We looked at S3 first — it's the obvious default and the DX is solid. But egress fees are a hidden tax that quietly compounds as your user base grows. Every file your users download costs you extra, and it adds up fast.
R2 gives you the exact same S3-compatible API, so there's no learning curve and no SDK migration. We literally just updated our config. Zero egress costs since day one, backed by Cloudflare's global network we were already using. It wasn't even a close call.
Clairic is Cloudflare’s new extraction model that pulls structured answers out of text and images in a single pass, without generating output tokens.
Problem: Many repetitive structured-output jobs, like invoice totals, delivery dates, repair costs and short-term traffic forecasts, get sent to frontier LLMs. That is slower and costlier than it needs to be, and the answers can come back mistyped or hallucinated.
Solution: Clairic takes a request with a state (text, plus up to four optional images) and a map of typed fields. It answers all the fields in one parallel pass and returns each value with a confidence score.
What makes it different: It doesn’t write an answer token by token. It extracts, computes, estimates and forecasts directly, and it returns null when the input doesn’t contain the answer. Estimates come back as a median with an 80% interval, so you can see how sure it is. Per Cloudflare, in a claims pipeline, a Clairic → Clef → GLM-5.3 cascade took 19.3 s against 61.7 s for GLM-5.3 on every step, at $6,032 against $11,740 per million claims.
Key features:
Field types include text, number, date, time, datetime, duration, list, estimate and forecast
Extracted text spans with character offsets
Computed values, such as totals and date differences
Estimates and forecasts with 80% intervals
Confidence and null probability on every answer
Charged for input tokens only
Who it’s for: Developers building document processing, claims, invoicing or forecasting workflows who want fast, cheap structured outputs and can put a larger model behind it for the harder steps.
Cloudflare blog post, the model docs and the live demo.