A typed AI decision still needs business authority and cost evidence
This week's Runtime and cost updates make the boundary between a model decision and an authorised business action clearer.
DataInbox
Architecture & Research
OpenAI made its Decisions API generally available on October 22 for GPT-6 Luna. It returns typed choices, scores or probabilities from supplied text, images or JSON. That is useful input to a workflow; it does not decide which customer record is authoritative, who may change it, or whether an action has been approved. We clarified that boundary in the existing Agent Runtime explanation. This is a proposed DataInbox Companion pattern, not a shipped Decisions integration.
OpenAI's October 22 API changelog
See the Agent Runtime explanation
Consider a service team routing a return request. Before: an agent classifies the message, then someone checks the order state, eligibility rule and approval trail across systems before issuing a refund. After, the typed classification can enter a governed flow that retrieves the current order, applies the business rule and pauses for approval where required. DataInbox would own the state, rules, authority and evidence in this proposed architecture. The retailer or payment system still owns the actual refund operation.
If the integration and permissions work, a pilot could test fewer manual handoffs, less duplicated rule maintenance and easier reconstruction of a decision. Count complete authorised resolutions, handoffs, corrections, approval wait time and total operating cost. Those benefits are DataInbox hypotheses, not outcomes OpenAI reports or guaranteed savings.
Compare the cost of an outcome
We also clarified in the AI Cost Benchmark Lab that a Decisions request has its own price structure and is outside the ordinary input, cached-input and output model-call estimate. The eleven selected Standard model tariffs and four team-seat rates were rechecked without a USD change. The current euro illustration now uses the ECB reference rate of October 23, EUR 1 = USD 1.1193. For 380 example GPT-5.6 Terra uses, the estimated provider-token cost remains $59.22, about €52.91 rather than €52.85 at the earlier rate. The change is currency conversion, not a cheaper or dearer model.
Explore the AI Economics Dashboard
The dashboard's stored DataInbox Runtime team scenarios were calculated on August 24 and have not been rerun. Its historical price comparison keeps 1,000 questions with 10,000 input and 500 output tokens each. Equal token counts do not prove equal model quality. Tools, cache writes and storage, grounding, regional or faster processing, tax, review and failed attempts can add cost. A team should compare completed outcomes under the same conditions, including the cost of checking and approving them.
The locally verified DataInbox CLI is rc.54 with a matching skill and command contract; offline support for reviewing intended configuration changes does not prove a new server capability.