OpenAI introduced GPT-6 Sol and Luna on September 22, 2026, extending its GPT-6 family into less expensive tiers. The launch prices are $2 input and $10 output per million tokens for Sol, and $0.10 input and $0.50 output for Luna.
The announcement attributes gains to training and serving improvements. The Sol model reference and Luna reference should guide endpoint and capability checks. Vendor comparisons do not establish that a cheaper model is appropriate for every task.
What the pricing makes possible
Lower token prices can change the economics of high-volume work. Classification, extraction, and short drafting tasks may benefit when the model completes them without costly retries or review.
Calculate the actual request pattern. Input length, repeated context, output length, and extra tool calls all influence cost. A workload dominated by long inputs has a different budget from one that produces large artifacts. Do not apply the output price to the entire conversation or ignore the cost of failed attempts.
Compare Sol and Luna against a defined result
Choose an acceptance criterion before running the evaluation. For extraction, check exact fields and missing-value handling. For coding, inspect the diff and required tests. For an agent, verify the final action in the system it operates.
Include ambiguous and incomplete inputs. A model that asks a clarifying question can be preferable to one that produces a confident answer with invented details. Lower cost is useful only when the system’s error and review burden remains acceptable.
Migration is more than changing a model name
Review tool definitions, context limits, supported inputs, structured-output behavior, and effort settings in the relevant model documentation. Preserve a stable evaluation set and record the configuration with every result.
Vision comparisons need an additional qualification: OpenAI’s September 25 changelog records an image-encoding fix for these models. Teams relying on early vision results should determine whether that issue affected their evaluation before using it to justify a lasting model choice.
Route only when the boundary is understandable
A smaller tier may be useful for bounded work, with a stronger model reserved for tasks that need deeper reasoning. Routing should have an explicit criterion and measurable effect. An automatic fallback that repeatedly retries an underspecified task can consume more than it saves.
Nerova’s assessment is that Sol and Luna widen the options for economical deployment. The durable decision is cost per accepted result, with quality, latency, and review included. Preserve that measurement as models and pricing change.