On September 22, 2026, OpenAI announced two new additions to the GPT-6 family: GPT-6 Sol and GPT-6 Luna. Both models are priced at half the promotional rates of their GPT-5.6 counterparts: Sol costs $2 per million input tokens and $10 per million output tokens, while Luna costs $0.10 per million input tokens and $0.50 per million output tokens. This is the first time since the GPT-6 Astra launch in early September that OpenAI has brought its new-generation flagship capabilities down to cheaper tiers.
How the three tiers line up
The GPT-6 lineup now has three tiers: Astra handles the hardest end-to-end tasks, Sol targets complex coding and agentic workflows, and Luna covers high-volume clerical work. OpenAI defines Luna's turf precisely: well-defined, high-volume tasks such as document summarization, information extraction, and quick Q&A.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Positioned for |
|---|---|---|---|
| GPT-6 Astra | $10 | $50 | Flagship, the hardest tasks |
| GPT-6 Sol | $2 | $10 | Complex coding, agentic workflows |
| GPT-6 Luna | $0.10 | $0.50 | High-volume clerical tasks |
Why they can cost half as much
OpenAI attributes the price cut to better caching and inference efficiency, saying the savings are passed directly to users. Both models were trained with methods similar to Astra's, spanning reasoning, factual reliability, coding, computer use, and alignment. OpenAI claims that on an internal factuality evaluation built from real user conversations, GPT-6 Sol makes about half as many mistakes as its predecessor, reaching Astra-level reliability. Note that this is the company's own framing; independent benchmarks have yet to weigh in.
First stop: Codex and ChatGPT Work
The rollout scope is worth a close look. Both models are coming first to Codex and ChatGPT Work, covering Plus, Pro, Business, Enterprise, and Edu users; Free and Go users can access Luna in the desktop app. Regular ChatGPT chat is excluded for now, and enterprise admins must enable the new models manually. In other words, this launch is aimed at getting work done, not casual conversation.
The takeaway: a cost war broke out on the same day
Notably, Anthropic released Claude Opus 5.5 on the very same day, claiming Fable 5.1-level performance at 40% lower running cost (see our earlier coverage). Two leading labs playing the "stronger and cheaper" card on the same day signals that the battleground is shifting from pure benchmark scores to the cost of making flagship capability the default choice.
Two details are worth noting for developers. First, the comparison baseline is GPT-5.6's "promotional" pricing, which is temporary by nature — the real discount depends on future standard pricing. Second, OpenAI's claim that both models beat Anthropic's Fable and Opus across tasks is vendor benchmarking at its most selective; independent evaluations will be the safer verdict.
