Cloudflare announced Clairic on October 19, 2026, describing it as the first model in a new System One extraction class. The preview runs on Workers AI as @cf/cloudflare/clairic, callable from a Worker or REST API, with open weights planned later. Pricing is $0.10 per million input tokens, with no charge for output tokens. The pitch for business is direct: typed data from messy inputs at higher accuracy, lower latency and lower cost than large language models.
How Clairic extracts typed data in one pass
Clairic does not generate text. It reads a state containing text plus up to four images and returns a map of typed fields in a single parallel forward pass. Supported types include text, number, date, time, datetime, duration, list, estimate and forecast, each with optional plain-language instructions. The documentation lists a 65,536-token context window and vision support for PNG, JPEG or WebP files up to 4 MiB each. Fields that exceed an 8,192-token pass are split across parallel passes that each see the full input, then merged.
Every scalar answer carries p_null, the probability the value is absent, plus a confidence score. Missing values return null, with a separate guess field showing what the model would have selected, which avoids fabricated entries. Estimate fields return a distribution inside caller-declared bounds, with a median and an 80 percent interval from p10 to p90, and all 99 percentiles on request. Cloudflare says responses typically arrive in under a second because there is no reasoning step and no generated tokens. Billing therefore covers input only.
Cloudflare positions extraction alongside two other model categories. Decision models classify an input among listed options, while language models generate open-ended text. TypeSafe introduced its Jev decision model about a month earlier, and Cloudflare released Clef and Clef-flash on October 1, 2026, open-sourcing them on Hugging Face under Apache 2.0. Clef was post-trained on Qwen base models and is Jev-API compatible. In this scheme, extraction feeds decisions, and decisions feed the final generated reply, with Clairic handling the first stage.
What faster extraction means for operations
For companies processing documents, emails and photos, the change is fewer language-model calls for routine reading tasks. A support-email example computed a $701.10 total from $649.00 plus $52.10 tax, derived 11 days between September 28 and October 9, 2026, listed cracked screen and dented corner as exact spans, and returned null for a missing tracking number. A receipt photo yielded order number, datetime, quantity, unit price, subtotal, tax and tip with confidence scores, and null for a discount not shown. Small firms gain structured intake without custom parsing, while larger operations can standardize extraction before routing to review.
The trade-offs center on verification and scope. Cloudflare calls the endpoint an early checkpoint and notes that one- and two-step arithmetic looks reliable, while longer dependent chains and aggregation across many table rows remain difficult without a scratchpad. Estimates require evidence in the input, calibration is measured rather than guaranteed, and intervals can run narrow on unfamiliar data. Forecasting is described as basic, not a replacement for dedicated time-series models. Buyers should therefore ask how p_null thresholds, confidence scores and intervals map to auto-approve versus manual review.
A useful marker is the open-weights release after further checkpoint work, for which no date was given, alongside independent tests of the claimed pipeline gains. Cloudflare cites a claims workflow where a Clairic plus Clef plus GLM-5.3 cascade finished in 19.3 seconds versus 61.7 seconds and cost $6,032 per million claims versus $11,740. It also cites a 21-day TSA forecast test with 3.3 percent average error. If later checkpoints hold accuracy while weights become openly available, extraction could shift from hosted language-model pipelines to a cheaper first stage.
