AI, in your work.

What changes with general availability

On October 22, 2026, OpenAI announced general availability of Decisions API, built with GPT-6 Luna. It targets bounded decisions, such as classifying an input or selecting the next step in a process. OpenAI announcement.

The update adds native JSON inputs, context caching and GPT-Live integration. The maximum rises from 200 questions during the public beta to 400 per request. The guide identifies gpt-6-luna as the only currently supported model. Decisions guide.

Three answer types

The API reference defines these results. The examples in the final column are hypothetical.

Pricing and a caching discrepancy

Uncached input costs $0.10 per million tokens, with no output-token charges. Cache reads cost $0.01 per million tokens, and 30-minute writes cost $0.125. These are base rates: regional processing premiums and long-context multipliers may apply. Decisions guide, official pricing.

Nexo hypothetical example: one million requests averaging 500 billed input tokens per request would consume 500 million tokens. At the uncached base rate, that would cost $50. The calculation assumes that total consumption, including instructions, and excludes retries and premiums; it is neither an actual bill nor a measurement.

The announcement mentions 12-hour cache writes, whereas the technical reference only lists ttl: 30m. Because of this discrepancy, the 12-hour option remains unconfirmed for this endpoint in the documentation reviewed. Announcement, technical reference.

Access, limits and data handling

Requests are sent to POST /v1/decisions using API-key authentication. Any question can receive a refusal even when other questions in the same request receive answers. API reference.

The maximum of 400 questions per request is separate from a requests-per-minute quota. Usage limits apply at organization and project level and should be checked in the relevant console. Rate limits guide.

OpenAI documents availability in all supported API regions and regional processing in the United States and Europe—the EEA and Switzerland. Access from a region does not mean inference takes place there. By default, abuse-monitoring logs are retained for up to 30 days; Zero Data Retention is available to eligible customers, subject to limitations. Data controls.

How to read the speed claim

OpenAI says Decisions can respond up to ten times faster than GPT-6 Luna through Responses API. The announcement also includes selected customer testimonials. These are provider claims; they do not establish equivalent performance in a Nexo workflow or replace an evaluation of that workflow. OpenAI announcement.

Editorial example: evaluating document classification

A hypothetical pilot could classify documents as invoices, contracts or other and use the label to suggest a work queue. That classification alone would not authorize a payment or acceptance of a contract.

Nexo’s interpretation is that a bounded answer makes integration easier, but its format does not demonstrate that the decision is correct. A useful evaluation could follow these steps:

• Prepare representative documents labeled by people, holding back a sample that is not used to adjust the instructions.
• Compare category-level errors, their consequences and the proportion of documents requiring review against the current procedure.
• Measure median and 95th-percentile latency alongside billed tokens, retries and cost per resolved document.
• Send “other” results, refusals and cases below a confidence threshold established using the labeled sample to review.