A short classification job and a complicated research assignment do not ask the same thing of a model. OpenAI’s July 9 release of GPT-5.6 makes that distinction explicit with three models: Sol as the flagship, Terra for a balance of capability and efficiency, and Luna as the most cost-efficient option in the family.
The launch announcement describes general availability following a limited preview. OpenAI emphasizes performance per unit of work, including the number of tokens required to complete tasks. Those are reported results from its evaluations, not a promise that every application will see the same improvement.
The cheapest request may create more work
A model choice affects more than the price of an individual response. If a task needs repeated attempts or substantial correction, the total cost includes those extra steps. A more capable model can be worthwhile for some assignments even when its first response costs more.
The reverse can also be true. A tightly defined task with an easily checked answer may not need the flagship model. The sensible comparison is the cost of an acceptable completed result, including any review, rather than the cost of producing text once.
For a small team, a useful evaluation set can include the actual kinds of work it delegates: extracting fields from a document, resolving a difficult code issue or preparing a sourced brief. Each needs its own acceptance criteria. Success on one category does not automatically establish that another should use the same model.
More parallel work needs a way to come together
The release also introduces an ultra setting that coordinates multiple agents for demanding tasks. OpenAI describes a tradeoff in which additional token use can improve results and shorten the time to an answer.
Parallel work is useful when parts of an assignment can genuinely be investigated separately. It also creates a coordination problem: overlapping conclusions have to be reconciled, and one agent’s unsupported assumption should not become evidence merely because another repeats it.
That makes the final review as important as the division of work. A research output needs inspectable sources. A changed application needs checks that exercise the finished behavior. The model’s ability to perform more steps does not establish that the outcome meets the original requirement.
Choosing among AI models can start with a small set of recurring assignments. Record the review effort alongside the model’s output and usage, so a cheaper first attempt does not hide a more expensive finished task.
Reporting & Sources
Based on OpenAI’s July 9, 2026 GPT-5.6 announcement. Later price changes on the living source page are excluded from this launch analysis. Performance claims are OpenAI’s; evaluation suggestions are ours.
