Choose a model
Compare System One models (Jev, Clef Flash, Clef Omni, Clairic), costs, and when to switch.
A model is the part of System One that evaluates your context and answers your questions. Different models can give different results for the same task. A higher price does not guarantee a better answer for your particular use case.
Answer-ready
Q: Which models does System One provide?
A: System One provides Jev, Clef, Clef Flash, Clef Omni, and Clairic for hosted decisions and extraction.
Q: When should I use Clairic vs Clef Flash?
A: Use Clairic when you need typed field extraction; use Clef Flash for fast Choice/Score/Noul decisions. See the Clairic SDK page for extract examples.
Find the current selection
Open the model directory to see the models currently offered. After signing in, the Playground lets you select a model and try it on an example or your own task.
The selection can grow and change. This guide explains how to choose; it is not a fixed list of every model. Before using a model, check its current information in the product:
| What to check | Why it matters |
|---|---|
| Model name and ID | The name helps you recognize it. The exact ID identifies it when connecting another tool. |
| Supported inputs and question types | Make sure it accepts the material and questions you want to submit. |
| Input limits | Check request size, number of questions, choices, scoring levels, and any required instructions. |
| Credit price | The model you select affects the request cost. Check the estimate before running. |
Only use capabilities explicitly offered by System One for that model. A feature described elsewhere does not necessarily mean it is available here. If the directory cannot be loaded, do not guess a model ID or price; try again later or contact support.
Images and per-model limits
Read capabilities.inputs for image and the corresponding limits.images (count, bytesPerImage, totalBytes). Model names, ID prefixes and SDK versions do not grant capabilities. Image-capable catalog entries such as Clef, Clef Flash and Clef Omni accept images; text-only entries such as Jev and Decider reject nonempty images. Clairic is an extraction model (typed values with confidence and p_null), complementary to those decision models; public model pages list it as live. It does not appear in GET /v1/models until it is open for API calls. Call it through System One with the public name clairic and SDK extract(). The hosted catalog can add other compatible models without changing application code. Image gateway ceilings are PNG/JPEG/WebP, four images, 4MB decoded each, 8MB decoded total and 13MB complete UTF-8 JSON; text-only entries remain bounded by 64KB. Catalog limits can be tighter.
In Playground, select an image-capable model before adding local image files. Review the attached images in their submitted order, remove unwanted files, and recheck the estimate: base64 image bytes contribute to credits. State remains required but can be an empty string. An empty image list is a text request; remote image URLs, audio and video are unsupported. Switching to a text-only model requires removing the images rather than silently dropping them. With decision images present, use the HTTP/curl example and explanation instead of an evaluate() SDK snippet. Clairic images use extract(). See the image request contract and Fetch example and privacy guidance.
Compare models on your task
- Choose a few representative examples, including one that is ambiguous or difficult.
- Write clear questions and candidate answers. See question types.
- Select a model in the Playground. Click Model information beside the selector to review its limits and rate, check the estimated credits, then run the request.
- Keep the context and questions the same, select another model, and review the new estimate before running again.
- Compare the answers with your own expected outcomes. Consider mistakes, uncertainty, elapsed time, and credits together.
Each run is a separate request and can consume credits. Results and response times can vary; one successful example is not evidence that a model will always be correct or fast. Start small before using a model for an important or repetitive task.
Switching models
After changing the selected model, check the input and estimate again. A model may require more choices, shorter question names, or instructions that another model did not need.
If an input is rejected, follow the displayed limits rather than repeatedly submitting it unchanged. If a model is no longer available, choose from the current directory and test the replacement before relying on its results. Do not assume different models are interchangeable.
Select a model in SDK or HTTP requests
Copy the exact model ID from the current directory into your SDK or HTTP request's model field. Do not substitute the display name, an old example ID, or a guessed abbreviation. Select the model explicitly when you want to control which model and price are used; leaving it out uses the service's current default.
Retrieve the current directory with public GET /v1/models. It does not require an API key and does not consume inference credits. Each entry can provide the model ID, capabilities, limits, and credit pricing. Browse the same enabled models at /en/models. See API reference for the response format and SDK documentation for client integration. A direct SDK connection to another service has that service's own model catalog; the hosted directory does not describe every SDK adapter.
Changing the model creates a different task for retry tracking: use a new idempotency key. If the previous request's outcome is unknown, check or recover it first rather than switching models to retry it blindly.