GPT-6 Sol and Luna: OpenAI halves its mid-tier prices 19 days after Astra

Nineteen days earlier OpenAI put a $10-input, $50-output price on GPT-6 Astra and called it the most intelligent model it had built. On 22 September 2026 it published a very different number for the tier below: GPT-6 Sol and GPT-6 Luna, priced at half of what the GPT-5.6 versions cost — and, OpenAI told VentureBeat, permanently. A flagship launch tells you what a lab thinks is possible. A mid-tier price cut 19 days later tells you what a lab thinks it can actually afford to serve at scale, which is the more useful number for anyone building a product on top of it.

The reduction is exact and OpenAI states it plainly: both models are 50% cheaper than their GPT-5.6 predecessors on input tokens, and Sol is 50% cheaper on output too, while Luna's output falls 58%. Cached input — the discounted rate for tokens a request has already sent before, which is most of what a long-running coding agent or chat session actually pays for — falls further still: $0.20 per million for Sol and $0.01 per million for Luna, according to OpenAI's published API pricing.

OpenAI's explanation is an efficiency story, not a capability trade-off. It says Sol and Luna were trained with the same methods as Astra and carry the same alignment work, and that the price only fell because serving them got cheaper — improved prompt caching that lifts the default cache-hit rate, plus inference gains that cut the compute per token. That distinction matters for anyone who has watched a lab quietly shrink a 'cheap' model's underlying weights: this is presented as the same model family getting cheaper to run, not a smaller model wearing the same badge.

It also matters because OpenAI has form on time-limited discounts. GPT-5.6 Sol's own promotional pricing — the $4/$20 rate this cut replaces — is itself guaranteed only through 21 November 2026, a reminder that 'sale price' has been the norm rather than the exception in this market. Sol and Luna's new rates are explicitly not that: an OpenAI spokesperson told VentureBeat the figures are permanent, which changes what a developer can safely build a unit-economics model around.

Sol and Luna did not land in isolation. xAI shipped Grok 4.7 the day before, on 21 September, holding its price at $2 input / $6 output — the same rate as Grok 4.6 — while pushing benchmark gains through a larger base model instead. Anthropic then released Claude Opus 5.5 on the morning of 22 September, roughly 90 minutes before OpenAI's own launch, at $4 input / $20 output: a 20% per-token cut on Opus 5, and a claimed 40% saving on typical workloads through fewer tokens per task. Three labs repriced within 48 hours of each other. That is not a coincidence of calendars; it is what happens once one lab's inference costs fall far enough to make a public price cut the cheapest way to defend developer share.

The practical upshot for a team already spending real money on the API is a portfolio, not a single model choice. Luna's new rate — $0.10/$0.50, with cache reads at a cent — is aimed squarely at tightly-scoped, high-volume work: summarising, extracting, classifying. OpenAI's own numbers show why that recurs: it says GPT-5.6 Luna usage grew more than tenfold after an 80% price cut in July, and that cheaper inference let Replit offer a Free Mode to millions of users without touching their normal usage credits. Sol, at $2/$10, is priced to sit exactly alongside Anthropic's Claude Sonnet 5 and is aimed at the coding and agent work a team runs constantly rather than occasionally. Astra remains the model for the hardest, highest-stakes jobs, at five times Sol's input price and unchanged since 3 September.

What to watch next is whether this was OpenAI reacting to Anthropic and xAI, or the other way round — the timing of all three launches inside two days argues that none of them wanted to be the lab left holding the expensive price once the others moved. Anthropic has already said Claude Sonnet 5.5 and Haiku 5.5 follow 'in the coming weeks,' with similar cost claims attached. If the pattern holds, the next cut worth watching is not the flagship. It is whatever sits below it.