SHARED FORMAT
What carries over
The state + questions structure and question types noul/choice/score match the published Jev schema; probability answers follow the same documented shape for single-question requests.
Jev alternatives · comparison
A Jev-compatible decision API on open models
Decision Models accepts the documented Jev request shape — state plus typed questions — matching TypeSafe’s published OpenAPI schema, checked 28 September 2026. Migration is not drop-in: the launch differences below apply. Jev is a trademark of TypeSafe AI and Decision Models is not affiliated with TypeSafe.
Get an API key
Sign in with Google, create an API key and send text decisions right away; free allowance is available by invitation. Top up from EUR 1. Private individuals and businesses worldwide can sign up and top up. Taxes are shown in checkout before you pay. Image decisions use s1-vision (one image per request). Launch rates are listed on the pricing page.
At a glance
Decision Models vs the published Jev list
Dated, sourced figures — last checked 27 September 2026.
| Decision Models | TypeSafe Jev (published list) | |
|---|---|---|
| Input tokens | $0.032–$0.040 / 1M, by model and tier (text requests; s1-vision requests with an image use the image rate on the pricing page) | $0.042 / 1M list price |
| Output tokens | Free — decisions return no generated prose | Free per published list |
| Request format | state + up to three typed text questions for s1-pro, or one for s1-fast and s1-vision (noul, choice, score) per request | state + typed questions |
| Models | Open-weight decision models: Plumb-4B (s1-fast) and Winnow-12B Q8 for text. s1-pro and s1-vision run on Surogate Rune 26B-A4B v3; Winnow-12B Q8 is the fallback for s1-pro (text). s1-vision has no fallback on the EU tier or on Peer-to-Peer without worldwide opt-in: if Rune is unavailable, requests get a retryable 503 model_unavailable. Peer-to-Peer keys with worldwide processing opt-in may be served by on-demand worldwide GPUs running the Gemma 4 12B image implementation when capacity is busy. Peer-to-Peer-tier requests may use Rune when spare capacity is available. s1-fast runs on a batched engine that processes concurrent requests together, so probabilities for identical requests can differ slightly. | Proprietary Jev model family |
| EU processing | EU tier: contracted EU operators only, inference in the EU; DPA at signup | See TypeSafe documentation |
| Benchmark evidence | Compare model implementations across five equally presented sources, including weaker results and missing coverage. See the full benchmark matrix and method → | |
| Access | Sign in with Google, free allowance by invitation — get an API key | TypeSafe account signup |
Sources: TypeSafe Jev list price, checked 27 Sep 2026; Five-source matrix, methods and conflict disclosure; checked 30 September 2026; approved Decision Models launch rates.
The format
Keep questions structured
At launch, each request carries a state and up to three named text questions for s1-pro, or one for s1-fast and s1-vision, of type choice, yes/no or score. Answers include a selected value and, where supported, probabilities.
Request shape
{"state":"Mia owns a red bicycle.","questions":{"color":{"type":"choice","instructions":"Which color is the bicycle?","criteria":{"red":null,"blue":null}}}}Comparison
What changes when you migrate?
The shared request shape matches TypeSafe’s published OpenAPI request and answer schema for the documented decision types (checked 28 September 2026); the surrounding service contract is Decision Models-specific.
SEPARATE SERVICE
What differs
Up to three text questions for s1-pro; one for s1-fast and s1-vision; score rubrics need at least two levels (the published Jev schema also accepts one). Decision Models model IDs, bearer keys, the S1-Region tier header, idempotency behavior, billing and limits. Byte-for-byte compatibility is not claimed beyond the documented request shape.
Jev and TypeSafe are trademarks of TypeSafe AI. Decision Models is not affiliated with TypeSafe AI, does not resell Jev, and references the marks only to describe compatibility (nominative use). Read the full migration notes →