GPT-5.6 Sol, Terra and Luna: choose for the task
Keep the names Sol, Terra and Luna unchanged. Test the GPT-5.6 family without confusing the model, reasoning effort, working tool and cost.
Sol, Terra and Luna are model names in the GPT-5.6 family. They should not become “Sun” and “Earth” when a website is translated. Keeping the names intact matters because the choice must correspond to the identifier used in documentation and in the product available to the team.
The documentation positions Sol for complex professional work, Terra as a balance of capability and cost, and Luna for cost-sensitive, higher-volume workloads. These descriptions guide an initial selection. They do not replace testing the work that actually needs to be delivered.
Model, effort and tool are separate choices
The model determines part of the capability. Reasoning effort, where configurable, changes how a request is processed. The working tool supplies context and actions such as inspecting files or running tests. Two people using the same model can have different experiences because these other choices differ.
Record the combination being evaluated: exact name, product or API, configuration and task. Do not turn one application experience into a promise about every environment. Check plans, limits and integrations where the work will actually happen. A feature visible in a demonstration does not establish access for a particular account.
Begin with recurring work
If the team classifies messages or extracts document fields, prepare ordinary examples and known exceptions. Check whether the output follows its required format and handles missing information appropriately. A less expensive model may suffice, but that conclusion depends on accepted results observed in the test set.
Investigating a poorly documented defect is a different workload. The agent must support hypotheses, locate evidence and explain the change. Savings per request may be small compared with the time lost following a wrong lead. Choose a unit of work that represents this situation, instead of judging it by the same criteria as simple extraction.
Increase capability for an identifiable reason
Consider a hypothetical extraction task failing because the document does not contain the requested date. Changing models will not create the missing fact. The correct behavior may be to return an empty field with the explanation required by the output contract. Before increasing expenditure, classify the failure: missing data, ambiguous instructions, a tool limitation or a reasoning difficulty.
When the difficulty warrants it, test another configuration under the same criteria. Record what improved and what remains incorrect. Avoid sending every task to the most expensive model by default, or persisting with the cheapest option when repeated repairs are necessary. The workflow should follow the evidence actually available.
Keep routing simple enough to operate
A small team can begin with a default model and an alternative for difficult work. Numerous routing rules invented before measuring cases add complexity without a demonstrated benefit. Revisit the choice when the work changes, observed cost shifts or availability changes in the purchased service.
The cost-per-task guide includes retries and review effort in the calculation. The article on adopting Codex examines the working tool and its integration with a project. Separating these decisions makes it easier to identify what actually improved when a configuration works well.
About the author
Tiago F SantiagoComments
No comments yet
Share a question or an experience related to the article.
Related articles
