Jev vs OpenAI Decisions API
Unofficial. Differences only, from sources opened for this page. Not a ranking, and not affiliated with TypeSafe or OpenAI.
Documented differences between TypeSafe Jev and the OpenAI Decisions API: boundaries, uses, and access. No ranking.
This note lists differences first. It does not say which product to pick. If a field is not in a source opened for this page, it is marked not published.
As of October 15, 2026, jev-latest points at jev-1.13.0 and is the SDK default (Models). TypeSafe will move jev-latest to jev-1.14.0 around October 21, 2026 (Jev 1.14 release). See Jev 1.14: What’s New and How to Upgrade.
Differences
| TypeSafe Jev | OpenAI Decisions API | |
|---|---|---|
| What the docs call it | TypeSafe’s flagship System One model | A Decisions API that focuses Luna on user-defined questions |
| Availability in the source opened | Public docs for POST /v1/systemone; alias jev-latest points at jev-1.13.0 | “Available in limited preview today with a broad release planned in the coming days.” |
| Input the source describes | Text only: string, JSON object, or array of text. Images, audio, and video are not supported (yet) | “Developers supply context using text or images” |
| What you send | state plus a map of typed questions (noul, choice, score) | “A specific set of user-defined questions with finite pre-defined answers,” plus that context. Request schema: not published |
| What you get back | Structured answers (chosen option, score, or a 0–1 noul). Choice and Score also return probabilities and confidence | “Answers they can use to classify content, route requests, or choose an agent’s next action.” Response fields: not published |
| Separate confidence field | Choice and Score: yes. Noul: the API reference shows noul only, not a separate confidence | not published |
| Several questions in one call | Yes. Questions in one request are evaluated in parallel and in isolation against the same state | The recap says a set of questions. How they are packed into a request: not published |
| Text generation | System One docs: no replies, no code, no explanations of reasoning | not published |
| Speed wording | Use-case map: “real-time speeds (150ms)” for real-time applications. This page does not treat that phrase as a result against OpenAI | Recap: “real-time decision-making.” Graphic tagline: “Take action in near real-time.” A numeric latency: not published |
| Price | Models page, Jev 1.13: $42 per Btok / $0.042 per Mtok, charged per input token; output tokens free | not published |
| Rate limits | Models page, Jev 1.13: 100K tokens per second / 40 requests per second, and the page says those limits can change | not published |
| Context length | Models page, Jev 1.13: 64k tokens per request; 32k for state plus the longest question | not published |
| HTTP endpoint | POST https://api.typesafe.ai/v1/systemone | not published |
| Auth | Authorization: Bearer API key (TYPESAFE_API_KEY) | not published |
| Client SDKs | Python typesafe-sdk; JavaScript @typesafe-ai/sdk | not published |
| Try-without-code surface | Playground at console.typesafe.ai/playground | not published |
Capability boundaries
TypeSafe Jev
From the Introduction, System One, API reference, and Models pages:
- Jev evaluates typed questions against a state and returns structured values. It is not documented as a chat or code-writing model.
- Three question types can be mixed in one call: Choice (one option from a set you define, maximum 255 options), Score (an ordered rubric, at least two levels, API accepts up to 10), and Noul (yes/no as a number from 0 to 1).
- Choice returns
choice,probabilities, andconfidence. Score returnsscore,legend,probabilities, andconfidence. Noul returnsnoul. - The same weights serve every account. Jev is not fine-tuned or LoRA-adapted with customer data. Domain behavior is shaped in
state,instructions, andcriteria, then combined in your code. - English is the primary training language. Other languages, including CJK scripts, are handled but not equally; the models page says to test non-English content.
- Jev is not trained on customer requests or responses. The models page points to the Legal page for the DPA, privacy policy, and enterprise zero-data-retention details. This page did not open that Legal page, so those terms are not restated here.
OpenAI Decisions API
From the Decisions API section of the DevDay 2026 recap (opened 2026-10-01):
- It “enables real-time decision-making by focusing Luna’s intelligence on a specific set of user-defined questions with finite pre-defined answers.”
- Developers supply context using text or images and get back answers for classifying content, routing requests, or choosing an agent’s next action.
- It is in limited preview, with a broad release planned in the coming days. The recap does not name an endpoint, model snapshot id, schema, score, price, or limit.
- The section’s graphic says “Take action in near real-time.” It does not add a numeric latency or a field list.
The GPT-6 Luna model page was opened in the same pass and does not mention a Decisions API. These URLs returned 404: developers.openai.com/api/docs/guides/decisions, .../guides/decisions-api, developers.openai.com/api/reference/resources/decisions, and openai.com/index/introducing-the-decisions-api/. The API docs index that was searched had no Decisions API guide.
Not in the sources opened
For the Decisions API, the following are not published in the OpenAI pages opened for this note: endpoint path, request and response schema, confidence or probability fields, numeric latency, price, rate limits, context length, auth, SDKs, a playground, regions, a maximum number of answers, and whether the preview uses a specific Luna snapshot id.
Use cases
Published jobs overlap in wording only. Both sources describe classification and routing. That is not a ranking.
OpenAI’s recap names three uses for the answers: classify content, route requests, or choose an agent’s next action. No other Decisions API use case is on that page.
TypeSafe’s example use-case map groups examples under these headings: AI automation software, real-time applications, AI map-reduce over big data, universal verification, harness engineering, search and retrieval, scientific discovery, model routing, LLM guardrails, semantic code linting, feature extraction for predictive modeling, recruiting, lead generation, customer support, insurance claims, financial crime, legal and compliance, e-commerce marketplaces, moderation and trust and safety, advertising, gaming, risk assessment, demand forecasting, and graphs and knowledge graphs. The same page lists decision shapes: classification, detection, scoring, routing, search, retrieval, ranking, verification, ML feature extraction, and structured data extraction. The coding-agents page says to reach for Jev to route among a fixed set of destinations, score a rubric, check whether a statement is true of a document or record, or replace a prompt that asks a chat model to return JSON.
Entry channels
Jev, as documented:
- Playground: console.typesafe.ai/playground, after login. Paste text as state and add questions.
- HTTP:
POST https://api.typesafe.ai/v1/systemonewith a Bearer key from the keys dashboard.GET /v1/modelslists names the account can send. - Python SDK:
typesafe-sdk(Python 3.10+). The client readsTYPESAFE_API_KEYand defaults tojev-latest. - JavaScript SDK:
@typesafe-ai/sdk(Node.js 20+), same environment variable. - Agent skill: a drop-in skill for Claude Code, Codex, and other coding agents so they can write TypeSafe calls. The coding-agents page says this does not turn the coding agent’s own chat model into Jev.
Decisions API, as documented: limited preview, with broader release planned. How to request access, which host to call, which SDK to install, and whether a console exists are not published on the pages opened here.
Related on this site
Sources
Opened before the claims above were written:
