OpenAI Decisions API Turns Response Time Into a Buying Category

OpenAI made its Decisions API widely available to all customers on October 22, 2026. According to TechCrunch, the fast, low-cost decision model was built by reconfiguring ChatGPT Luna, cutting inference from 1–3 seconds to roughly 100–180 milliseconds. The announcement puts a specific response-time range at the center of a product category rather than leaving latency as a supporting model specification.

TechCrunch names Box, Datadog, Ramp, and Speak as early customers that replaced LLMs in high-throughput workflows. The report also places OpenAI alongside TypeSafe's Jev, Liquid's D1 Ultra, and Base Lab's Elephant-1. It describes a decision-model category created by TypeSafe's Jev, with competition now broad enough to make model selection a separate procurement question.

For builders, the distinction is between generating a flexible response and making a bounded decision inside an application. The announcement does not establish that every existing LLM task belongs in the latter category. It gives teams a reason to identify where their applications are waiting for a decision rather than benefiting from an open-ended response.

Latency becomes part of the product definition

The reported move from 1–3 seconds to 100–180 milliseconds changes the interval a developer is being asked to evaluate. Those numbers are attributed inference figures, not a promise that an entire application interaction completes in the same time. Any evaluation should keep model execution separate from the rest of the workflow so that the comparison remains meaningful.

A fast model cannot by itself establish that a decision is correct. Builders still need a definition of acceptable output and a way to identify cases that should not proceed automatically. That requirement is especially important when speed makes a result easier to place directly in an execution path. Faster delivery changes when an answer arrives; it does not determine what the application should do with it.

The low-cost description is also qualitative in the supplied reporting. There is no price schedule here from which to calculate a workload budget. A buyer should request the applicable commercial terms rather than infer a saving from the label. The defensible published comparison is the inference range, while the business comparison must be made against the buyer's actual workload requirements.

Customer names establish interest, not interchangeability

Box, Datadog, Ramp, and Speak are useful evidence that the category is being used in high-throughput work. Their inclusion does not disclose a common task definition or a shared acceptance standard. A builder cannot take the presence of those names as proof that a different production workflow can replace its existing LLM without evaluating behavior.

The competitive set makes that evaluation more important. TypeSafe's Jev, Liquid's D1 Ultra, and Base Lab's Elephant-1 offer named alternatives within the same category. The fact sheet does not supply comparable latency measurements or prices for those products. Ranking them against OpenAI on either dimension would therefore go beyond the available evidence.

Instead, use the category to refine the buying brief. State which decision is being requested, what inputs are available, and what should happen when the output is unsuitable for automatic use. This is a specification exercise before it is a vendor comparison. A narrowly defined decision gives a team something it can test; a general demand for faster AI does not establish a meaningful acceptance condition.

The trade-off is speed without a complete quality contract

The 100–180 millisecond figure is attractive precisely because it is concrete. The corresponding question about task quality is not answered by that figure. Reconfiguring ChatGPT Luna for decisions establishes the product's stated origin, but it does not provide a workload-specific error profile. Teams need to discover where behavior differs from the model they already use.

Replacing an LLM can also change the assumptions surrounding its response. An application may currently rely on explanatory text, an opportunity for human review, or a path for handling ambiguous inputs. These are possibilities to inspect in your application, not documented omissions in the Decisions API. A migration should identify those dependencies rather than presume that a faster answer is a compatible answer.

The cost of that inspection belongs in the decision to migrate. Existing workflows contain examples of what users actually ask and what the application actually accepts. Reusing those examples for comparison is more informative than adopting a vendor's category label as a complete requirement. The public customer list supports taking the product seriously; it does not replace the work of establishing fit.

What To Do About It

• Map the decision boundary: Identify where your application needs a bounded decision rather than an open-ended response, and compare that requirement with the reported starting inference range of 1–3 seconds.
• Measure the complete path: Record application response time separately from model inference so a faster model is not confused with a guaranteed user experience, using the reported inference target of 100–180 milliseconds.
• Compare actual outputs: Run representative existing workflow inputs through the candidate and the current model, retaining disagreements for review before approving a production change to the Decisions API.
• Request comparable terms: Obtain pricing and workload-specific evaluation conditions instead of calculating a saving from the phrase low-cost, and keep the vendor shortlist grounded in the named alternative, TypeSafe's Jev.

The Bottom Line

OpenAI's October 22 availability announcement makes low-latency decisions a product buyers can evaluate directly. The reported 100–180 millisecond range is a reason to inspect time-sensitive workflows, not a reason to replace every LLM call. Customer adoption establishes interest without establishing your acceptance criteria. Buy the decision behavior your application needs, then verify that the speed claim survives the complete path.