Back to Blog
    AI & AutomationNov 19, 2026

    Cloudflare Clairic: Open-Weight AI That Reads Your Business Documents

    Amir

    Amir Farid

    7 min read

    Cloudflare Clairic: Open-Weight AI That Reads Your Business Documents

    Cloudflare Clairic reads your business documents and returns structured data in under a second. Launched on 18 November 2026, Clairic (27B) and Clairic-flash (9B) are open-weight models that extract typed information from text, images and PDFs — receipts, invoices, tickets, leases and expense audits. Both run on Workers AI, with weights on Hugging Face under Apache-2.0.

    The short version: if paperwork is your bottleneck, a small specialist may now beat a big general LLM on cost, speed and accuracy. Clairic costs $0.24 per 1M input tokens, Clairic-flash $0.10, output tokens free — both $0.10 under a promotion until 1 December 2026. Details: Cloudflare's own announcement.

    Cloudflare Clairic: Open-Weight AI That Reads Your Business Documents

    What Cloudflare Clairic actually is

    Clairic is a document-extraction model, not a chatbot. Give it a document and a schema — the fields you want back — and it returns structured, typed data, with confidence scores, character spans and probabilities, according to Cloudflare. Clairic (27B) suits the hardest documents; Clairic-flash (9B) suits high-volume routine work. Both are open-weight, from Cloudflare's Workers AI superintelligence team.

    What it reads

    Clairic handles text, images and PDFs, up to four images or PDF pages per request — receipts, invoices, tickets and leases. Line items come back as rows, keeping description, quantity and price together.

    Why a small specialist can beat a big general LLM

    Hands sorting paper receipts beside a laptop

    General models are brilliant generalists, but extraction rewards narrow competence. A big model reads a receipt well — and charges big-model prices per page, waits longer, and may invent a plausible total. A specialist does one task: find the fields, type them, and say how sure it is.

    That trade-off is familiar: route routine work to a cheaper model, reserve the frontier model for hard jobs — see our GPT-6.2 Astra and Sol guide. Extraction is repetitive and easy to verify, so cost per document matters more than reasoning power. If your AI automation pipeline re-reads the same invoice formats all day, a specialist is the obvious saving.

    Speed and price, in plain numbers

    • Latency: Cloudflare reports sub-second responses — p50 of 155–691ms for Clairic and 155–393ms for Clairic-flash in its own tests, not independent measurements.
    • Price: Clairic $0.24 per 1M input tokens; Clairic-flash $0.10. Output tokens are free, so you pay only for what goes in.
    • Promotion: both cost $0.10 per 1M input tokens until 1 December 2026 — a cheap window to pilot.

    At $0.10 per 1M input tokens, a thousand short receipts cost cents — measure cost per completed task, per our AI agent cost guide for the UK.

    Built for numbers, dates and line items

    Clairic's architecture targets the ways documents go wrong: parallel slot autoregressive generation, plus special heads for estimates, forecasts, durations, calculated dates and arithmetic, per Cloudflare. In plain English: dedicated machinery for the fiddly bits — a date from "30 days from invoice", or adding up a column.

    Confidence scores are the real feature

    A model that says "I am 62% sure this is the total" beats one that states a wrong total confidently. Auto-accept high-confidence fields, route low-confidence ones to a human, and log the spans so a reviewer can check the source in one click — the pattern we use building AI agents on real data, and behind good RAG chatbot design.

    Where Clairic fits in a real business

    Start with the paperwork your team hates. Accounts payable: invoices arrive as PDFs, emails and photos, and someone retypes them. Expense audits: receipts in every format, totals to reconcile, limits to check. Tickets, leases and loan pricing follow the same pattern — semi-structured, high volume, copied by hand.

    Five use cases worth piloting first

    • Invoice processing: supplier, date, line items, tax and total into your accounting flow, low-confidence totals routed to a reviewer.
    • Expense receipts: merchant, date, amount and category from photos, policy checks applied automatically.
    • Support ticket triage: structured fields from ticket text and attachments, so routing works on data, not guesswork.
    • Lease abstracts: key dates, amounts and obligations into a searchable register.
    • Restaurant and retail paperwork: supplier invoices reconciled against orders — the grind our restaurant ERP vs spreadsheets guide describes, minus retyping.

