OpenAI opens Decisions API to all customers for low-latency classification
The AMW Read
Broad availability extends OpenAI into specialized decision workloads, meaningfully updating its product position while customer results remain workload-specific.
OpenAI opens Decisions API to all customers for low-latency classification
OpenAI made its Decisions API available to all customers on October 22, adding voice and visual modalities to a product announced September 29. The API adapts ChatGPT Luna for faster, cheaper inference rather than using a separately trained decision model. These systems select among supplied options and return confidence ratings. OpenAI reported response times of roughly 100–180 milliseconds. Speak, an early customer alongside Box, Datadog, and Ramp, said it reduced the time between actions from 1.4 seconds to 0.13 seconds.
The launch brings a general-purpose model provider into a market established by TypeSafe's Jev, where the buying criteria emphasize latency, cost, and confidence calibration for bounded tasks. It illustrates competitive pressure between specialized models and frontier labs repackaging existing capabilities through their APIs. OpenAI's reported customer results show potential for replacing slower LLM calls, but do not establish category leadership: its decision model placed eighth and fourteenth on two public indices cited by TechCrunch. Multimodal support adds another dimension to that competition.
For builders, the concrete opportunity is to test whether frequent classification or routing calls need a full LLM response. Speak's result offers a workload-specific reference, rather than a general performance guarantee. Compare candidates on end-to-end latency, cost per correct decision, and whether confidence scores reliably predict errors using representative inputs. The article provides no API pricing or controlled comparison across the named customers, so broader savings remain unquantified.
