OpenAI Decisions API vs Jev AI: Two New Approaches to Machine-Native Decisions
OpenAI Decisions API and TypeSafe AI’s Jev both target structured, software-consumable decisions, but differ in model approach, input support, and ecosystem.
Key takeaway
OpenAI Decisions API provides predicate, choice, and score answers through gpt-6-luna and supports text plus images. Jev is TypeSafe AI’s System One model, designed around typed decisions and calibrated probabilities for automation workflows; its public comparison claims should be treated as vendor-reported benchmarks.
Updated on Oct 9, 2026
01
Abstract
A practical comparison of OpenAI’s Decisions API and TypeSafe AI’s Jev: two approaches to returning structured decisions for software workflows instead of relying on free-form text generation.
02
Introduction
OpenAI Decisions API and TypeSafe AI’s Jev both reflect a shift from asking AI to write a response toward asking AI to return a decision that software can consume. This is useful for classification, routing, prioritization, and other workflows where downstream code needs a constrained answer rather than a paragraph.
OpenAI’s Decisions API is a public-beta endpoint using gpt-6-luna. It evaluates shared text or text-and-image input against named questions and supports three question types: predicate (probability that a condition is true), choice (select from supplied options), and score (rate against ordered levels). OpenAI positions it as a faster alternative to using the Responses API for these decision-shaped tasks.
Jev is TypeSafe AI’s System One model, offered in early access. TypeSafe describes it as built specifically for structured decisions, with typed outputs and confidence/probability information. The company attributes its approach to a different model architecture, parallel sampling, and Reinforcement Learning for Calibrated Decisions (RLCD).
The key distinction is not simply “which model is smarter.” It is the product contract and system design: OpenAI offers a decision-focused API inside its broader model ecosystem and accepts images as well as text; Jev is purpose-built around machine-consumable decisions and TypeSafe’s workflow-oriented approach. Performance claims from either provider should be tested on the workload that matters to your application.
03
Methodology
How structured decisions drive incident response
The diagram progresses through triage, disposition, containment, and playbook selection. Model readings describe the alert and incident; code combines them with context and thresholds to close, queue, or act. Containment examines eleven incident properties before policy selects an applicable response or escalation. This illustrates TypeSafe’s approach: decompose judgments into probabilistic questions while software controls branching. Read TypeSafe’s explanation.
Source: TypeSafe AI, Introducing System One Models & Jev. Vendor-reported evidence; screenshot supplied by the site owner.
04
Results
Capability comparison
| Dimension | OpenAI Decisions API | TypeSafe AI Jev |
|---|---|---|
| Availability | Public beta, according to OpenAI’s documentation | Early access, according to TypeSafe |
| Model/API | Dedicated Decisions endpoint; documentation currently names gpt-6-luna | System One model exposed through TypeSafe’s API |
| Inputs | Text, or user messages containing text and images | State supplied as text, structured state, or arrays of state, according to the API docs |
| Decision types | Predicate, choice, and score | Yes/no, choice, and rating-style questions; confirm exact schema against the current API docs |
| Output contract | Named answers with type-specific fields such as probability, choice, score, and confidence/probability distributions | Typed decision values with confidence/probability information, as described by TypeSafe |
| Design emphasis | A decision endpoint integrated into OpenAI’s API and SDK ecosystem | A model and serving stack designed specifically for structured decisions and automation |
| Broader generation | Use the Responses API when you need generated prose or arbitrary schema-shaped objects; use function calling when the model must request tools | Jev explicitly gives up general string generation in exchange for a narrower typed-decision interface |
What the published evidence says
OpenAI documents Decisions as roughly 10× faster than using its Responses API for these decision-shaped tasks. TypeSafe reports 70–500 ms response times and much larger speed/cost gains on its own workflow evaluations. These are not directly comparable universal benchmarks: the providers describe different baselines, test harnesses, and workload assumptions. TypeSafe itself notes possible evaluation bias and says some headline workflow gains represent the high end of observed results.
The practical result is that both approaches can reduce the amount of free-form text your application must parse. Neither removes the need to validate business rules, set decision thresholds, measure false positives/negatives, and route uncertain or high-impact cases to human review.
Workflow accuracy versus cost
Higher and farther left is better; cost uses a logarithmic axis. Diamonds represent workflows, circles prompted alternatives. TypeSafe evaluates four fixed workflows against reference probabilities averaged from Astra and Fable, rather than independently labeled ground truth. Jev occupies the low-cost frontier shown. The comparison uses wrapped LLM calls and does not establish a direct Decisions API benchmark. TypeSafe acknowledges possible evaluation bias. Read TypeSafe’s explanation.
Source: TypeSafe AI, Introducing System One Models & Jev. Vendor-reported evidence; screenshot supplied by the site owner.
05
Limitations
Type safety versus decision correctness
Lower bars mean fewer errors. The left panel concerns structured outputs; the right concerns tool calls. TypeSafe says the comparison figures come from OpenRouter and may reflect differences in routed query difficulty. Jev’s plotted zero is a claimed schema guarantee, not an empirical error-rate estimate. Schema-valid outputs can still encode incorrect judgments; these charts do not establish zero factual or decision errors. Read TypeSafe’s explanation.
Source: TypeSafe AI, Introducing System One Models & Jev. Vendor-reported evidence; screenshot supplied by the site owner.
Continue reading