OpenAI has made its Decisions API available to all customers, adding voice and visual inputs for applications that need fast, high-throughput AI decisions. Early users reported shorter response times, but developers should test latency, decision quality, and cost with their own workloads.
Key takeaways
- OpenAI opened the Decisions API to all customers on October 22 after announcing it at its September 29 developer day.
- The API accepts voice and visual inputs, extending beyond text-based requests.
- OpenAI positions the API for fast, cost-effective decisions at higher request volumes, rather than as a universal replacement for larger language models.
- Speak CTO Andrew Hsu reported a decrease in the time between actions from 1.4 seconds to 0.13 seconds.
- Reported performance is not guaranteed; developers should test representative tasks, network conditions, decision quality, latency, and cost.
OpenAI has made its Decisions API available to all customers, adding voice and visual inputs to a product designed for fast, high-throughput AI decisions. The company announced the API at its September 29 developer day and opened access on October 22, giving developers a new option for applications where response time matters as much as the quality of an answer.
The release puts OpenAI into a growing market for lower-latency AI models: systems intended to make quick decisions without the cost or delay of routing every task through a larger, general-purpose language model.
What OpenAI Announced and What the API Adds
The public version of the OpenAI Decisions API accepts voice and visual inputs, expanding how developers can pass information into the model. That makes it relevant to applications that need to interpret spoken requests or visual context as part of a rapid response, rather than relying only on text prompts.
OpenAI describes the API as fast, cost-effective intelligence for higher-throughput decisions. In practice, that could mean using it for frequent, bounded tasks that previously depended on the company’s larger language models. The intended role is not necessarily to replace a general-purpose model in every application, but to give developers another option when latency, volume, or operating cost is a constraint.
The company’s announcement also includes a practical signal about how it expects the API to perform across markets: OpenAI said it targeted inference resources geographically to reduce latency in regions around the world. That matters for applications whose users are distributed internationally, since a model’s responsiveness can depend not only on its design but also on where requests are processed.
How Faster Responses Could Support Real-Time Decisions
The early performance figures shared by customers and OpenAI point to the API’s focus on speed. Speak CTO Andrew Hsu said it reduced the time between actions in the company’s product from 1.4 seconds to 0.13 seconds. OpenAI product lead Nikunj Handa described response times falling from one to three seconds to roughly 100-180 milliseconds.
Those changes could make an interaction feel more immediate, particularly in voice applications, where pauses are noticeable. A system that can interpret an input and respond quickly may also support workflows involving repeated decisions, such as selecting an action during a live interaction or processing frequent updates.
The figures are reported examples, not a guarantee that every application will see the same improvement. Actual latency depends on the task, implementation, network conditions, and how the API is combined with other systems. Developers evaluating the OpenAI Decisions API should test it against their own workload, including the quality of its decisions as well as speed and cost.
Early Customers and Competition in Fast AI Models
Box, Datadog, Ramp, and Speak used the API in initial trials. OpenAI also applied it to improve the responsiveness of Dots, its voice assistant. Together, the examples span enterprise software, financial services, and voice interaction, suggesting the API is intended for a range of products rather than a single narrow use case.
The launch also reflects increasing competition around fast, inexpensive AI intelligence. TechCrunch reported that TypeSafe’s Jev model is competing in the same category and that TypeSafe reached a $7.6 billion valuation. The broader contest is about more than benchmark performance: developers need models that can handle high request volumes, fit into real-time products, and provide useful results at a sustainable cost.
For developers, the key question is whether the API’s speed and support for voice and visual inputs match a specific application’s needs. Reviewing OpenAI’s API documentation and testing with representative inputs can help determine whether it is a fit, especially for products where a delay of even a second changes the user experience.
By the numbers
Speak reported that the time between actions decreased from 1.4 seconds to 0.13 seconds.
Speak CTO Andrew Hsu shared this customer example in connection with the API. It is a reported result from Speak's product, not a general performance guarantee.
OpenAI product lead Nikunj Handa described response times falling from 1-3 seconds to roughly 100-180 milliseconds.
This is an example shared by Handa about the API's response times. Results for other applications may vary based on task, implementation, and network conditions.
TypeSafe reached a $7.6 billion valuation.
TechCrunch reported this valuation while describing TypeSafe's Jev model as a competitor in the market for fast AI models.
Step by step
- 01
Identify latency-sensitive tasks
Choose bounded, frequent decisions where response time, request volume, or operating cost is a constraint, such as voice interactions or repeated workflow decisions.
- 02
Prepare representative inputs
Create realistic voice, visual, and other relevant examples that reflect the conditions and variety of inputs your application will handle.
- 03
Integrate the API into a test workflow
Follow OpenAI's API documentation and connect the Decisions API to the relevant parts of your application without assuming it must replace a general-purpose model.
- 04
Measure latency, quality, and cost
Test the same representative workload across your implementation and compare response times, decision quality, and operating costs under realistic network conditions.
- 05
Decide where the API fits
Use the results to determine which tasks benefit from the API and whether other tasks should continue using a larger model or a different approach.
Frequently asked questions
What is the OpenAI Decisions API?
The OpenAI Decisions API is an API designed for fast, cost-effective AI decisions at higher request volumes. Its public version accepts voice and visual inputs in addition to text-based information. OpenAI presents it as another option for latency-sensitive tasks, not necessarily as a replacement for larger general-purpose models.
What inputs does the OpenAI Decisions API support?
The public version of the API supports voice and visual inputs. This lets developers include spoken requests or visual context in applications that need a rapid response. Developers should check OpenAI's API documentation for implementation details and supported input formats.
How fast is the OpenAI Decisions API?
Early examples reported response-time improvements, but actual performance depends on the application and its conditions. Speak CTO Andrew Hsu said the time between actions fell from 1.4 seconds to 0.13 seconds, while OpenAI product lead Nikunj Handa described response times of roughly 100-180 milliseconds. These reported results are not a guarantee for every workload.
How should developers evaluate the OpenAI Decisions API?
Developers should test the API with representative inputs and measure latency, decision quality, and cost. Results can vary with the task, implementation, network conditions, and how the API works with other systems. Testing helps establish whether its speed and input support meet the needs of a specific product.
