Cloudflare Opens Clairic Weights: 27B Costs 2.4x the 9B

Cloudflare has released open weights for two of its own AI models, Clairic and Clairic-flash, the company said on its blog on 18 November.

Cloudflare lists the 27-billion-parameter Clairic at $0.24 per million input tokens from 1 December, 2.4 times the $0.10 it lists for the 9-billion-parameter Clairic-flash, by our arithmetic.

Business Pill · Vertical Integration

An evergreen Business Engineer explainer on vertical integration. It teaches the general idea only and does not discuss Cloudflare or Clairic.

The key insight: Cloudflare open-sourced its 27B Clairic and 9B Clairic-flash extraction models under Apache 2.0 and priced the larger one at 2.4 times the smaller per input token from 1 December, by our arithmetic.

What Cloudflare Released

Both models do one job: pulling structured fields, such as a supplier, line items and a total, out of text, images and PDFs.

Both run on Workers AI, Cloudflare’s hosted inference service. The full weights are on Hugging Face under an Apache 2.0 licence, per the Clairic and Clairic-flash model pages.

Cloudflare opened the post with a new self-description: “Cloudflare has become a neolab.” It says it is “building our own frontier AI models for our customers and the open-source community.”

The team behind the models now calls itself the Workers AI superintelligence team, the post says.

After the promotion ends on 1 December, the 27B Clairic costs $0.24 per million input tokens, 2.4 times the 9B
After the promotion ends on 1 December, the 27B Clairic costs $0.24 per million input tokens, 2.4 times the 9B Clairic-flash’s $0.10 (Cloudflare prices; ratio is our arithmetic).

The Pricing

Cloudflare charges only for input tokens. “we eat the cost of the output token generation”, the post says.

Until 1 December both models cost $0.10 per million input tokens as a promotion. Cloudflare says the 27B Clairic then resumes its $0.24 list price, while Clairic-flash is listed at $0.10.

The preview model that launched at the @cf/cloudflare/clairic endpoint was the 9B version. Cloudflare says that model is moving to the clairic-flash endpoint, and the 27B model takes the clairic name.

What Cloudflare Says Changed

Cloudflare says it redesigned the architecture after the preview. Text fields are now generated with what it calls parallel slot autoregressive generation, with special heads for estimates, dates and arithmetic.

On its synthetic receipt test, Cloudflare reports 576 of 576 fields right for both new models, against 369 for the preview. It says image reading went “from 207 errors to 0” on its internal benchmarks.

Cloudflare calls these evals “self-reported” and says there are no public extraction benchmarks yet.

On speed, it reports a median of 358 milliseconds for Clairic and 155 for Clairic-flash on its fresh text test. That makes the small model about 2.3 times faster at the median, by our arithmetic.

Bar chart: receipt fields read correctly out of 576 - Clairic-flash 9B 576, Clairic 27B 576, Liquid AI D1.2 plus OCR 397, Clairic preview 369, ScaleDown 280. Data: Cloudflare internal evals, self-reported.
On Cloudflare's own synthetic receipt test, both new models read all 576 fields, against 369 for the preview; Cloudflare calls the evals self-reported.

Where This Fits

We covered the preview in Cloudflare Says Clairic Cut AI Pipeline Cost 49% in Its Test and its earlier model in Cloudflare Cuts Clef-flash to $0.038 by Shrinking Context.

Cloudflare’s stated aim is to help customers “augment their expensive LLM usage with smaller, faster, cheaper models that are purpose-built for a specific task”.

The Structural Read

As we read the post, Cloudflare is selling the cost of a narrow task rather than a general model: it charges only input tokens and absorbs the output.

As we read it, the neolab language is a positioning claim; the post gives no team size, usage or revenue to size it.

Cloudflare blog, 18 Nov 2026

“Cloudflare has become a neolab.”

Three Implications

FOR DEVELOPERS Two task-specific models on Workers AI, both $0.10 per million input tokens until 1 December, per Cloudflare.

FOR OPEN-SOURCE USERS Full weights under Apache 2.0 on Hugging Face, per the model pages.

FOR CLOUDFLARE A second open-weight family after Clef, from a team it now calls the Workers AI superintelligence team, per the post.

What Is Not Established

The accuracy and speed figures are Cloudflare’s own tests, not independent benchmarks. Cloudflare did not give the size of the superintelligence team, usage numbers or revenue from Workers AI in the post.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The Map of AI

A Business Engineer map of 200+ AI companies across 9 layers.

Read the Map of AI →

The Bottom Line

Cloudflare open-sourced two extraction models and called itself a neolab. Its list price for the 27B Clairic after 1 December is 2.4 times the 9B Clairic-flash’s per input token, by our arithmetic.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. The release, prices, promotion, architecture, evals and latency figures are from Cloudflare’s 18 November blog post; the licence and parameter counts are from Hugging Face. The evals are Cloudflare’s own, which it calls self-reported. The 2.4x and 2.3x figures are our arithmetic.

Sources: Cloudflare, ‘An updated Clairic and Clairic-flash are now open-weight, from the new Workers AI Superintelligence team’, 18 Nov 2026 · Hugging Face, Cloudflare/clairic model page · Hugging Face, Cloudflare/clairic-flash model page

Scroll to Top
FourWeekMBA