GPT-6 Astra shipped on 3 September. On 22 September 2026, OpenAI's two lower tiers followed on OpenRouter: GPT-6 Sol and GPT-6 Luna. Each replaces its GPT-5.6 namesake from July, and each makes a different trade.
GPT-6 Sol: same price, faster
The price has not moved, and the index moves by one point. The real change is speed. Sol streams 37% faster and returns its first token 1.7 seconds sooner. For interactive use, that latency change is the upgrade.
GPT-6 Luna: the price cut
Luna goes the other way. It is one point lower on the index, and it costs half as much for input and 58% less for output. At $0.50 per million output tokens with a 1.05M context, it undercuts DeepSeek V4.1 Flash ($0.15 / $0.60) on both input and output. DeepSeek scores 40 on the same index to Luna's 37, and streams at 142 tok/s.
Where they sit in the GPT-6 line
All three accept text, images and files, and all take a reasoning_effort parameter.
Pro variants. OpenRouter also listed *GPT-6 Sol Pro* and *GPT-6 Luna Pro* on the same day, at the same list prices as the base models. Neither has a published Intelligence Index score yet. We only add a model once it has an external score to anchor its rating, so they are not in the catalog yet.
Which one
For interactive use, move from GPT-5.6 Sol to GPT-6 Sol: same bill, less waiting. For high-volume work where GPT-5.6 Luna was good enough, GPT-6 Luna cuts the bill by more than half, and one index point is unlikely to show up in bulk extraction or classification. Test it on your own prompts first.
Prices are list prices on OpenAI's endpoint. Speed figures are OpenRouter medians over a trailing thirty-minute window, read 23 September 2026. Our category ratings for both models are editorial estimates, labelled Estimated, until accepted benchmark results replace them.
