Home / Blog

· 2 min read · Citrusworks

Jev vs OpenAI's Decisions API: models that decide instead of write

In Industry news

Short answer: two products now sell decisions instead of text. TypeSafe AI opened early access to Jev on September 15, 2026. OpenAI put its Decisions API into public beta on October 6. Both take a question with a fixed set of answers and return a choice with a probability, at a fraction of the price of a generative call. For any workload that is really classification, test both against your own evals before you change anything.

What changed

Most production AI calls do not need prose. They need a yes or no, a category, or a score that another system can act on. Until now teams paid generative prices and waited for generated tokens to get one word back.

Decision models skip the writing. You send a typed question and get a typed answer.

Jev, from TypeSafe AI

  • Three output types: a choice among predefined options, a score against your rubric, and a yes or no with a probability.
  • Price: $0.042 per million input tokens, with no charge for output.
  • Speed: 70 to 500 milliseconds, according to the company.
  • The headline claim: about 194 times faster and 445 times cheaper than an LLM alternative, in TypeSafe's own evaluations. The company describes those figures as the top of the realistic range.

OpenAI's Decisions API

  • Announced at DevDay on September 29, 2026, with public beta on October 6.
  • Three output types: predicates, choices and scores. The same shape as Jev.
  • Runs on GPT-6 Luna only.
  • Price: $0.10 per million input tokens, input only.
  • Speed: up to 10 times faster than the same model through the standard API, according to OpenAI. Coverage notes the claim shipped without a published benchmark.

On list price per input token, that makes Jev about 2.4 times cheaper.

What nobody has shown yet

  • Calibration. A probability is only useful if 80 percent means right four times in five. TypeSafe says it trains for this. We found no published calibration data for the Decisions API, and no independent head-to-head.
  • Your labels. Vendor benchmarks use vendor tasks. Business-specific categories are where general models struggle most.
  • Manipulation. Documentation for one Jev integration warns that text inside the input can steer the answer. Keep your deterministic checks.

What to do this month

  1. Find the calls in your stack that return one word or one label. Those are the candidates.
  2. Run both products against the eval set you already have. If you do not have one, build that first.
  3. Compare pass rate, cost per decision and latency, then check whether the reported confidence matches how often each model is right.
  4. Keep the generative model for the reasoning steps and use a decision model for the branch points.

Our read

This is a price cut on classification with a cleaner interface, and it is real. It does not remove the need for evals. It makes them more valuable, because switching vendors is now a one-day test for anyone who has them.

Sources