneman’s distinction between fast, intuitive judgment (System One) and slow, deliberate reasoning (System Two, where the chatbot-style LLM sits).
Jev, Clef and now Clairic #
TypeSafe, a San Francisco lab founded by Diogo Almeida, a co-author of the InstructGPT paper that underpins ChatGPT’s training method, launched the first such model, Jev, on September 15. Jev answers typed questions, a choice among up to 255 options, a score on a described scale, a yes/no probability, with a probability attached to each possible answer, in 70 to 500 milliseconds against 3 to 329 seconds for a frontier LLM doing comparable reasoning, priced at $0.042 per million input tokens, according to TypeSafe’s launch post. Demand was high enough that TypeSafe briefly lost the ability to serve its own API, TechCrunch reported, with Vercel citing results five to 18 times faster than the LLM it replaced on a safety-classification task.
Cloudflare answered on October 1 with Clef and Clef-flash, its own decision models (27 billion and 9 billion parameters, built on Qwen3.8-27B and Qwen3.5-9B), open-weighted under Apache 2.0 and priced at roughly $0.24 and $0.09 per million input tokens with output unbilled. On the BANKING77 intent-classification benchmark, Cloudflare reports Clef scoring a macro-F1 of 94.20 against Jev’s 79.74, and Clef-flash answering in a median 38.8 milliseconds against Jev’s 524.1 milliseconds, in Cloudflare’s own announcement. Those are Cloudflare’s numbers, on Cloudflare’s chosen benchmark, testing a rival’s model; TypeSafe has not published a competing BANKING77 score for Jev. Clairic is the third leg Cloudflare has now built: Clef picks among a fixed set of options, Clairic pulls out or computes numbers and typed values, and an LLM still writes the prose a human reads.
How a request works #
A Clairic call sends a state, text plus up to four images, and a map of named fields, each with a type (text, number, date, time, duration, list, estimate or forecast) and an optional plain-language instruction. Every scalar answer returns with p_null, the model’s estimated probability that the field isn’t actually present in the input, and a confidence score. Estimates come back as a median plus an 80% interval (p10 to p90), with all 99 percentiles available on request; forecasts run up to 64 steps ahead, each with its own interval. There is no cap on the number of fields: a request too large for one pass gets split across parallel calls that each see the entire input, then merged, according to Cloudflare’s developer changelog.
Cloudflare’s own worked example is a customer email: an order arrived damaged, the customer paid $649.00 plus $52.10 tax, and wrote eleven days after delivery. Clairic correctly computed the combined total ($701.10) and the day count (11) as derived values, not just copied text, and returned the order number with p_null of 3.4 × 10⁻⁶ and confidence 0.999957. Asked for a tracking number that the email never mentions, the model still internally flagged the tax figure, $52.10, as a tentative guess, but assigned it only 0.416 confidence and correctly returned null rather than a wrong but well-formatted answer. That is the point of the confidence score: it flags its own best wrong guess rather than silently emitting one.
The extract-decide-write pipeline How Cloudflare's claims-processing demo chains three model types
| From | To | How |
|---|---|---|
| Customer claim text | Clairic (extracts typed values, 0.46s) | text + fields |
| Clairic (extracts typed values, 0.46s) | Clef (picks a decision, 0.22s) | typed values |
| Clef (picks a decision, 0.22s) | GLM-5.3 (writes the reply, 18.6s) | decision |
| GLM-5.3 (writes the reply, 18.6s) | Reply to customer |
Based on Cloudflare The diagram shows where the time actually goes in Cloudflare’s benchmark: extraction and decision are fast regardless of which model does them, but the final step, turning a decision into a written reply, still needs an LLM and still takes seconds.
