Cloudflare is positioning itself as a neolab, building frontier AI models for its customers and the open-source community. Its latest release covers an updated Clairic alongside a new, smaller sibling, Clairic-flash. Both are hosted on Workers AI and ship with full open weights on Hugging Face.
The group responsible for them is now known as the Workers AI superintelligence team, the frontier AI research unit behind Clef, Clairic, and further work in progress.
Clairic revisited: a redesigned architecture
Feedback since the preview release showed customers using extraction for tasks such as pulling data from receipt images or reading values out of support tickets. With extraction as the deliberate focus, the team went back to the drawing board and reworked the architecture behind this version.
For ordinary text and number fields, the model now relies on parallel slot autoregressive generation rather than a single forward pass. Every field conditions its next token on the tokens it has already produced and halts at its own end marker, with all fields advancing together in the serving batch so per-request latency stays low. The change improved extraction accuracy for both text and images, and the accuracy gain outweighs the small latency cost. Dedicated heads were trained for computation, estimation, and extraction, covering estimates, forecasts, durations, calculated dates, and explicit arithmetic.
The typed API interface retains the elements that make answers usable: confidence scores, character spans, and probabilities remain part of every request. The model is also still fast and cheap. Public extraction benchmarks don't exist yet, so the team's evaluations are internal and self-reported, run against other emerging extraction models such as Liquid AI's D1.2 (pipelined with Apple Vision for optical character recognition) and Scaledown's Extract. The results should be read as directional.
Eval | Clairic-preview | Clairic | Clairic-flash | D1.2 + OCR | ScaleDown |
Fresh text fields, 96 requests | 295/336 | 335/336 | 328/336 | 304/336 | 259/336 |
Internal eval (mchen) - mixed-input fields, 36 requests | 207/226 | 221/226 | 213/226 | 116/226 | 156/226 |
Tracked customer assertions, 139 request shapes | 1,033/1,263 | 1,263/1,263 | 1,263/1,263 | 973/1,263 | 653/1,263 |
Synthetic receipt fields in text and images | 369/576 | 576/576 | 576/576 | 397/576 | 280/576 |
Latency and document coverage
Latency tests measured the full HTTP response from a self-hosted server with all request fields included. Despite the architectural redesign, responses come back in under one second, well ahead of typical LLM responses for comparable workloads.
Eval set | Requests per model | Clairic p50 / p95 | Clairic-flash p50 / p95 |
Fresh text fields | 96 | 358 / 736 ms | 155 / 328 ms |
Fresh image fields | 48 | 453 / 976 ms | 305 / 605 ms |
Commercial leases | 180 | 691 / 803 ms | 379 / 519 ms |
Loan pricing | 600 | 395 / 407 ms | 188 / 192 ms |
Traffic forecasting | 114 | 532 / 585 ms | 243 / 261 ms |
Expense audit, development | 60 | 632 / 742 ms | 393 / 512 ms |
PDF support is new to the family. A request can ask for supplier, line items, and total in one go even when those details span pages, and up to four images or PDF pages can be supplied in total. Line items can be returned as rows that keep description, quantity, and price together. On internal benchmarks explicitly covering text and images, image reading accuracy improved from 207 errors to 0.

Endpoints, sizes and promotional pricing
The preview endpoint at @cf/cloudflare/clairic hosted the 9B variant at $0.10 per M input tokens. That 9B model now moves to @cf/cloudflare/clairic-flash, while the 27B model takes the @cf/cloudflare/clairic endpoint at $0.24 per M input tokens. The 9B Clairic-flash stays at $0.10 per M input tokens. Output tokens are technically generated, but the System One pricing principles are unchanged: billing covers input tokens only, and output generation is absorbed by Cloudflare.
To make the swap seamless, a promotional discount runs until December 1, 2026, pricing both models at $0.10 per M input tokens so the higher-quality variant can be tried at the same rate. Applications already pointed at the preview Clairic endpoint automatically receive the improved extraction capabilities, since the new model is rolled out in place. Cost-sensitive workloads can shift to the cheaper 9B variant. After December 1, the promotional discount ends and Clairic returns to $0.24 per M input tokens.
Open weights and availability
Both models are open-weight on Hugging Face, continuing the support for teams and research that followed the Clef release. The aim behind these System One models is to let people augment expensive LLM usage with smaller, faster, cheaper models built for one specific task, and the open-weight release is intended to encourage others to experiment and build on the ideas.
Clairic and Clairic-flash are available on Workers AI as @cf/cloudflare/clairic and @cf/cloudflare/clairic-flash. The developer documentation holds the full API schemas, unchanged from the preview endpoint, and the open-weight releases are published on Hugging Face.



