--- title: "Clairic vs Jev and Clef: Extraction vs Decision Models, and How to Chain Them" description: "Clairic pulls typed values out of text and images; Jev and Clef decide among options you list. One table of the differences, how billing differs, and the extract, decide, then reply pattern." canonical: "https://system-one.dev/en/blog/clairic-vs-jev-clef" language: "en" --- # Clairic vs Jev and Clef: Extraction vs Decision Models, and How to Chain Them **What's the difference between Clairic, Jev and Clef?** Clairic is an extraction model that answers "what's the value?". From text or JSON, plus up to 4 images, it pulls typed values such as text, numbers, dates, durations, lists, estimates and forecasts. Most values carry a confidence and p_null; null is returned when p_null exceeds null_threshold, and guess may still hold a value. Jev and Clef are decision models that answer "which one?": they run Choice, Score and Noul questions over options you list and return probabilities. **Which should you use?** Use Clairic to get values and Jev or Clef to make judgments. Many workflows need both: extract with Clairic, decide with Jev or Clef, and hand off to an LLM when a reply needs writing. ## The differences in one table This table compares interfaces, inputs and outputs only, not speed or accuracy. | | Clairic | Jev | Clef Flash / Clef | Clef Omni | Decider | |---|---|---|---|---|---| | Model type | Extraction | Decision | Decision | Decision | Decision | | Output | Typed values, most with a confidence and p_null; estimate and forecast have neither | Choice, Score or Noul with probabilities | Same | Same | Same | | Input | Text, JSON, plus up to 4 images | Text, JSON | Text, JSON, images | Text, JSON, images | Text, JSON | | When the answer isn't there | null, when p_null exceeds null_threshold | Probabilities across your options | Same | Same | Same | ## When to use Clairic When the answer is a value you can't list in advance: - The amount, invoice date and invoice number on an invoice. - The order number, quantity and delivery date in an email. - A duration in a description, or an interval estimate: a median plus p10–p90. Clairic has a 65,536-token context window. The official blog lists these preview limitations: one- and two-step calculations are reliable, but long chains of calculations and multi-row aggregation are not there yet; multi-step forecasts drift toward the average; and well-formed output isn't necessarily correct. ## When to use Jev or Clef When the answer can be written down in advance as options, a scale, or yes/no: - Which team should get this email? (Choice) - How risky is this reply? (Score) - Should this go to a human? (Noul) Choosing between Jev and Clef comes down to input type and credit rate: | Model | Input types | Credits per block | |---|---|---| | Jev | Text, JSON | 1 | | Decider | Text, JSON | 1 | | Clef Flash | Text, JSON, images | 3 | | Clef Omni | Text, JSON, images | 4 | | Clef | Text, JSON, images | 6 | | Clairic | Text, JSON, plus up to 4 images | 2, per request, not multiplied by questions | How billing works: > Each 4KB counts as one block, and anything under 4KB still counts as one block. Jev and Clef: blocks × questions × rate. Clairic: blocks × 2, per request. Source: [system-one.dev pricing](https://system-one.dev/en/pricing) · 2026-10-20 ## Chaining: extract, decide, then reply The recommended pattern is Clairic to extract, Clef to decide, and an LLM to write the reply. Take a payment email with an invoice screenshot attached: 1. **Extract**: Clairic reads the email and screenshot and pulls the amount, invoice date and invoice number, usually with a confidence and p_null. 2. **Check**: fields with a high p_null or a low confidence go straight to a person and don't move on. 3. **Decide**: pass the extracted values to Jev or Clef as JSON and ask "which team gets this?" (Choice) and "does this need human review?" (Noul). 4. **Reply**: when the customer needs an answer, let an LLM write it. Each step does one job, and you can tune each threshold separately. ### Example from the official blog: handling a damage claim in three steps A customer emails to complain about a damaged product: 1. **Extract**: Clairic pulls the order number, total paid, days from delivery to the email and the list of damage, and uses an estimate field to give a range for the repair cost. 2. **Decide**: Clef checks these values against the policy and chooses between refund and repair. The example policy: refunds within 30 days of delivery; repair is preferred when it costs less than 60% of the price and the damage doesn't stop the product from working, otherwise refund. 3. **Reply**: an LLM writes the reply from the decision and the values it needs. Step 2 sees structured values, not the raw email, so it is less easily swayed by the wording of the email. ## Using Clairic on system-one.dev Clairic is available on system-one.dev. For the model ID, see `GET /v1/models`. How Clairic is billed: each 4KB of request content counts as one block, and anything under 4KB counts as one block; each block costs 2 credits. You're charged once per request, no matter how many fields you extract. This differs from Jev and Clef, whose cost is also multiplied by the number of questions. See [pricing](https://system-one.dev/en/pricing) and the [Clairic model page](https://system-one.dev/en/models/clairic). ### Calling Clairic and Clef Omni through the System One API Clairic and Clef Omni are both models system-one.dev provides through the same System One API and the same SDK client. Use `extract({ state, fields })` with model `clairic` to get values, then send those values as state to a decision model. Clef Omni, like Clef and Clef Flash, can also read images when it decides. See the [Clairic SDK docs](https://docs.system-one.dev/en/docs/sdk/clairic) and the [Clef Omni model page](/en/models/clef-omni). New accounts get 100 free credits on sign-up. [Try it online (sign-in required)](https://system-one.dev/en/demo/invoice-extraction). ## Related reading - [What is Clairic? Extraction models explained](https://system-one.dev/en/blog/what-is-clairic) - [Clairic model page](https://system-one.dev/en/models/clairic) - [What is a decision model?](https://system-one.dev/en/blog/what-is-a-decision-model) - [system-one.dev pricing](https://system-one.dev/en/pricing) - [Screenshot to JSON for AI agents](https://system-one.dev/en/use-cases/screenshot-to-json) - [Invoice & receipt extraction API](https://system-one.dev/en/use-cases/invoice-receipt-extraction) Last verified: 2026-10-22 ## Frequently asked questions ### Can I use Clef alone for extraction? Not really. Clef only answers within the options you list. Amounts, dates and reference numbers can't be listed in advance, so they belong with an extraction model. ### Can Clairic do classification? Clairic returns values, not probabilities over options. To judge among fixed options, use [Jev](https://system-one.dev/en/models/jev) or [Clef](https://system-one.dev/en/models/clef). ### If I use both, isn't that redundant? No. Clairic handles "what's the value?" and Jev or Clef handles "what do we do about it?". Each owns one step. ### How do Clef Flash and Clef Omni work with Clairic? The same way as Clef: pass Clairic's output to them as JSON for the decision. If there are images, Clef Flash, Clef Omni and Clef can all read them. Test on your own samples to pick one. ### Which is more accurate, Clairic or Jev? They don't do the same job, so they can't be compared directly. Test each on its own step with your own data. ### Does it work with languages other than English? Jev handles Chinese, but it is less accurate than in English. For Clairic, test with your own data in your language first.