Two faster, cheaper AI models create a familiar problem people choose the “best” one once, then use it for everything.
OpenAI introduced GPT-6 Sol and GPT-6 Luna on September 22 as lower-cost members of the GPT-6 family. OpenAI says both cut API pricing by 50 percent compared with the promotional pricing of their GPT-5.6 counterparts. Sol is positioned for more capable professional work; Luna is the lighter, cheaper option. GPT-6 Astra remains the highest-capability model in the family.
For everyday work, the useful question is not which model wins overall. It is which task deserves more capability?
Use a three-lane rule
Fast lane routine drafting, extraction, rewriting, classification, formatting and straightforward summaries. These tasks are clear and easy to check.
Deep lane ambiguous analysis, multi-step reasoning, difficult debugging, synthesis across many sources and recommendations where a subtle error matters.
Max lane the hardest work you can justify giving to the strongest model available, especially when the task is complex and the cost of a weak answer is high. Even here, human review stays in the workflow.
Escalate after evidence, not instinct
Start routine work in the lighter lane. Move up when the result fails a defined quality check, the task has important ambiguity, or the cost of a mistake is high.
For example, extracting dates from a clean document does not need the same model choice as comparing conflicting evidence across a large research pack. A first-pass rewrite does not need the same reasoning budget as a recommendation that will influence a client decision.
Define the failure condition first
Before switching models, decide what failure looks like. It might be missing a source, losing an instruction, producing inconsistent calculations or failing to identify an important trade-off. If the lighter model passes the check, the stronger model may add little value for that task.
Classify this task as Fast, Deep or Max. Consider ambiguity, number of reasoning steps, verification difficulty and cost of error. Tell me the lightest capable starting point, the quality check to run, and the exact failure condition that should make me escalate to a stronger model. Do not recommend extra capability unless it solves a specific risk.
Do not pay for capability you cannot measure
If you cannot state what a stronger model needs to improve, you cannot tell whether the upgrade helped. Save a representative task and rerun it across models using the same inputs and review checklist. Compare usable output, correction time and failure rate rather than how impressive the first answer sounds.
Availability matters
OpenAI says Sol and Luna are available in ChatGPT Work and Codex for eligible paid plans, with Luna also available to Free and Go users in the desktop app. They are also available through the API. Product access can vary by plan and surface, so check what you can actually use before rebuilding a workflow around a model name.
Make routing a habit, not another decision
You do not need to debate model choice before every prompt. Create a default routine tasks start light, complex tasks start deeper, and only the hardest or highest-consequence work goes to the maximum-capability lane. Revisit the rule when prices, access or model performance changes. The goal is fewer model decisions during the day, not a more complicated model picker.
Three examples of routing in a normal day
Email cleanup extracting requests from a thread and rewriting a reply is usually Fast work because the input and expected output are clear. Research synthesis comparing conflicting sources and deciding which evidence matters belongs in Deep because the model has to preserve context and trade-offs. A consequential recommendation a decision involving substantial money, legal exposure or a difficult-to-detect error may justify Max capability plus explicit human verification.
The important part is that the same project can move between lanes. Research collection can start light, synthesis can move deeper, and only the most consequential decision may need maximum capability. Route the step, not the entire job.
Human boundary
A stronger model can reduce some errors; it does not become accountable for the decision. Keep human review where the consequence matters, and treat model choice as one part of a reliable workflow rather than the workflow itself.
Sources
OpenAI – Introducing GPT-6 Sol and Luna, September 22, 2026
