OpenAI released two new models, GPT-6 Sol and GPT-6 Luna, on September 22. Their API model IDs are gpt-6-sol and gpt-6-luna, and both became available to developers the same day. Pricing for both is exactly half of what OpenAI charged for the equivalent GPT-5.6 models, and both come with a 1.1 million-token context window.
On the ChatGPT side, the Work, Pro, Business, Enterprise, and Edu plans all get access to both new models. Luna is also rolling out to the free tier and the Go plan on desktop, making it the more widely available of the two, while Codex is picking up the update alongside ChatGPT Work. OpenAI has not published model weights for either release — Sol and Luna are API-only, with no open-weight option.
Official benchmark results
OpenAI's own release materials point to several benchmark results as evidence of the upgrade:
- AutomationBench 1.0.6, which measures professional work tasks: Sol posted 33.2% accuracy on the xhigh reasoning tier, at a cost of $0.27 per task. Luna, run on the high tier, outscored its predecessor by 5.4 percentage points while costing 58% less to run.
- DeepSWE v1.1, a coding benchmark: Sol reached 68.8% on its max reasoning tier, and Luna reached 66.6%.
- OSWorld 2.0's offline version, which tests computer-use tasks: Sol scored 60.5% on the xhigh tier, which OpenAI says is comparable to Opus 5 running on a mid reasoning tier, but at 80% lower cost. Luna, on its max tier, outperformed the previous-generation Sol model outright, at roughly a tenth of the cost.
- Agents' Last Exam: Sol scored 56.4% on the max tier.
The release also ships with an overhauled prompt-caching system. OpenAI's reasoning is that agents tend to resend the same system instructions, tool definitions, and conversation history on every single turn, so the new system is designed to raise the default cache-hit rate across the board, which in turn cuts the cost of cached reads by as much as 90%.
Third-party read: prices down, scores flat
Independent testing from Artificial Analysis tells a somewhat different story on the capability side, diverging from what OpenAI's own materials emphasize. The firm's testing puts both models' intelligence index and coding-agent index scores at roughly the same level as GPT-5.6. On the coding-agent index specifically, Sol scored 57 on its max tier, two points higher than its predecessor, while Luna scored 41, two points lower than the model it replaces.
Where this release actually moves the needle, according to Artificial Analysis, is cost. Running the same intelligence-index benchmark suite, Sol costs $1.06 per task on its max tier, compared with $1.99 for the previous generation — nearly half. Luna comes in at $0.07 per task, down from $0.18.
Doing the math
Extrapolating loosely from Artificial Analysis's cost figures gives a sense of scale: a team running roughly 10,000 similar tasks a day would save around $9,300 daily by switching to Sol, which works out to nearly $280,000 over a month. Moving to a lighter-weight tier like Luna cuts costs even further — the same volume of daily tasks would drop from about $1,800 to $700.
That math is a direct extrapolation from benchmark cost data, not a guarantee — actual bills will depend heavily on task length, which reasoning tier a team chooses, and how often cached content gets reused. The cache-hit rate matters more than it might seem: for agent applications that keep resending the same system prompt turn after turn, each 10-percentage-point improvement in cache-hit rate can save more money than the headline price cut itself.
The timing lines up with a similar move from a rival: Anthropic released Claude Opus 5.5 on the same day, cutting its overall running cost by 40% compared with Opus 5, with pricing set at $4 per million input tokens and $20 per million output tokens. Both companies pushed flagship pricing down on the same day — Sol's per-token price works out to half of Opus 5.5's, while Luna pushes the lightweight-tier price point all the way down to $0.10.
For developers in China, how much of this price cut actually reaches them depends on which route they use to call the API. Teams that call OpenAI's official API directly get the discount immediately and in full. Those going through relay services or cloud-provider resale channels will have to wait and see whether those providers pass the same reduction through in their own markup — that's each provider's call to make, not OpenAI's. Agent products that bill by task cost are likely to feel the shift first in their margin structure: with inference spending cut in half for the same functionality, designs that previously throttled call volume just to keep costs under control now have room to loosen those limits — though rival products built on the same models can do exactly the same thing.
Sources: OpenAI release materials, Artificial Analysis benchmark report, CocoLoop, MarkTechPost; third-party benchmarks verified per-task cost and coding-agent index scores.