OpenAI shipped GPT-6 Sol and GPT-6 Luna on September 22, with API rates cut by half or more. The independent measurements barely moved: one point gained on the intelligence index for Sol, nothing at all for Luna.
Key Takeaways
- Input falls to 2 dollars per million tokens on Sol and 0.10 dollar on Luna.
- Artificial Analysis scores Sol at 48 against 47 before, with regressions on knowledge work.
- The new rates are presented as permanent, with no expiry date attached.
Have an AI Sum Up This Article
ChatGPTGPT-6 Sol Cuts the Invoice and Leaves the Scores in Place
OpenAI laid out both models in the launch post it published on Tuesday. GPT-6 Sol now runs at 2 dollars per million input tokens and 10 dollars on output. Luna drops to 0.10 and 0.50 dollar.
Cache reads take the steepest cut of all, at 90 % off. A re-read token costs 0.20 dollar on Sol and a single cent on Luna.
One word carries the whole announcement: permanent. This is not an entry rate designed to climb once teams are locked in, which changes how a technical director can build a twelve-month budget around it.
A fixed rate also reshapes how a contract gets negotiated. A cut framed as final flows straight into the forecast, where a temporary discount forces everyone to keep a safety margin. The pattern goes back to spring, when the two labs started matching each other rate for rate.
Cache pricing is the quiet part of the change. Any workload that replays a long system prompt across thousands of calls now carries a fraction of the cost it did last week, and that shows up before any benchmark does.
On performance the picture is far less dramatic. Artificial Analysis puts GPT-6 Sol at 48 points on its intelligence index against 47 for the previous version. Luna holds at 37.
The movement is not even uniform. Gains show up on some evaluations and regressions on others, including GDPval-AA v2.1, which scores knowledge work rather than code.
OpenAI leans on a different number: roughly half as many mistakes on its own factuality evaluations. The lab played this same card in July, when the Luna line saw its price fall by 80 % in one move.
Luna Takes the Volume, Sol Keeps the Heavy Jobs
The two models are not aimed at the same work. Sol is positioned on complex tasks, coding included. Luna is built for volume: document summaries, information extraction, repetitive clerical work.
The argument lands on cost per task rather than price per token. On FrontierCode 1.1, GPT-6 Sol scores 49.3 % at 2.14 dollars a task, while Fable 5.1 reaches 50.3 % at 12.83 dollars.
One point of score for six times less money: the trade-off shifts from the leaderboard to the invoice. On AutomationBench, Sol at maximum effort beats Claude Opus 5 while costing 9 % as much per task.
On DeepSWE v1.1, Sol reaches 68.8 % at maximum effort. Luna matches that score in the same configuration for 0.22 dollar a task, which makes the choice between the two less obvious than the positioning suggests.
Effort level is doing a lot of work in those figures. A score posted at maximum effort and a score posted at default settings describe two different products, and the gap between them lands on the invoice.
So the announced split between the two holds up less well than it looks. A team pushing Luna to its ceiling sometimes lands where it would have gone to Sol, at a tenth of the spend.
OpenAI has spent the month stacking announcements, between the general intelligence claim Astra carried in early September and this double pricing release. The pace reads like a line-up still looking for its shape.
On OSWorld 2.0, which scores desktop control, GPT-6 matches Claude Opus 5 at roughly 80 % less. Availability follows the same split: ChatGPT Work, Codex and the API for both, with Luna reaching free accounts through the desktop app.
More articles on Horizon
- AI Agents Cheated Their Evaluator, the UN Warns
- Claude Opus 5.5 Decides How Hard to Think
- Daily AI Use Doubled in Six Months Across the US
The Fight Now Happens per Task, Not per Token
The timing says everything. GPT-6 Sol and Luna landed about ninety minutes after Anthropic announced Claude Opus 5.5, which carried its own 40 % cut on operating cost.
A team already in production now measures what a full task costs, real effort included, on a workload that actually resembles their own. Benchmark rankings sit one rung below that.
That shift also makes comparison harder. Two models posting the same score can burn very different amounts of reasoning to get there, and only an end-to-end measurement exposes the gap.
Enterprise buyers will read this as leverage. Two flagship cuts inside the same afternoon give any procurement team a fresh argument at renewal, whichever vendor they were leaning toward.
There is a floor to this, though. Serving these models still costs real money, and a price war that runs past the point of margin tends to end in usage caps rather than in further cuts.
For open models and low-cost providers the message is rougher. Their main argument rested on the price gap with proprietary flagships, and that gap just narrowed without them changing a thing.
The real turn will happen on the consumer side, where swapping the default model inside ChatGPT reshuffled everyday use without warning anyone. A model that costs half as much always ends up trickling down into the product.
Follow the story on Horizon.
