Clairic vs Jev vs OpenAI Decisions vs Clef vs Microsoft-Decision-1: Which System One Model Fits?

The new class of "System One" models, fast single-pass models that answer without generating text, now has two branches. Decision models such as Jev, OpenAI's Decisions API, Cloudflare's Clef, Perplexity's pplx-decider, Liquid's d1 and Microsoft-Decision-1 pick from options you list. Extraction models, introduced by Cloudflare's Clairic, return typed values read, computed, estimated or forecast from the input. This post lines them up so you can tell which one fits which job, and how they combine.

Every price and benchmark below comes from the vendors or from our earlier write-ups of them, and we have not re-run any. They use different benchmarks and baselines, so none of the numbers are directly comparable across vendors. For background, see our explainer on what decision models are and our ranking of the top ten.

The key distinction: values versus choices

If you can list the possible answers in advance, you want a decision model. If you cannot, because the answer is an amount, date or free value, you want an extraction model.

Side-by-side comparison

Prices are as reported in our earlier posts and may have changed. The pplx-decider price was not on its model card, so confirm it on Perplexity's pricing page.

What each one is best at

Clairic is the only model here that returns arbitrary typed values. It gives text spans with character offsets, numbers with set decimals, dates and times, durations, lists, and distributions with a median and an 80 percent interval for estimates and forecasts. Each scalar answer carries p_null and confidence. Cloudflare says all fields are answered in one parallel pass, so thirty fields cost about the same as three, and there are no output tokens to bill. Its own demonstrations include a damaged-order email, a receipt photo, a home-price estimate and a 21-day TSA checkpoint forecast with a 3.3 percent average error. See our Clairic analysis.

Jev started the category and has the most field use. It chooses among up to 255 options or returns a score, and TypeSafe reports 70 to 500 ms end to end. See how Jev works and its launch.

OpenAI Decisions API is the procurement-friendly option. If you already buy from OpenAI, it adds a decision endpoint with zero data retention and HIPAA for eligible customers, per day-one reports. The case rests on convenience, not price. See the developer verdict versus Jev.

Clef pairs naturally with Clairic because both come from Cloudflare and run on Workers AI. Clef is open-weight with a Jev-compatible API, and Clef-flash targets the latency-sensitive hop. See the Clef launch.

pplx-decider is the strongest open option on Perplexity's own Decision Index and takes images, but needs a large GPU to self-host. See the pplx-decider write-up.

Microsoft-Decision-1 is a Qwen3.5-9B-based model scoring fixed options, priced like Jev, and available in Foundry. It is attractive if you are already on Azure. See the Microsoft-Decision-1 post.

Liquid d1 is the best fit for edge devices and non-text inputs, with open 3B and 600M models. See Liquid d1.

Where Jev and Clairic overlap, and where they do not

There is one genuine overlap: both answer in a single pass, charge for input only and report calibrated confidence. There are two important differences.

1. Arithmetic and dates. A summary of Jev's documentation says it is unreliable at counting, arithmetic and date comparison, which is the kind of work an extraction model is built for. Clairic computes sums and day counts, and in Cloudflare's example it added a price and tax and counted days between two dates. But Cloudflare is candid that longer chains and aggregation are hard for any one-pass model and recommends letting code do the math.
2. Open versus listed answers. Jev cannot return an invoice total, because the total is not on a list. Clairic cannot say which policy applies, because that is a choice among options. Trying to force one into the other's job produces buckets, which are coarse and add tokens, or a model that guesses.

How they combine

Cloudflare's own example is a three-stage pipeline that is easy to generalize:

Cloudflare reports 49 percent lower cost and about 3.2 times lower latency for this pipeline against a single LLM doing all three steps, with the decide step at 0.22 s and the extract step at 0.46 s. We noted that the baseline step times do not sum to the stated total, and that this is one pipeline from the vendor, not an independent benchmark. The sensible takeaway is the shape: remove the LLM from the steps that do not need prose.

Nothing requires the stages to come from one vendor. Clairic's typed output can feed Jev or any other decision model as the state, since the decision step takes text or JSON.

Which should you pick?

How to test before you commit

1. Collect real examples with verified answers. Two hundred or more from your own pipeline.
2. Separate the tasks. Test extraction fields and decision labels independently so you know which stage fails.
3. Score per type. Exact match for text and numbers, tolerance for dates and estimates, and log-loss or calibration plots for probabilities.
4. Check calibration. Confirm low-confidence answers are the wrong ones, and see what happens when the value is genuinely missing.
5. Measure latency as a distribution, not a single median. Field testers disagreed with vendor numbers for more than one model.
6. Compute cost per completed task at your volume. Clairic has no published price yet, so wait for the Workers AI pricing page.
7. Plan a fallback. Route null or low-confidence answers to an LLM or a human.

What we could not confirm

• Clairic's price and open-weights date.
• Any independent, general extraction benchmark. The evidence is worked examples and one vendor pipeline test.
• Head-to-head accuracy between Clairic and an LLM on your document types.
• Cross-vendor decision-model comparisons on a common benchmark.

Bottom line

Clairic and the decision models are complements, not substitutes. Use a decision model when the answer is a choice, an extraction model when the answer is a value, and an LLM when the answer is words. The practical advice is the same across all of them: pilot on your own labeled data, read the calibration, and keep a fallback for the cases the fast model cannot handle.

Related reading

• Cloudflare Clairic: claimed versus verified
• Top 10 decision models ranked
• What are decision models?
• Jev speed and cost claims fact-checked
• OpenAI Decisions API versus Jev: developer verdict
• Cloudflare Clef
• Microsoft-Decision-1

Prices, benchmarks and limits are as reported by vendors and in our earlier posts as of October 20, 2026. Nothing here is independently verified.