OpenAI expanded its GPT-6 family on September 22, 2026, launching GPT-6 Sol and GPT-6 Luna — two new models that undercut their predecessors by 50% or more on API pricing, making GPT-6-class intelligence genuinely economical for developers building at scale.
Sol is designed for complex reasoning and coding tasks. Its pricing lands at $2 input / $10 output per million tokens, down from $4/$20 for GPT-5.6 Sol. Luna targets high-volume structured work — document summarization, classification, data extraction — at $0.10 input / $0.50 output, down from $0.20/$1.20. The output token cut on Luna reaches 58%, the steeper discount of the two.
Both sit below GPT-6 Astra — the flagship model OpenAI launched earlier in 2026, the one that scored a perfect 42/42 at the International Math Olympiad. Astra remains the ceiling; Sol and Luna are the pragmatic API tier.
What this means for developers
The practical impact is that GPT-6-generation reasoning is now priced within striking distance of workloads that previously justified GPT-4-era models on cost grounds. At $0.10 input, Luna undercuts a wide band of older models still running in production pipelines purely because switching wasn't worth the cost savings.
OpenAI positions Sol as offering "fewer mistakes" compared to its GPT-5.6 predecessor at equivalent capability — suggesting quality improvements alongside the price drop, not a lateral move. Luna is explicitly framed for tasks with a clear, structured goal rather than open-ended reasoning.
This adds to an increasingly competitive API landscape. Anthropic's Claude Fable 5.1 cut cache-read pricing earlier this month; Google's Gemini 3.8 Flash holds strong benchmark performance at low cost. The trajectory is clear: capable models are becoming cheap enough that cost is no longer the primary reason to stay on legacy tiers.
Availability
Both GPT-6 Sol and GPT-6 Luna are available today through the standard OpenAI API. No waitlist, no preview program — they're live alongside the existing GPT-6 Astra and GPT-6 Astra mini options. Teams already calling the GPT-5.6 Sol or Luna endpoints can evaluate a model switch with a straightforward A/B test.
OpenAI has not yet published a detailed side-by-side benchmark against Astra, but the pricing structure makes the tradeoff legible: Astra for when maximum capability matters, Sol for complex tasks where cost must be managed, Luna for predictable, high-throughput jobs where the query is well-defined.