Skip to content
NewsProduct launch

OpenAI's Decisions API exits beta after Datadog reports 40x faster risk checks

· by

The short version

OpenAI made the Decisions API generally available on October 22, 2026, and published customer results, including Datadog's claim of 40x faster p95 latency for screening risky agent commands.

OpenAI's Decisions API exits beta after Datadog reports 40x faster risk checks

OpenAI moved its Decisions API out of public beta into general availability on October 22, 2026, and used the milestone to publish production numbers from early customers rather than new vendor benchmarks: Datadog said the endpoint scores whether an AI agent's shell command is safe to execute with p95 latency 40 times faster than the frontier models it replaced, at comparable accuracy, per OpenAI's GA announcement.

What

The endpoint still runs on gpt-6-luna through the dedicated POST /v1/decisions call that OpenAI opened for beta testing on October 6, and pricing is unchanged at GA: $0.10 per million input tokens, with cached input at $0.01 per million and no charge for output tokens, per the same announcement. Pondero's earlier coverage of the beta launch has the original predicate/choice/score mechanics and the vendor-claimed 10x speed figure versus a standard Responses API call. Since beta, OpenAI added regional processing in the US and EU, support for native JSON inputs and images inside question instructions, and raised the per-request question limit to 400, double the prior cap. The GA post also confirms Zero Data Retention and HIPAA eligibility for qualifying customers. A native GPT-Live voice integration, which lets a live conversation trigger a decision mid-sentence, had already shipped on October 15 (Pondero's coverage of that update) and is folded into this GA release rather than new to it.

Customers are reporting double-digit latency and cost cuts

The GA announcement is thinner on new vendor claims and heavier on named customer results, a shift from the beta rollout's own-benchmark framing. Sierra said its customer-support routing runs the endpoint at under 280 milliseconds p95, about three times faster and 90% cheaper per call than its prior approach, with no change in customer outcomes, per quotes OpenAI published in the GA post. Ramp reported the same switch made AI cost attribution four times faster and 30% cheaper than routing the task through GPT-6 Luna directly, and Box said it saw roughly 5x faster document-routing decisions on a 30-case insurance-certificate test with matching accuracy, per the same quotes. Speak, a language-learning app, cut the time between a user's utterance and the next action from 1.4 seconds to 0.13 seconds inside a GPT-Live session, also per OpenAI's post. These are OpenAI-selected, OpenAI-published customer quotes rather than independent benchmarks, but they give teams evaluating the endpoint for routing, moderation, or agent-tool-selection a concrete range to test against their own workloads instead of relying solely on OpenAI's 10x headline figure.

What to watch next

Watch for independent latency or cost benchmarks from developers outside OpenAI's selected customer list now that the endpoint is unrestricted, and whether the newly doubled 400-question-per-request cap rises again as adoption grows. OpenAI has not said whether regional processing will extend beyond the US and EU.

Sources