OpenAI releases GPT-6 Sol and Luna — and cuts token prices in half

OpenAI has added two new GPT-6 models, Sol and Luna, to the lineup alongside the flagship GPT-6 Astra. There’s no GPT-6 Terra, at least not yet. The headline, though, is pricing: OpenAI cut the per-million token cost by half or more compared with the prior version, and says the new prices are the default rather than a promotion.

GPT-6 Sol is priced at $2 for input and $10 for output tokens, down from $4 and $20 for GPT-5.6 Sol. GPT-6 Luna lands at $0.10 and $0.50, versus $0.20 and $1.20 before. OpenAI says better caching and inference let it serve the models more cheaply, and it is passing that along to users and customers.

On benchmarks, the new models do better, but not by some heroic leap. On Zapier’s AutomationBench, Luna is up 5.4 percentage points. On DeepSWE v1.1, Sol basically lands next to Anthropic’s Fable: 68.8% at max effort versus 69.9% for Fable 5 at xhigh effort, though OpenAI is quick to point out the cost difference. Luna also posts scores in the same ballpark as Claude Opus 5 and Fable 5, again with a lower bill attached.

Anthropic complicated that comparison the same day by releasing Opus 5.5 and cutting its own pricing to $4 and $20 from $5 and $25. OpenAI still argues Sol is cheaper per task, while Anthropic says Opus 5.5 uses fewer tokens and ends up about 40% cheaper than Opus 5 on typical workloads. Nobody has run Sol and Opus 5.5 head-to-head yet, which leaves the usual fog around agent pricing intact. Per-token numbers are tidy. Real task costs are messier.

The parts developers may care about most are the plumbing changes. OpenAI says GPT-6 improves prompt caching with higher hit rates by default, discounts of up to 90% on cached input tokens, and a new ability to change reasoning effort and tool availability without blowing away the cache. There are also explicit breakpoints plus a dashboard and diagnostics tool to show what is cached. OpenAI also says the models answer more directly now, with fewer odd turns of phrase and slightly shorter replies.

OpenAI is also leaning hard on alignment. On an internal coding deception test, Sol’s misleading-claim rate fell to 1.3% from 10.4%. When faced with a broken search tool, it failed to disclose the problem 4.9% of the time, down from 77.5%. But when researchers used an “access denied” warning, Sol still tried to work around it in 64.4% of runs, only a small improvement from 68.2%. Availability is limited for now: Sol and Luna are in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, while Free and Go users get Luna in the desktop app.