OpenAI opens the Decisions API to all customers, using GPT-6 Luna to cut decision responses to 100 milliseconds

Zhihui Dongxi October 23 news, according to a TechCrunch report yesterday, OpenAI has opened its Decisions API to all customers. This API is based on the reasoning model GPT-6 Luna and optimizes reasoning speed and cost; in some scenarios, response time can be shortened from 1 to 3 seconds to 100 to 180 milliseconds.

Decisions API is mainly aimed at tasks with a clear range of answers. Developers can predefine questions and optional answers, and the model outputs probabilities, options, or scores, for use in content classification, request routing, and deciding what action an AI agent should take next.

On September 15, AI startup TypeSafe launched the Jev decision model, which can quickly make judgments based on preset options and give confidence levels, helping developers complete classification and task routing at lower cost and faster speed.

OpenAI's opening of the Decisions API this time also means competition in the AI decision model field is further heating up.

1. OpenAI uses existing models to speed up and enter the fast decision market

Nikunj Handa, OpenAI's API product lead, said that after the team saw the Jev model, it realized this type of decision model has a large number of potential application scenarios.

The team then decided to adjust the reasoning approach based on the existing foundation reasoning model GPT-6 Luna to shorten response time and reduce cost, rather than training a dedicated decision model from scratch.

Handa said the most notable improvement of the Decisions API is speed, with response time for some tasks shortened from 1 to 3 seconds to 100 milliseconds, 120 milliseconds, or 180 milliseconds.

On September 29, OpenAI announced the Decisions API at its Developer Day event.

▲OpenAI announced the Decisions API at its September 29 Developer Day event. Image source: OpenAI.

On October 6, the API opened public testing. In a video introducing the Decisions API, Handa also mentioned that in the future OpenAI may conduct dedicated training for this type of task to further improve the model's performance; currently, related capabilities are mainly achieved by adjusting the existing Luna model.

▲OpenAI opened public testing of the Decisions API on October 6. Image source: OpenAI.

On October 22, OpenAI announced the official opening of the API to all customers. Decisions API supports receiving text, images, and structured data, and can process up to 400 questions per request, twice the limit during public testing.

▲OpenAI announced on October 22 that the Decisions API is officially open to all customers. Image source: OpenAI.

Enterprises such as enterprise document management platform Box, cloud monitoring platform Datadog, fintech company Ramp, and language learning platform Speak have used the Decisions API to replace some large model calls to improve task processing speed.

Cloud monitoring and security platform Datadog uses the Decisions API to assess the risk of commands generated by AI agents, with P95 latency about 40 times faster than available frontier models, meaning 95% of requests can complete a response within that time, and only 5% of requests take longer.

Enterprise financial management platform Ramp attributes AI call costs to specific work tasks, increasing processing speed by 4 times and reducing costs by 30%.

Language learning platform Speak shortened the waiting time between user actions in GPT-Live from 1.4 seconds to 0.13 seconds, a reduction of more than 90%.

▲OpenAI shows application cases of the Decisions API at enterprises such as Datadog, Ramp, and Speak. Image source: OpenAI.

Among them, enterprise document management platform Box showed on its official website its process of using the API to classify business documents and assign files to corresponding processing stages.

▲Box shows the application of the Decisions API in an insurance certificate review process. Image source: Box official blog

2. Two decision models compete on the same stage, with different rankings in third-party evaluations

OpenAI's launch of the Decisions API puts it in competition with TypeSafe AI's Jev model. Both products target software automation tasks and can quickly output decision results based on preset options.

TypeSafe AI was founded in 2024 and mainly develops decision models for software automation. Its co-founder and CEO Diogo Almeida previously worked as a researcher at OpenAI and participated in development related to improving language model instruction-following capabilities.

▲Photo of the TypeSafe AI co-founder team. Image source: TypeSafe

On September 15, TypeSafe released its first System One model, Jev. This model can directly output preset options and their corresponding probabilities for software calls. Internal test results published by TypeSafe show that Jev achieved 68% accuracy in 8 workflow tests, tied for 3rd with OpenAI's Tera model.

▲Ranking of Jev model decision-making capability. Image source: TypeSafe

On October 14, TypeSafe released Jev version 1.14.0, adding features such as image input and updating model capabilities. But Almeida mentioned that benchmark results should be treated with caution and suggested developers evaluate model performance in the context of their own workflows.

Rankings on third-party leaderboards show different results.

On the System One Index leaderboard published by AIMultiple, the latest Jev model ranks 7th, and OpenAI's Decisions The API ranks 14th; the list comprehensively evaluates model performance across different tasks and calculates a composite index based on the scores of various tests.

▲System One Index decision model rankings, image source: AIMultipl

In the Decision Index list released by Artificial Analysis, the Jev model ranks 2nd, and OpenAI's Decisions API ranks 11th. This list mainly evaluates model performance from four aspects: knowledge and hallucination control, reasoning, tool calling, and classification.

In addition to test scores, OpenRouter's model usage ranking also provides another reference: TypeSafe's products remain in a leading position on this list.

▲OpenRouter model usage ranking, image source: OpenRouter

Beyond comparing composite rankings on public lists, there are also tests that examine decision models' continuous judgment ability through actual gameplay.

AI infrastructure company Opper AI developed the JevMan benchmark, having decision models such as Jev play the classic game Pac-Man in real time to observe whether the models can continuously make judgments based on the game situation.

The benchmark has each model play 100 matches against the ghosts in the game and ranks them by average score, while also recording survival time, response latency, fallback action ratio, and per-match cost.

Opper AI Chief Growth Officer Felix Wunderlich said Pac-Man is suitable for testing decision models because the model must decide the next direction within a few hundred milliseconds at every intersection.

▲The JevMan benchmark evaluates decision models' continuous decision-making ability through the Pac-Man game, image source: Opper AI

According to Opper AI's JevMan list, Jev 1.14 has an average score of 3,597 points, ranking 1st; OpenAI's GPT-6 Luna has an average score of 2,568 points, ranking 8th.

▲JevMan benchmark model average score ranking, image source: Opper AI

Speaking about future development, Handa believes that OpenAI's models may have advantages in multimodal capabilities, especially in image classification tasks, where Luna demonstrates strong capability and confidence calibration. He also said that OpenAI customers hope the Decisions API can more flexibly use images as output in the future.

Handa also mentioned that faster response speeds may expand application scenarios for more decision interfaces. If interface response time can be further reduced to about 60 milliseconds, developers may be able to call the interface on every frame of a video to make judgments in real time.

As decision models' response speeds increase and usage costs decrease, developers are expected to apply them to more scenarios that were previously difficult to realize due to excessive latency or cost.

However, whether decision models can play a role in more business areas still depends on their accuracy, confidence calibration, and cost performance.

As OpenAI enters this market, developers will have more choices, and actual application results will also become an important criterion for testing the competitiveness of different models.

Source: TechCrunch