    Clairic slots into the step where a document becomes data, before the rest of the workflow runs — see our n8n vs Make vs Zapier comparison and our AI automation guide for logistics companies.

    What Cloudflare claims — and what is not yet proven

    Honesty matters. Cloudflare says there are no public extraction benchmarks yet, so its eval tables are self-reported internal evals. Image-reading errors falling from 207 to 0, and 1,263 of 1,263 customer assertions passing, are Cloudflare's own claims — reasons to pilot, not proof.

    How to test it honestly

    Take 200 real documents — the messy ones, not clean samples. Define the fields and correct answers in advance, then score each field: right, wrong, or missing. Track how often the confidence score predicts an error; that sets your human-review load. Our GDPx benchmark post showed even the best AI agents score only around 12% on long-horizon tasks independently — measure Clairic on your own documents, not vendor slides.

    Workers AI or self-hosted: which route to take

    Workers AI is the fast route: call @cf/cloudflare/clairic or @cf/cloudflare/clairic-flash as an API, pay per input token, and let Cloudflare handle hosting. Start here — while the promotion makes the experiment nearly free.

    Self-hosting is the control route: Apache-2.0 licensing lets you run the models on your own infrastructure — useful for regulated work or data-residency rules in the UK, UAE and Saudi Arabia. The trade-off: you own the GPUs, updates and uptime. Many teams pilot on Workers AI, then self-host where control justifies it. Our custom software development team builds these pipelines; our custom software cost guide for the UK explains the budget.

    A simple four-week pilot plan

    • Week 1: pick one document type, 200 samples, fields and ground truth; run Clairic-flash on Workers AI.
    • Week 2: score accuracy field by field and set confidence thresholds for auto-accept vs human review.
    • Week 3: wire the output into your workflow and measure cost per document and time saved.
    • Week 4: decide: scale on Workers AI, test Clairic (27B) on hard documents, or self-host sensitive work.

    Keep a human in the loop: extraction feeding money, contracts or customer records should never run unattended in its first month — the same rule as for local personal agents and ecommerce AI agents.

    Clairic vs the alternatives

    Clairic is not the only way to read a document. Big general LLMs can do it, and OCR-plus-LLM pipelines have existed for years. Compare per-document: accuracy, cost, latency and engineering effort. A general LLM may win on a weird one-off; Clairic is built for the thousand that look alike.

    Cloudflare's pitch: extraction runs where your traffic already runs. If your stack is on Cloudflare, integration is cheap; if not, the open weights mean no lock-in. If your documents are mostly chat — customer questions, not paperwork — a different tool fits: see our RAG chatbot guide and Mistral Large 4 guide. For build costs, read our custom CRM development cost guide.

    Frequently asked questions

    What is Cloudflare Clairic?

    Clairic is Cloudflare's open-weight document-extraction AI, launched 18 November 2026 as Clairic (27B) and Clairic-flash (9B). It extracts structured, typed data from text, images and PDFs, with confidence scores and source spans. It runs on Workers AI; weights are on Hugging Face under Apache-2.0.

    How much does Cloudflare Clairic cost?

    Clairic costs $0.24 per 1M input tokens and Clairic-flash $0.10, with output tokens not charged. Until 1 December 2026, both cost $0.10 under promotion. Those are Cloudflare's published prices; check Cloudflare's announcement for current terms.

    Is Clairic really faster than a general LLM?

    Cloudflare reports sub-second responses — p50 of 155–691ms for Clairic and 155–393ms for Clairic-flash in its own tests, not independent measurements. Run your own pilot before promising speed to stakeholders.

    Can I self-host Clairic?

    Yes — the weights are open on Hugging Face under Apache-2.0, so you can run them on your own infrastructure. Pilot first on Workers AI.

    Should my business switch to Clairic now?

    Pilot first, on your own documents. Clairic looks strong for high-volume paperwork — invoices, receipts, forms and tickets. Measure accuracy over a few hundred samples and scale only if it beats your current process. For help, contact Coheser — we build document and automation pipelines for the UK, USA, UAE, Saudi Arabia and Pakistan.

    #Cloudflare Clairic#Document AI#AI Automation#Open-Weight AI