What is Clairic and Its Position in the AI Ecosystem
Cloudflare has launched a preview of Clairic, an extraction model classified as 'System One'. Unlike large language models (LLMs), which generate text token by token and require deliberate reasoning, Clairic is designed to respond in a single pass with typed values directly extracted, calculated, or estimated from the input.
The model integrates into a three-layer architecture proposed by Cloudflare: extraction models (like Clairic) convert raw input into typed values; decision models (like Clef) transform those values into decisions; and LLMs (like GLM-5.3) draft the final response. This separation allows each component to perform the task it is optimized for, reducing the cognitive and computational load on generalist models.
System One vs. System Two Architecture Comparison
Key differences between direct extraction and text generation.
| Feature | Clairic (Extraction) | LLM (Generation) |
|---|---|---|
| Output | Typed values (numbers, dates, text) | Free text (requires parsing) |
| Process | Single parallel pass | Sequential token generation |
| Output Cost | None | Per generated token |
| Hallucination Risk | Limited by data type | High (can invent data) |
Technical Architecture: Single-Pass Extraction
The core mechanic of Clairic is the elimination of intermediate text generation. The developer sends a 'state' (text and images) and a schema of typed fields. The model processes this information in a single inference pass and returns a valid value for each field, or 'null' if the data is not present in the input.
Each scalar response includes a confidence score and 'p_null', the probability that the value is not determined by the input. For text fields, the model returns character offsets (spans) that allow tracing the original fragment in the document. This traceability is crucial for audits and corrections, as it eliminates the ambiguity of knowing where an extracted data point originated.
Cost and Performance Analysis vs. LLMs
Clairic's price is $0.10 per million input tokens, with no charge for output tokens. This cost structure is significantly lower than that of LLMs, which charge for both input and generated output. In Cloudflare's reference test, a damaged product claims pipeline combining Clairic, Clef, and GLM-5.3 cost $6,032 per million claims, compared to $11,740 for a pipeline based exclusively on GLM-5.3.
In terms of latency, extraction with Clairic took 0.46 seconds, while GLM-5.3 took 16.1 seconds for the same task. The complete pipeline with Clairic finished in 19.3 seconds, 69% faster than the 61.7 seconds of the LLM-only workflow. It is important to note that most of the remaining time (18.6 seconds) was dedicated to drafting the response by the LLM, not to extraction.
Performance Metrics in Claims Pipeline
Direct comparison between the hybrid workflow (Clairic + Clef + LLM) and the 100% LLM workflow.
| Metric | Hybrid Workflow | LLM Workflow (GLM-5.3) |
|---|---|---|
| Total Latency | 19.3 s | 61.7 s |
| Extraction Time | 0.46 s | 16.1 s |
| Cost per Million | $6,032 | $11,740 |
| Cost Reduction | 49% | - |
Optimal Use Cases and Known Limitations
Clairic excels in structured data extraction tasks from semi-structured documents, such as invoices, support emails, claim forms, and incident reports. Its ability to handle up to four images per request makes it useful for processing photographs of receipts or scanned documents with low quality.
However, Cloudflare warns that the current model is an early checkpoint. Arithmetic chains of more than two steps (such as summing multiple rows of a table) are not reliable, and it is recommended to perform those operations in code. Additionally, estimates require evidence in the input (such as comparable prices); without it, confidence intervals can be too wide. Time-series forecasting is basic and does not replace specialized models trained on specific data.
Implications for Developers and Adoption
For developers, Clairic offers a path to reduce the complexity of AI pipelines. By separating extraction from generation, it eliminates the JSON validators and retries that are usually necessary when using an LLM for structured tasks. Integration is direct through Workers AI via env.AI.run() or the REST API.
The availability of open weights is planned for the future, which will allow organizations to fine-tune or run the model locally. For now, adoption should be gradual, starting with low-risk extraction tasks and closely monitoring confidence scores. The open question is how the model will perform with non-standard document formats or in languages other than English, a critical point for adoption in markets like Latin America.
Recommended Integration Process
Steps to implement Clairic in an existing pipeline.
- 01Identify repetitive extraction tasks in the current workflow.
- 02Define the schema of typed fields (text, number, date, etc.).
- 03Implement the call to Clairic and capture confidence scores.
- 04Establish confidence thresholds to route doubtful cases to review.
- 05Perform complex arithmetic operations in code, not in the model.
Open questions
- No independent standard accuracy benchmark has been published; metrics come from Cloudflare's internal tests.
- The model's behavior with documents in languages other than English (Spanish, Portuguese) is not documented in primary sources.
- The exact date for the release of the model's open weights has not been specified by Cloudflare.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction