OpenAI Decisions Model API Fully Launched, Entering the Fast Inference Market
nashnova research
OpenAI opened its Decisions API to all users on October 22, entering the fast decision-model market pioneered by startup TypeSafe — cheaper and faster than LLMs for classification and routing tasks, though its public benchmark rankings lag behind rivals.
What is a "decision model," and why does it matter now?
A decision model is an AI built for multiple-choice tasks — give it a set of options, it picks one fast and returns a confidence score. It handles classification, routing, and filtering, not open-ended generation like writing or chat.
This means → it skips the heavy, token-by-token reasoning of a large language model, running an order of magnitude faster at a fraction of the cost.
TypeSafe's Jev model defined the category. The company is now valued at $7.6 billion.
How did OpenAI build it — a new model from scratch?
No. OpenAI reconfigured its existing ChatGPT Luna model, accelerating inference to match the category's speed requirements rather than training a purpose-built model.
Product lead Nikunj Handa said the team saw Jev and realized there were "millions of use cases," then moved quickly to follow.
In plain terms = OpenAI chose to retune an existing engine rather than build a new one — faster to ship, lower investment, but the underlying model was not purpose-trained for decision tasks.
How much faster is it?
The core gain is latency: from 1–3 seconds down to 100–180 milliseconds.
Language-learning app Speak saw an even sharper drop — response gaps fell from 1.4 seconds to 0.13 seconds. OpenAI also used the API internally to speed up its Dots voice assistant.
Handa floated an even more aggressive target: at 60 milliseconds, the API could run a call on every frame of a video.
The benchmark rankings look weak — what does that signal?
On the AIM System One Index and the Artificial Analysis Decision Index, OpenAI's decision model ranks eighth and fourteenth, respectively. Liquid's D1 Ultra and Base Lab's Elephant-1 lead both lists.
This means → as a "reconfigured, not purpose-trained" product, Luna does not yet lead on specialized benchmarks — a direct consequence of the technical path OpenAI chose.
TypeSafe CEO Diogo Almeida himself distrusts benchmarks and advises users to test in their own workflows. This reflects a broader reality: the evaluation standards for this category have not settled.
What is OpenAI betting on to differentiate?
The general-availability release adds voice and vision modalities and distributes inference resources geographically to cut global latency.
Handa said Luna is "very strong" in image classification with "better calibration," and noted that customers want the API to support images as output.
In plain terms = OpenAI is betting on multimodality — while rivals handle text-only choices, it wants the model to see images and hear audio before deciding. Whether that edge can offset weaker benchmark scores is the next thing to watch in this market.
市场有风险，内容仅供研究参考，不构成投资建议。