Decision APIs, Compared
A neutral directory of APIs that return a decision instead of prose. Three products below are comparable typed-decision APIs; a second group uses the word “decision” for a different job and is labeled accordingly.
Last updated: September 30, 2026
Typed-decision APIs
| Provider | What it returns | Pricing model | Best fit | Docs |
|---|---|---|---|---|
| OpenAI Decisions API | The chosen answer from a finite set of pre-defined answers | Not yet confirmed (limited preview) | Classification and routing inside an existing OpenAI stack | Docs |
| TypeSafe Jev | Choice, score, or yes/no values with calibrated probabilities | $0.042 / M input tokens; output free | Fast typed decisions through a hosted API | Docs |
| Laya | Choice, score, or yes/no from a single encoder pass | Open weights, self-hosted (no per-token API price) | Self-hosted or data-locality decision workloads | Docs |
OpenAI Decisions API
Announced at DevDay on September 29, 2026, in limited preview. It focuses Luna's intelligence on user-defined questions with finite pre-defined answers, taking text or image context and returning the chosen answer.[1] Pricing and latency figures were not published at the time of writing. Read the OpenAI Decisions API breakdown.
TypeSafe Jev
A purpose-built System One model, announced September 15, 2026.[3] It returns choice, score, or yes/no values with calibrated probabilities in a single pass.[2] TypeSafe reports 70–500 ms latency and $0.042 per million input tokens with output free.[5][4] See what Jev is and how it works.
Laya
An open-source, Apache 2.0 decision model from Convai Innovations, published days after Jev. It shares the same choice / score / yes/no interface but ships as open weights you self-host on your own hardware.[6][7] Convai reports about 39.5 ms per question on a T4 GPU; treat that as a vendor claim rather than an independent benchmark.[7]
Adjacent “decision” products
These use the word “decision” for a different job from the typed-decision APIs above and are included here so the comparison does not quietly conflate them.
- OmegaEngine — an agent-governance platform whose “Decision API” authorizes an agent's action with an approved / denied / escalated verdict, on a bring-your-own-keys model. It judges whether an action should run; it is not a text-classification decision model.[8]
- Wizerr — decision infrastructure for hardware engineering: cross-referencing electronic components and optimizing bills of materials. A different domain, not a text-decision API.[9]
- AI/ML API — an API gateway / aggregator that resells many models under one key (including Jev through gateways), rather than a decision API of its own.[10]
Names we could not verify
We investigated two further names that appear in the decision-API search space and could not verify them as comparable typed-decision APIs. OpenPM is an OpenAPI package registry for AI plugin manifests, not a decision model.[11] We found no verifiable decision-API product under the name Effect and have left it out rather than guess. For the point-by-point verdict, see the OpenAI Decisions API vs Jev vs Laya comparison.
Sources
- OpenAI — DevDay 2026 recap
- TypeSafe AI — System One (concepts documentation)
- TypeSafe AI — Introducing System One models and Jev
- TypeSafe AI — Models
- Eigent — What Is Jev? TypeSafe AI's System One Model, Explained
- Laya — Laya (GitHub repository)
- Convai Innovations — Laya (Hugging Face model card)
- OmegaEngine — Pricing & Limits
- Wizerr — Hardware Intelligence for Component & BOM Decisions
- AI/ML API — AI/ML API (documentation)
- OpenPM — OpenPM (OpenAPI package registry)
Last updated: September 30, 2026