OpenAI’s guide recommends matching model capability, reasoning effort and speed to the work. It also asks developers to measure task success, latency and cost, keep instructions consistent, and define what an agent may do without further input. The retained version names GPT-6.1 Sol alongside Astra and Luna. That current table should not be presented as the original lineup on the guide’s earlier publication date.
A creative studio can use the guide to write an evaluation brief before choosing a model. The following example is hypothetical. We have not run the test, compared model performance or measured a saving, and the brief does not authorize a new paid run or a change to anyone’s model budget.
Define one deliverable
Suppose a producer wants to turn an approved product-shoot brief into an asset delivery list. Supply the approved brief and a fixed output format: the required asset, its destination and the instruction that supports including it. Tell the assistant to flag missing dimensions or conflicting directions instead of filling them in.
Write the acceptance test before looking at the output. The list must include the mandatory assets, retain the supplied product names and link each requirement to its place in the brief. It must identify unresolved instructions. The producer can then assess a proposed output against the same reference rather than reward whichever answer sounds most confident.
Measure the whole task when a run is authorized
For a later authorized comparison, retain the input and model settings, the first output, corrections and the accepted delivery list. Record elapsed time and billed usage alongside the human review needed to reach acceptance. Reuse the same task and acceptance conditions when comparing another configuration. We report no results from that proposed comparison here.
The guide discusses caching, compaction and parallel work as ways to manage a workflow. Their value depends on the task and the application’s implementation. Keep the evidence a reviewer needs, and give the agent clear boundaries around client files and external actions. A written acceptance test gives the studio something useful to prepare now, even while execution and a purchasing decision remain open.