OpenAI Launches 'Decisions API' to Compete with Fast Decision Models

The AI landscape is undergoing a notable shift away from an exclusive focus on large language models toward a new, more specialized and faster category known as "Decision Models." In a move that underscores the seriousness of this shift, OpenAI has announced that its Decisions API is now available to all customers, after previously being limited to a handful of pilot partners. The step places the company in direct competition with emerging players who have built substantial market value on a simple premise: not every task requires a full language model.

What Are Decision Models and Why Have They Emerged?

Decision models can best be described as "super classifiers." They receive a set of defined options, then select the most suitable one in a fraction of a second, attaching a confidence score that reflects their level of certainty. This mode of operation—which some liken to fast, intuitive thinking, or what is often called "System 1"—differs fundamentally from language models, which generate open-ended text and consume far more time and resources.

Developers quickly discovered that a great many workflows within their applications don't actually require text generation at all, but rather a decision between alternatives: Is this email a phishing attempt? Which path should the user's request follow? Does this case warrant human intervention? Replacing a language model with a decision model at these junctures means a sharp reduction in cost and latency without any loss in output quality.

OpenAI's Approach: Retuning Rather Than Building from Scratch

According to Nikunj Handa, who oversees API products at OpenAI, the team did not build a model purpose-trained for decision-making; instead, it took an existing model and reconfigured it to accelerate inference and cut costs. Handa says the most significant gain was in latency: response times dropped from a range of one to three seconds down to between 100 and 180 milliseconds.

The general release wasn't limited to speed improvements. OpenAI also added new capabilities, including:

• Support for audio and visual inputs alongside text.
• Geographic distribution of inference resources to reduce latency across different regions of the world.
• Greater flexibility in handling images as inputs, with plans to support them as outputs as well.

Field Results from Early Trials

Early trials with companies in document management, technical monitoring, payments, and language learning pointed to tangible improvements in processing rates. One of the most striking examples came from an AI-powered language learning platform, which managed to cut the interval between actions from roughly 1.4 seconds to just 0.13 seconds—a difference that transforms automated conversation from a stilted experience into something approaching natural interaction. OpenAI also used the same model to improve the response speed of its own voice assistant.

Fierce Competition and Unsettled Benchmarks

The core problem in this emerging market is that the definition of a "good decision model" remains murky. The required balance between intelligence, cost, calibration accuracy, and speed varies from one use case to another, making any general ranking open to debate. For this reason, some founders in the space urge developers to ignore published benchmarks and test models within their actual production environments—even as those same companies continue to publish their own rankings.

Public benchmarks specialized in measuring "System 1" model performance currently tilt in favor of models recently released by competitors, while OpenAI's model fluctuates among the middle rankings depending on which index is used. On the other hand, the leading models in this category remain the most widely used in practice across request-routing platforms—a practical measure that may be more telling than theoretical tables.

Among the most amusing yet genuinely revealing tests is one that forces decision models to play the classic game Pac-Man. The test's developers justify their choice by noting that the game represents a continuous stream of rapid, high-stakes decisions: at every intersection, the model must determine its route within a few hundred milliseconds.

Where Is This Headed?

OpenAI's stated ambition is to reach a response time close to 60 milliseconds—a threshold that would open the door to invoking the model on every frame of a video, enabling full real-time visual analysis. If that is achieved alongside continued price declines, we will be witnessing a direct application of a well-known economic principle: the more efficiently a resource is used, the more widely it spreads and the broader its applications become.

The bottom line is that decision models are not a replacement for language models, but a complementary layer that intelligently redistributes the workload: open-ended tasks go to the large models, while repetitive, latency-sensitive decisions go to lightweight, fast ones. For companies building AI-powered products, auditing existing workflows for points that could be swapped out for a decision model may be the fastest route to cutting costs and improving the user experience at the same time.

✦ بقلم فريق دروب أيديا