OpenAI's decision endpoint, announced at DevDay 2026. It focuses GPT-6 Luna on developer-defined questions that have a finite set of predefined answers, with text or image context, for classification, routing and an agent's next action.
An embedding model turned decision model. Qwen3-Embedding-4B with a merged LoRA, driven by the open JevEmbed toolkit, which scores Choice, Score and Noul questions from embedding similarity and returns a distribution without generating text.
Decides
choice, classify, route
choice, score, noul, classify
Architecture
gpt-6-luna
jevembed
Fine-tuned from
—
qwen/qwen3-embedding-4b
License
proprietary
apache-2.0
Availability
Hosted API
Open weights
Hosted by
OpenAI
—
Input price
—
—
Decision accuracy
—
85.9%
Calibration error
—
—
Valid action rate
—
—
Median latency
—
—
p95 latency
—
—
Figures are from each model’s manifest; accuracy and latency are what the publishers report, on their own suites and hardware. Add a third model.
Questions
What is the difference between decisions-api and jevembed?
decisions-api is from OpenAI and jevembed from HIT-TMG (Lychee Team). decisions-api is only available as a hosted API; jevembed has open weights you can download and run. Both answer choice and classify questions. Only decisions-api answers route. Only jevembed answers score and noul. decisions-api is licensed proprietary; jevembed, apache-2.0.
Which is more accurate, decisions-api or jevembed?
Only jevembed publishes an accuracy figure (85.9% on JevEmbed-Data test split (64,110 hard-label questions), final LoRA checkpoint before merging); decisions-api does not, so there is no comparison to make without your own test.
Which is cheaper, decisions-api or jevembed?
decisions-api: Hosted, price not published. jevembed: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run decisions-api or jevembed locally?
jevembed yes — systemone pull hit-tmg/jevembed downloads its weights. The other is only served as a hosted API.
Evaluation suite
—
JevEmbed-Data test split (64,110 hard-label questions), final LoRA checkpoint before merging