What the benchmark shows, and what it doesn’t #
Cloudflare ran a cascade, Clairic extracts, Clef decides, GLM-5.3 writes the reply, against a baseline doing all three steps with GLM-5.3 alone, on a claims-processing demo. The cascade finished in 19.3 seconds against 61.7 seconds for the all-LLM version, 69% less time, about 3.2 times faster, at a total cost of $6,032 per million claims against $11,740, a 49% reduction, per Cloudflare’s post. Broken down by step: extraction took Clairic 0.46 seconds against GLM-5.3’s 16.1 seconds; the decision step took Clef 0.22 seconds against GLM-5.3’s 21.9 seconds.
Where the speed gain comes from Time per pipeline step, seconds
| | System One model (s) | GLM-5.3 (all-LLM) (s) |
|---|---|---|
| Extraction | 0.46 | 16.1 | | Decision | 0.22 | 21.9 |
Source: Cloudflare Notice what the chart leaves out: the final write-the-reply step, done by the same GLM-5.3 in both versions, actually took longer in the cascade (18.6 seconds) than in the all-LLM run (16.4 seconds). All of the pipeline’s speed gain comes from skipping two LLM calls, not from any faster writing. Cloudflare also reports that Clairic returned the identical repair-cost estimate across repeated runs on the same input, where GLM-5.3 returned $250, $250 and $220, with a different interval each time, a reproducibility gap that matters for anything an auditor might later check. Separately, Cloudflare tested Clairic forecasting 21 days of daily TSA checkpoint traveler counts: the model’s median forecasts were off by 3.3% on average, and 17 of the 21 actual counts, 81%, fell inside its stated 80% interval, close to what a properly calibrated model should produce. The four misses clustered around Columbus Day weekend, a holiday effect the model apparently hadn’t learned. Cloudflare itself flags weak multi-step arithmetic, miscalibrated intervals on unfamiliar data and weak long-range forecasting, and calls the current release “an early checkpoint.” None of these benchmark numbers come from an outside evaluator; they are Cloudflare measuring its own product against a baseline, GLM-5.3, that Cloudflare chose.
The money behind the model class #
TypeSafe’s bet on this category has already been repriced once, hard. The company raised $40 million in seed funding led by DCVC when it launched Jev on September 15, at a reported $200 million valuation. Twenty-four days later, on October 9, it closed an $870 million Series A led by Andreessen Horowitz, with Sequoia Capital and DCVC also participating and a16z partner Martin Casado joining the board, a $7.5 billion valuation, according to TypeSafe’s own announcement and a16z’s investment post. That is a roughly 37-fold step-up in valuation in under four weeks.
The adoption numbers behind that step-up don’t quite agree with each other. TypeSafe’s own blog says “a third of the Fortune 500” use Jev; a16z’s post, investing in the same company, puts the figure at 25%. a16z also calls Jev the fastest-growing model it has tracked, citing 1 trillion tokens generated in the three days after launch; separately, Almeida has described the business running at roughly 1 trillion tokens per day, a different and much larger quantity than a one-time three-day total, and coverage of his remarks has not reconciled the two figures. Taking a16z’s lower, three-day number as a steady daily rate and applying Jev’s own list price of $0.042 per million input tokens (treating input and total volume as the same, since output tokens aren’t billed) works out to about $42,000 a day, or $15.3 million a year, against a $7.5 billion valuation: roughly 489 times that run rate. None of this makes Jev’s underlying technology slower or less accurate than claimed. It does mean the valuation is pricing in a great deal of future growth that a single self-reported token count, measured by the company being valued, cannot itself confirm.
Cloudflare has not disclosed revenue or usage figures for Clef or Clairic, and neither model has attracted a funding round because Cloudflare is a public company bankrolling the research itself. The open question is the same one that applies to every number in Cloudflare’s own launch post: whether the 3.2-times speedup and the 49% cost cut survive contact with a workload nobody at Cloudflare picked, and whether the promised open weights for Clairic actually ship, as opposed to staying, like TypeSafe’s reconciled token counts, a figure quoted once and never checked again.