Decision Models
A decision model returns a typed answer, such as a choice from a fixed set, a probability or a score, instead of text. It replaces the step where code asks a chat model a question and parses the reply.
The shift
Until September 2026, code that needed a small judgment called a chat model. The prompt listed the allowed answers, the code parsed the reply, retried when it came back malformed, and set temperature to zero to keep it stable. Every answer cost output tokens and a full generation's latency.
A decision model takes the question and the allowed answers and returns one of them with a calibrated probability in a single pass. There is nothing to parse and little or no output to pay for. TypeSafe calls these "System One" models: fast judgments that sit inside ordinary code, where a chat model does the slower thinking. TypeSafe
Who ships them
- TypeSafe Jev, early access from 2026-09-15. On Vercel's AI Gateway it reached about 13% of teams on its first day, twice the GPT-5.6 family and six times Fable 5.1. Source: X/@vercel, 2026-09-18
- OpenAI Decisions API, beta on 2026-10-06 and generally available on 2026-10-22. It runs GPT-6 Luna on bounded questions and returns predicates, choices from a fixed set or rubric scores, up to 400 questions per request. OpenAI says it answers up to 10 times faster than the Responses API; Box reported about 5 times faster routing, and Sierra a p95 under 280 ms for choosing an agent's next action. Source: OpenAI, "Build more responsive apps with the Decisions API"; Source: OpenAI API changelog, checked 2026-11-03
- Liquid AI d1, in hosted and open-weight forms:
- The hosted d1 decision model (
d1on Liquid's API) launched around 2026-09-29. Its weights are not public. It reached Vercel's AI Gateway asliquid/d1by 2026-10-09. Source: Datanorth, 2026-09-30; Source: Liquid AI, "d1", dated 2026-10-05; Source: Vercel changelog, 2026-10-09 - The d1-3B and d1-omni-600M open weights followed on 2026-10-07. Source: Hugging Face, LiquidAI "Open d1", 2026-10-07; Source: Hugging Face LiquidAI/d1-3B; Source: Hugging Face LiquidAI/d1-omni-600M
- d1-2B and d1-Omni-2B came with LFM3 on 2026-10-27, for decisions on the device. Source: X/@liquidai, 2026-10-27
- The hosted d1 decision model (
- Laya, open weights for typed decisions on Apple Silicon. TypeSafe has the details for Laya and LFM3.
When to use one
Use a decision model when the answer set is known before the question is asked:
- classifying an input or a document
- routing a request to a handler, a model or a person
- choosing an agent's next action from a fixed list
- scoring against a rubric or gating a risky command
- matching, deduplicating and ranking
Keep a chat model for anything that must produce text, code or an open-ended answer. In an agent, the loop runs on a chat model and the small decisions inside it can move to a decision model. When the output is structured data rather than a choice, private page are still the tool.
Which skill runs it
The typesafe-ai skill builds features on TypeSafe's models. Read the vendor's latest docs before choosing, because pricing and speed figures so far come mostly from the vendors.
Timeline
- 2026-11-03 | Page created as the owner of the decision-model paradigm: Jev, OpenAI's Decisions API, Liquid's d1 models and Laya, with when to use one instead of a chat model. Source: TypeSafe and OpenAI docs; X bookmarks