Cloudflare is expanding its frontier AI research effort with two open-weight extraction models, as its newly introduced Workers AI superintelligence team pushes into building specialized artificial intelligence for customers and the open-source community. The redesigned Clairic and Clairic-flash models promise faster, more accurate extraction from text, images and PDF documents.
According to Cloudflare Blog, the releases represent a substantial redesign of Clairic following customer feedback on its initial preview. The company says the updated architecture dramatically improved image extraction, reducing errors from 207 to zero in its internal benchmarks. Both models are available through Workers AI, Cloudflare’s platform for running AI models, and their full weights have been published on Hugging Face.
A new research identity
Cloudflare is also giving its frontier AI research operation a more explicit identity. The company has reintroduced the group responsible for these releases as the Workers AI superintelligence team, describing it as the research organization behind Clef, Clairic and additional projects under development.
The announcement signals that Cloudflare wants to develop its own frontier AI models rather than rely exclusively on models built by other organizations. Its stated audience includes Cloudflare customers and the broader open-source community. The team says it is moving quickly to advance AI research while producing models with practical applications.
This is not simply a branding exercise around an existing extraction product. The company says customer feedback prompted it to reconsider Clairic's underlying architecture after the original preview release. The redesigned version reflects a narrower emphasis: making extraction, rather than general-purpose conversation, the central task.
Why Clairic was redesigned
The original preview attracted a range of customer applications. Some developers wanted to retrieve information from photographs of receipts, while others needed to identify values contained in customer support tickets. These requests highlighted a common problem: useful information frequently sits inside unstructured material rather than clean databases or consistently formatted records.
Cloudflare concluded that the model needed to become particularly effective at reading and extracting information from these varied sources. Instead of making incremental changes to the preview, the team went back to its architecture and rebuilt important parts of the system.
The central technical change concerns how Clairic generates ordinary text and numerical fields. The previous approach relied on a single forward pass. The redesigned model uses parallel slot autoregressive generation, a method that allows separate fields to generate their answers while progressing together during serving.
Each field uses the tokens it has already generated to determine what comes next. It then stops independently when it reaches its designated ending marker. Although the fields generate their contents separately, they advance together within the serving batch.
Cloudflare says this approach improves extraction accuracy without sacrificing the responsiveness expected from a specialized model. There is a small additional latency cost, but the company considers the accuracy improvement worth that trade-off. Its tests indicate that the resulting system remains capable of returning responses in under one second.
Specialized processing improves accuracy
The architectural changes also address requests that involve more than copying information from a document. Cloudflare trained dedicated processing heads for computational and estimation tasks, including forecasts, durations, calculated dates and explicit arithmetic.
That distinction matters because extraction systems often need to do more than identify a sequence of characters. A developer might want a model to read several values, calculate a total or estimate a duration. Dedicated components allow Clairic to handle these requirements as part of its extraction-oriented design.
The company has retained important elements of the original typed API interface. Every request can still return confidence scores, character spans and probabilities. These features help developers understand not only what information the model extracted but also where it appeared and how confident the system is about its result.
Character spans identify the relevant portions of the original input, while confidence information provides additional signals for applications that need to validate extracted data. Keeping these capabilities intact means the architectural redesign does not require customers to abandon the interface features they already use.
What the evaluations show
Cloudflare reports improvements across several categories, including text, images, PDFs and mixed-input extraction. Its published evaluation material includes an internal test involving 36 requests, tracked customer assertions covering 139 request shapes, and synthetic receipt fields represented in both text and images.
The company also compared its models with other emerging extraction products, including Liquid AI's D1.2, used alongside Apple's Vision technology for optical character recognition, and Scaledown's Extract.
There is an important qualification: Cloudflare says no public extraction benchmarks are available yet. Its comparisons rely on company-reported evaluations and should therefore be treated as directional evidence rather than independent proof of a definitive performance ranking.
The tests are intended to illustrate how Clairic performs across generalized extraction workloads. Cloudflare also argues that the emergence of competing extraction products supports its original view that developers need more efficient ways to retrieve information from documents and other inputs.
The underlying problem is familiar to users of large language models (LLMs), the systems designed to process and generate human language. General-purpose models can perform extraction, but doing so may involve unnecessary expense, slower responses and inconsistent accuracy. Smaller models dedicated to specific tasks offer another approach.
PDFs become first-class inputs
One of the most practical additions is PDF support. The earlier Clairic preview demonstrated how the model could interpret a photograph of a receipt. The updated model extends that capability to PDF documents, including cases where relevant information appears across multiple pages.
