# Cloudflare Clairic Clairic (`@cf/cloudflare/clairic`) is an extraction model that [[Cloudflare]] released in preview on [[Cloudflare Workers AI]] on 19 October 2026. Cloudflare calls it the first of a new class of [[System One Models|System One]] models: extraction models. [[Cloudflare Clef|Clef]] and [[Jev]] answer "which of these options?". Clairic answers "what's the value?". You give it some text (plus up to 4 images) and a map of typed fields, and it returns a value of the requested type for each field, or `null`. It never writes text. It "only processes input tokens with a single forward pass", so there's nothing to parse and no output tokens to pay for. Cloudflare is upfront about what that buys you: "The architecture makes every answer well formed, not necessarily correct." ## What it returns - **Extracted values.** A `text` field returns a span copied from the input, with its character offsets. You can highlight exactly where the answer came from - **Computed values.** `number`, `date`, `time`, `datetime`, `duration` and `list`, worked out from the input (e.g., the number of days since delivery). `reference_date` anchors dates written without a year - **Estimated values.** An `estimate` field returns a distribution within the bounds you declare (`min`, `max`, `step`, optionally a log scale): `median`, `mean`, an 80% interval (`p10` to `p90`), plus `p_below` and `p_above` - **Forecasted values.** A `forecast` field returns one estimate per future step, each with its own 80% interval Every scalar answer also comes with `p_null` (the probability that the input doesn't determine the value) and a `confidence`. That's the part I care about most, because you can route on it: low confidence or a high `p_null` goes to a human or an LLM ([[Confidence-Gated Routing]]). ```json { "state": "Invoice INV-2291 from Northwind Supply, issued 2026-09-02, net 30. ...", "fields": { "invoice_number": { "type": "text" }, "days_to_pay": { "type": "estimate", "min": 0, "max": 120, "step": 1 } } } ``` ## Specs - **Context:** 65,536 tokens. Fields are packed into 8,192-token passes that run in parallel, each one seeing the full input, so there's no limit on the number of fields - **Images:** up to 4 (PNG, JPEG or WebP, 4 MiB each, 8 MiB decoded in total). Remote URLs aren't accepted - **Price:** $0.10 per million input tokens, and no output charge - **Access:** the Workers AI binding (`env.AI.run("@cf/cloudflare/clairic", ...)`) or the REST API at `/ai/run/@cf/cloudflare/clairic` - **Status:** preview, "an early checkpoint", provided "as is". Cloudflare intends to release the weights "soon", with no date. Unlike for Clef, there's no word yet on the architecture, the backbone or the training ## Extract, decide, write Cloudflare's pitch is a three-stage cascade. Extraction models turn raw input into typed values, [[Decision Models (DMs)|decision models]] turn those values into decisions, and LLMs turn decisions into text. Their demo handles damage claims sent by email. Clairic pulls the customer name, the order ID, the amount paid, the days since delivery and the damage, and estimates the repair cost (USD, log scale between 10 and 5,000). Clef then picks a full refund or a free repair from a five-line summary plus the policy (refund within 30 days; repair when it costs under 60% of the price and the product still works). Finally, GLM-5.3 writes a reply of under 60 words. The numbers below come from Cloudflare, so treat them as vendor benchmarks: - **Speed:** 19.3 s end to end, against 61.7 s when GLM-5.3 handles every step (about 3.2x faster). Clairic's extraction took 0.46 s and Clef's decision 0.22 s - **Cost:** $6,032 per million claims, against $11,740 (49% less) - **Consistency:** Clairic returned the same repair estimate on every run. GLM-5.3 gave $250, $250, then $220. [[Gemma 4]] and [[Qwen]] failed the task in the all-LLM version - **Forecasting:** on TSA checkpoint data over 21 days, forecasts were off by 3.3% on average, and 17 of the 21 actual counts fell inside the 80% interval (three of the four misses were on the Columbus Day weekend) ## Limits Cloudflare lists these itself: - Multi-step arithmetic isn't reliable yet - Estimates need evidence in the input - Intervals "can run narrow on unfamiliar data". The calibration is measured, not guaranteed ([[AI Model Calibration]]) - Forecasting is basic. It doesn't replace a dedicated time-series model ## My take Clef turned the decision interface into a commodity. Clairic tries the same trick on [[LLM Structured Outputs|structured extraction]], and that's a bigger market. Most of the LLM calls in my automations boil down to "pull these five fields out of this email", and JSON mode never gave me per-field `p_null`, span offsets or the same answer twice. The span output also gets you most of what [[Candidate-Then-Select Extraction]] is for: a `text` answer is copied from the input, so it can't be invented. Computed and estimated values are another story. That's where I'd look for errors first. Would I use it today? For triage-style extraction on Cloudflare, yes, behind a confidence threshold. But it's a preview, the architecture isn't public, the weights aren't out, and every number so far comes from Cloudflare. Test it on your own documents before you trust a threshold. ## References - [Introducing Clairic, a new System One class of extraction models (Cloudflare blog, 2026-10-19)](https://blog.cloudflare.com/clairic-extraction-model/) - [Clairic model page (Workers AI docs)](https://developers.cloudflare.com/workers-ai/models/clairic/) - [Extraction models (Workers AI concepts)](https://developers.cloudflare.com/workers-ai/concepts/extraction-models/) - [Clairic pipeline demo](https://clairic-pipeline-demo.workers-ai-mle.workers.dev/) - [Introducing System One Models & Jev (TypeSafe AI)](https://typesafe.ai/blog/introducing-system-one-models-and-jev) ## Related - [[Cloudflare Clef]] - [[Cloudflare Workers AI]] - [[Cloudflare]] - [[System One Models]] - [[System One Primitives]] - [[Decision Models (DMs)]] - [[Jev]] - [[TypeSafe AI]] - [[Candidate-Then-Select Extraction]] - [[LLM Structured Outputs]] - [[Confidence-Gated Routing]] - [[AI Model Calibration]] - [[AI Open Weight Models]] - [[AI Model Cascades]] - [[LLM-Generated Features for Classical ML]]