> ## Content Index > Fetch the complete content index at: https://eazzytechnews.ghost.io/llms.txt > Use this file to discover other available public pages before exploring further. # What changes when OpenAI makes an AI choose from fixed answers? - URL: https://eazzytechnews.ghost.io/openai-decisions-api-preview/ - Published: 2026-10-02T00:12:59.000Z - Updated: 2026-10-02T00:12:59.000Z - Description: OpenAI is previewing a service that chooses from developer-defined answers. What that changes for applications, and what its speed chart leaves unanswered. - Author: Collins Anfo - Tags: OpenAI, Artificial Intelligence, Developer Tools, AI Agents OpenAI has introduced a limited-preview application programming interface (API) called Decisions API, letting developers ask an artificial intelligence (AI) model to choose from a fixed set of answers. An API lets programs communicate; here, the developer supplies a question, possible answers and text or image context for the model to assess. The [September 29 DevDay announcement](https://openai.com/index/devday-2026-recap/?ref=eazzytechnews.ghost.io) describes a service powered by Luna that assigns content to categories or directs requests to their next destination. It can also help an agent, software that selects steps and uses tools to pursue a task, choose its next action. Constraining the answer gives an application a defined set of outcomes to handle. The announcement establishes a preview, with wider release planned in the coming days. ## The developer defines the choices A free-form response can explain a situation in many different ways. Decisions API focuses the model on user-defined questions with a finite set of predefined answers. For a developer building a routing feature, that changes the design problem: decide which categories the application can handle, then ask the model to select among them. Consider a hypothetical support tool that offers the choices billing, technical support, account access and human review. The tool could send a customer's message as context and use the answer to pick a queue. This illustrates the announced routing capability; OpenAI's release does not report results for this example. Selecting an agent's next action leaves the application's execution, permission checks and handling of failures as separate concerns. The product description does not establish that a selected action was performed successfully. A fixed menu also leaves a design responsibility with the developer. If a request fits none of the listed categories, the menu needs a suitable way to handle it. A human-review option in the hypothetical example provides one such route. Membership in an allowed answer set alone cannot establish that the chosen answer is correct. [OpenAI's Developers account](https://x.com/OpenAIDevs/status/2105003318917697873?ref=eazzytechnews.ghost.io) identified the release as a limited preview on September 29\. [Neowin's September 30 report](https://www.neowin.net/news/openai-unveils-500-chatgpt-pro-plan-decisions-api-and-major-codex-upgrades-at-devday-2026/?ref=eazzytechnews.ghost.io) describes the same constrained-choice product. These accounts support the announced mechanism and preview status; they do not supply an independent accuracy evaluation. ## What the 150-millisecond chart establishes An OpenAI comparison [published by The Decoder](https://the-decoder.de/openai-baut-codex-und-api-zum-devday-mit-sicherheitsscans-decisions-api-und-ultrafast-aus/?ref=eazzytechnews.ghost.io) puts Decisions API at 150 milliseconds and GPT-6 Luna API at 1.6 seconds. The figure advertises 10x faster completion. Those are the vendor's published comparison values, with no independent measurement established by the report. ![OpenAI vendor chart: Decisions service 150 milliseconds versus GPT-6 Luna model service 1.6 seconds; claimed 10x faster task completion.](https://storage.ghost.io/c/09/53/09539035-af35-486d-b626-5b7c2dda5b1d/content/images/2026/10/decisions-latency-source.jpg) OpenAI's vendor comparison gives Decisions API, an application programming interface, a task-completion time of 150 milliseconds and the GPT-6 Luna model's interface 1.6 seconds. It advertises 10x faster completion without stating the sample, timing boundary or matched accuracy. Source: OpenAI, via The Decoder, September 29, 2026\. [Source report and original context](https://the-decoder.de/openai-baut-codex-und-api-zum-devday-mit-sicherheitsscans-decisions-api-und-ultrafast-aus/?ref=eazzytechnews.ghost.io). The horizontal axis says "Task completion time (s)", with s denoting seconds. A millisecond (ms) is one thousandth of a second, so 150 milliseconds is 0.15 seconds. The displayed labels are rounded and do not establish a precise or universal speed ratio. The vendor's 10x description should remain an attributed claim. The axis measures task completion, but the figure does not define when its clock starts and stops. It therefore cannot establish a general 150-millisecond round-trip response guarantee, a time to begin producing text, or the duration of a completed external action. Choosing a tool action and carrying it out are different steps; the chart does not document the latter's successful completion. Several missing comparison details affect how a developer could apply these numbers: - The figure gives no task count, selection method or range of workloads. It does not state the sizes of the text or image inputs, the output lengths, or whether both services received equivalent requests. - The timing boundary is unspecified. Whether the measurement includes network transfer, request handling, model computation or later application work is unknown. Hardware, server load and repeated-test conditions are also unstated. - No statistical spread accompanies the two bars. The figure does not say whether the values are averages or typical results, show a range of observed times, or report how often requests were much slower. It provides no measure of uncertainty around the estimates. - There is no matched-accuracy result alongside the times. The chart gives no comparable correctness, error-rate or decision-quality measurement, and cannot show that the faster service made equally good or equally safe choices. These omissions leave the conditions unknown; they do not prove that OpenAI failed to test them. The figure is useful as a statement of the company's speed claim. Evaluating that claim for a particular application would require representative tasks, a defined start and end for timing, and correctness measured alongside speed. A quick answer that sends a request to the wrong queue has a different practical value from a quick correct answer. ## Preview access leaves production questions open The DevDay recap says broad release was planned in the coming days. That is a forward-looking availability statement, with no firm calendar date in the Decisions panel. It does not establish current general availability, a price for this service or a published reliability guarantee. Before putting this selection step into a consequential workflow, a developer would need to establish how the service handles ambiguous requests, how often its choices are correct on that workload, and which actions the application permits after receiving an answer. Those are evaluation requirements, not results reported for the preview. OpenAI's next stated availability step is a broader release; its timing and the comparison conditions remain open. **Author:** [Collins Anfo, a digital product builder. My write-ups cover technology and artificial intelligence, from software, intelligent agents and robotics to chips, computing infrastructure, cybersecurity, consumer devices and space technology. I also cover technology-driven discoveries across mathematics, science and other research fields, product releases and announcements, and the business, finance, investment, governance and real-world adoption of technology.](https://collins-anfo-portfolio-2026.collinsanfo24.chatgpt.site/?ref=eazzytechnews.ghost.io) **AI assistance disclosure:** AI tools assisted with the research and drafting of this article. Material claims are linked to their sources for independent verification.