A developer can request a supplier, individual line items and a total in a single operation. The model can examine as many as four images or PDF pages in total. This makes it possible to extract related information without treating every page as an entirely separate request.
The output can also preserve relationships between individual items. For example, extracted line items can be returned as rows containing a description, quantity and price together. That structure is useful for applications that need to transfer document contents into databases, accounting software or other business systems.
Cloudflare attributes the substantial improvement in image accuracy to the redesigned architecture. Its internal testing recorded a reduction from 207 errors to zero, with the evaluation explicitly covering text and images. The result is one of the clearest numerical claims in the announcement, although it remains an internal benchmark rather than an independently reproduced public result.
Two models, different prices
The release also changes how Cloudflare distributes the model sizes. During the preview, the company hosted the 9-billion-parameter version through the Clairic endpoint because it was faster and easier to train.
Under the updated arrangement, the 9B model moves to the Clairic-flash endpoint, while the larger 27-billion-parameter version becomes the standard Clairic model. Parameters are the numerical values that determine how a trained neural network processes information; larger parameter counts generally mean a larger model, although size alone does not establish performance.
The standard Clairic model will cost $0.24 per million input tokens. Clairic-flash, the smaller model, will cost $0.10 per million input tokens. Tokens are the pieces of text that AI systems process, and input-token pricing measures the amount of material supplied to a model.
Cloudflare says it will continue following the pricing approach used for its System One models. Although Clairic generates output tokens, customers will only pay for input tokens. The company will absorb the cost associated with generating the output.
Existing customers get an upgrade
Cloudflare is offering a temporary discount to make the transition easier. Through December 1, 2026, both Clairic and Clairic-flash will cost $0.10 per million input tokens. That gives developers an opportunity to test the larger model without paying its eventual standard rate.
Applications already connected to the original Clairic preview endpoint will automatically receive the updated extraction capabilities as Cloudflare rolls out the replacement in place. Developers do not need to change their existing integration simply to benefit from the redesigned model.
Customers whose workloads prioritize lower costs may instead choose to move to Clairic-flash. The company specifically recommends considering the smaller model for applications where the reduced price is more important than using the larger variant.
After the promotional period, standard Clairic pricing will return to $0.24 per million input tokens. The smaller model will remain at $0.10. Developers therefore have a defined period in which to compare the two versions before the price difference takes effect.
Open weights invite experimentation
Both models are also being published on Hugging Face, a platform widely used to distribute machine-learning models and supporting resources. This gives researchers and developers access to the model weights beyond Cloudflare's hosted Workers AI service.
Cloudflare says its earlier release of Clef generated considerable experimentation and resulted in tuned versions created by other teams. It hopes Clairic will produce a similar response, allowing researchers and developers to explore the architecture and build applications around it.
The releases fit into a broader strategy that Cloudflare describes through its System One models. Rather than sending every task to an expensive, general-purpose language model, developers can use smaller specialized systems to handle defined workloads.
The company sees extraction as a particularly suitable application. Reading receipts, retrieving values from tickets and processing multi-page documents are relatively specific tasks. A model designed around those requirements could reduce the amount of general-purpose AI computing needed for routine business operations.
Why it matters
Cloudflare's announcement brings together two developments: a more capable document-extraction product and a growing frontier AI research operation. The Workers AI superintelligence team is presenting its work as part of a larger effort to build proprietary models while making their weights available to the research and developer communities.
The technical results could matter to businesses that process large volumes of semi-structured information. Receipts, invoices, support tickets and PDF records often contain information that must be manually transferred into other systems. Better extraction could make those workflows faster and reduce the need to send every request through a large, expensive general-purpose model.
However, the evidence currently comes primarily from Cloudflare's own evaluations. The absence of public extraction benchmarks means independent comparisons will be important in assessing how well the models perform against competing systems under consistent conditions.
The immediate milestones are practical. Developers can test the two models through Workers AI, explore their open weights on Hugging Face and compare the larger and smaller variants during the promotional pricing period. Existing preview integrations will receive the replacement automatically.
The next test for Cloudflare's superintelligence team will be whether its internal accuracy gains translate into independently measurable improvements across real-world extraction workloads. The December pricing change, adoption of Clairic-flash and the response from the open-source community will provide early indications of how successfully the company can turn its frontier research ambitions into widely used specialized AI tools.