Not every task needs the most expensive model available. With GPT-6 Sol and GPT-6 Luna, OpenAI is turning that idea into a clearer model family: Astra remains the maximum-capability option, Sol targets demanding professional work, and Luna aims for scale at a low cost.
The prices make the strategy clear. GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens. GPT-6 Luna costs $0.10 and $0.50, respectively. OpenAI says both are 50% below the promotional prices of their GPT-5.6 equivalents.
This matters for teams that have learned a simple lesson: using the best model for every step is an expensive architecture, not a quality strategy.
Sol for depth, Luna for volume
Sol is positioned for coding agents, automation across applications, and complex professional tasks. OpenAI reports improvements over GPT-5.6 Sol in software-engineering, factuality, and computer-use evaluations, while approaching larger models at a lower cost.
Luna is the high-volume option. It may fit classification, extraction, predictable transformations, first passes, and subtasks delegated by a primary agent. The savings are real only when the task fits the model. If the output requires several attempts or extensive corrections, the cheaper call starts charging interest.
The published benchmark results come from the launch and use specific effort levels, tools, and harnesses. They are useful for forming hypotheses. They should not make the team's model decision.
The advantage is in routing
The most useful change may not be replacing one model with another. It may be stopping the habit of treating every step as equal. A single request could use Luna to organize context and find candidates, Sol to implement and test a change, and Astra only when ambiguity or risk justifies the additional cost.
This design needs metrics. Compare total task cost, latency, first-attempt success, and human review time. For coding agents, also track commands executed, unnecessary file changes, and correction cycles. Token price without final quality is a clean spreadsheet telling the wrong story.
The family also inherits communication changes from Astra: shorter responses, less jargon, and clearer statements about what the model did or did not verify. That is hard to capture on a leaderboard, but it matters when a person must review many agent decisions.
Where they are available
Sol and Luna are coming to ChatGPT, the API, and Codex. GitHub has also announced both models in Copilot, with a gradual rollout and policy controls for organizations. Before making either model a team default, check availability, consumption multipliers, and plan rules in each product.
This launch does not end the race for maximum capability. It highlights another competition that is closer to production reality: who can deliver enough intelligence at the lowest total task cost. For AI Engineering, that question is often more useful than finding which model finished a few points ahead on a chart.