3.2x Faster and About Half the Cost Versus an All-LLM Cascade: Meet Clairic!

Hello, it's Shii-chan! Today's story from the Cloudflare Blog is about a rather unusual model.

Cloudflare Blog blog.cloudflare.com

What was announced?

The Cloudflare Blog introduced Clairic, a new class of extraction model. It is in preview on Workers AI as @cf/cloudflare/clairic, and open weights are planned for later. Compared with the usage-focused Changelog post, this one centers on how it works, real examples, and benchmarks.

Why it matters

When you hand everything to an LLM, it can be slow, expensive, and inconsistent from run to run. Cloudflare proposes a pipeline that splits the roles:

• Extraction model (Clairic): turns raw input into values
• Decision model (Clef): turns values into choices
• LLM: writes language

What changes

If dedicated models handle extraction and decisions, the whole pipeline may get faster and cheaper. The real numbers are in Dive Deep below.

Dive Deep

How it works:

• It reads the input in a single forward pass with no reasoning and no output tokens, so you pay only for input
• A request has a state (text plus up to four images) and named, typed fields, each with optional plain-language instructions
• All fields are answered in one parallel pass. If there are too many, the server splits them into parallel passes that each see the full input

Field types are text, number, date, time, datetime, duration, list, estimate, and forecast. Scalar answers include p_null and confidence, and every answer is a value of the requested type or null.

Examples from the post:

• Support email: extracted the order ID, delivery date, and damage spans, and computed the total paid ($701.10) and days since delivery (11). A tracking number that was not in the email came back as null
• Receipt photo: extracted line items, tax, and tip with confidence scores, and returned null for a missing discount
• Home pricing: from comparable sales, estimated a price of $418,000 with a $370,000 to $470,000 interval
• TSA throughput forecast: over 21 daily forecasts, it averaged 3.3% error and 17 of 21 actuals fell within the 80% interval. It missed mainly around the Columbus Day weekend because it does not model holidays

Benchmark on a damage-claim pipeline (Clairic → Clef → GLM-5.3 versus GLM-5.3 for every step):

• End-to-end latency: 19.3 s vs. 61.7 s (about 3.2x faster)
• Extraction: 0.46 s (Clairic) vs. 16.1 s (GLM-5.3)
• Decision: 0.22 s (Clef) vs. 21.9 s (GLM-5.3)
• Cost per million claims: $6,032 vs. $11,740 (about 49% less)
• Clairic's repair estimate was identical across runs, while GLM-5.3's varied

Preview limitations:

• Single-pass multi-step arithmetic and aggregation across many rows are weaker. Let Clairic predict values and let code do the math
• Estimates need evidence in the input
• Interval calibration is measured, not guaranteed, and may be narrow on unfamiliar data
• Forecasting is basic and not a replacement for a dedicated time-series model

Use confidence, p_null, and interval width to decide what to automate and what to review.

Wrap-up

• Clairic is an extraction model that returns typed values in one pass, billed on input tokens only
• Cloudflare proposes splitting extraction, decisions, and text generation across dedicated models
• In the damage-claim example, it was about 3.2x faster and about 49% cheaper than an all-LLM pipeline
• It is not yet suited to long calculations or serious time-series forecasting

If cost and latency in your document-to-values pipeline are a pain, this one is for you!