OpenAI halves prices with GPT-6 Sol and Luna. The benchmarks are its own
Sol costs $2 in and $10 out per million tokens, Luna $0.10 and $0.50. Every benchmark comparison at launch came from OpenAI, and the Claude model it compared against was replaced the same morning.
The prices are on OpenAI's pricing page. The Claude comparison went stale 90 minutes before launch.
OpenAI released GPT-6 Sol and GPT-6 Luna on Tuesday, September 22, 2026, in the API, Codex and ChatGPT. They replace GPT-5.6 Sol and Luna and sit below GPT-6 Astra, the flagship that launched on September 3. OpenAI says both were built with methods similar to Astra's, TechCrunch reported.
Sol is exactly half the old price. Luna halves input and cuts output by about 58 percent. Luna has now fallen twice in two months: it was $1 in and $6 out before the 80 percent cut in July. The numbers above are on OpenAI's pricing page, so they are the solid part of this launch.
Whose numbers
Every benchmark result at launch was OpenAI's own, and The Next Web notes that competitor scores came from published reports, not from OpenAI running the models itself. The headline claim, per VentureBeat: Sol scores 33.2 percent on AutomationBench at $0.27 per task, against 26.9 percent for Claude Opus 5 at 11.1 times the cost. OpenAI's other cost comparisons put Sol at 9 to 40 percent of Opus 5's cost. Outlets do not agree on one of the scores either: VentureBeat reports 60.5 percent on OSWorld, Digital Applied 64.4.
The timing matters too. Anthropic released Claude Opus 5.5 about 90 minutes before OpenAI's launch, at $4 in and $20 out, 9to5Google reported. So the comparison OpenAI shipped was against a model its rival had just replaced.
Why a build studio cares
A 50 percent cut is real money for an agent that makes thousands of calls a day, so it is worth rerunning the cost model. A vendor benchmark is not a reason to switch. The test that settles it: take a few dozen real prompts from your own logs, run them on both models, and compare cost per successful task, not cost per token. A cheaper model that needs a retry is not cheaper. The launch table tells you what OpenAI measured. Your logs tell you what you will pay.
Next step: check OpenAI's pricing page against your current bill. If you want your own workload tested on both models before you switch, write to us at hello@gattyworks.com.