Two new models, half the price
OpenAI added two models to its lineup on September 22, 2026: GPT-6 Sol, aimed at complex work such as coding, and GPT-6 Luna, built for high-volume, low-complexity tasks such as summarizing documents or answering routine questions. Both are cheaper than the tiers they replace. Sol now costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20 under GPT-5.6 Sol. Luna costs $0.10 per million input tokens and $0.50 per million output tokens, down from $0.20 and $1.20. That works out to a 50 percent cut across the board on Sol, and on Luna a 50 percent cut on input pricing with a steeper roughly 58 percent cut on output.
For comparison, OpenAI's top-tier model, GPT-6 Astra, is priced at $10 per million input tokens and $50 per million output tokens. Sol now costs a fifth of Astra's price, and Luna costs a small fraction of it.
One caveat on the discount: the $4/$20 and $0.20/$1.20 baseline prices being cut were themselves promotional introductory rates for GPT-5.6 Sol and Luna, not long-standing list prices. The headline "50 percent cheaper" framing is accurate against what customers were actually paying last week, but it is a cut from a price OpenAI had already discounted once, not from a stable long-term baseline.
Where the savings come from
OpenAI says both new models were trained using the methods behind GPT-6 Astra, its most capable model, and inherit much of Astra's gains in accuracy, coding and computer use, packaged into faster, cheaper models. The company attributes the price cuts to improvements in caching and inference rather than to a smaller or less capable model underneath.
On OpenAI's internal factuality evaluation, built from real conversations where users flagged mistakes, GPT-6 Sol reportedly makes about half as many errors as its predecessor while approaching Astra's reliability, at a much lower price.
The numbers behind the claims
OpenAI published several benchmark results alongside the launch. On AutomationBench, a test of professional-work tasks, Sol scored 33.2 percent accuracy at its highest reasoning-effort setting, at a cost of $0.27 per task; Luna improved 5.4 percentage points over its predecessor on the same benchmark while costing 58 percent less to run. On DeepSWE, a coding benchmark, Sol reached 68.8 percent and Luna reached 66.6 percent at maximum effort, a narrow gap that suggests Luna has closed much of the distance to Sol on coding despite the large price difference between them. Sol scored 60.5 percent on OSWorld 2.0, a benchmark of artificial intelligence (AI) models operating a computer interface, and 56.4 percent on Agents' Last Exam, a test of agentic task completion.
As with any vendor-published benchmark, these figures come from OpenAI's own testing and have not been independently reproduced.
Who gets what
Sol is rolling out to ChatGPT Work and Codex, OpenAI's coding-agent product, for most paid accounts, as well as through the application programming interface (API). Luna is available more broadly, including in OpenAI's desktop app and for Free and Go tier ChatGPT users, with the rollout to the ChatGPT app and website continuing through launch day.
Timing was not an accident
Anthropic released Claude Opus 5.5 roughly 90 minutes before OpenAI's announcement, on the same day, cutting Opus's running cost by 40 percent compared with Opus 5 and removing five-hour usage caps for subscribers. Neither company has said the timing was coordinated, but the back-to-back releases underline how directly the two labs are now competing on price as much as on raw capability. Sol's focus on coding work in particular puts it in direct competition with Opus, which Anthropic has marketed specifically for agentic coding.
What it means in practice
For developers already building on GPT-5.6 Sol or Luna, the upgrade is close to a straight cost cut for the same or better output, since OpenAI has not raised the ceiling on task complexity these models handle. Teams running high-volume, low-stakes work, document extraction, ticket triage, quick lookups, are the most likely to see their API bill drop immediately after switching to Luna. Teams doing serious coding or agentic work should weigh Sol against Opus 5.5 on their own workloads rather than on either company's benchmark tables, since both companies tested their models on their own choice of tasks.
