OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, as lower-cost tiers below GPT-6 Astra. Both cost half what their GPT-5.6 predecessors did: GPT-6 Sol at $2 input and $10 output per million tokens, and GPT-6 Luna at $0.10 input and $0.50 output. A week later, on September 29, OpenAI followed with a newer Sol model, GPT-6.1 Sol, so for most readers this release matters mainly because of Luna.
In its announcement, OpenAI says it trained Sol and Luna with methods similar to those used for GPT-6 Astra, which it had introduced earlier in September. The two models fill out a three-tier GPT-6 family:
The API identifiers are gpt-6-sol and gpt-6-luna. In ChatGPT, OpenAI made both available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and gave Free and Go users access to Luna in the desktop app. OpenAI said neither model was yet available in Chat at launch.
| Per million tokens (USD) | GPT-5.6 price | GPT-6 price |
|---|---|---|
| Sol input / output | $4 / $20 | $2 / $10 |
| Luna input / output | $0.20 / $1.20 | $0.10 / $0.50 |
Source: OpenAI announcement, September 22, 2026. OpenAI describes the GPT-5.6 figures as promotional pricing.
The GPT-6 Luna model page adds the details that matter for budgeting, as of October 10, 2026:
At these prices, a request with 100,000 input tokens and 2,000 output tokens costs about one cent on Luna at standard rates, against about 22 cents on GPT-6.1 Sol and $1.10 on Astra. That spread is why the tier choice matters more than small benchmark differences for high-volume pipelines.
OpenAI announced an improved prompt caching system for the GPT-6 family the same day, described in a separate post, Better prompt caching for GPT-6. The points that affect agent design:
OpenAI reports results on AutomationBench, Agents' Last Exam, FrontierCode, DeepSWE v1.1 and OSWorld 2.0, mostly comparing GPT-6 Sol with Anthropic's Claude Opus 5 and Claude Fable 5 and 5.1. On DeepSWE v1.1, OpenAI says GPT-6 Luna at maximum effort scored 66.6%, which it describes as comparable to Claude Opus 5 and Fable 5 at medium effort, at 93% and 96% lower cost per task respectively. On its internal factuality evaluation, OpenAI says Luna at higher effort matches GPT-5.6 Sol at about a hundredth of the cost.
These are OpenAI's numbers. The announcement says competitor results came from public reports, and that Fable 5 scores were used where Fable 5.1 scores were unavailable. The comparisons also predate Claude Opus 5.5, which Anthropic announced the same day, and Claude Sonnet 5.5, announced on September 28; see the Opus 5.5 and Sonnet 5.5 guides. This article does not reproduce them.
High-volume pipelines such as classification, extraction, routing and summarization are the natural fit. At $0.10 input and $0.50 output per million tokens, Luna competes with Claude Haiku 5.5, which lists at the same $0.10 and $0.50 for prompts up to 100,000 tokens, per Anthropic's Haiku 5.5 announcement and pricing documentation. See the Claude Haiku 5.5 guide.
Agent systems with a router can use Luna for simple steps and escalate to GPT-6.1 Sol or Astra when a step fails or needs deeper reasoning. The caching changes, which keep the cache warm across effort and tool changes, make that pattern cheaper than it was on GPT-5.6.
Teams on GPT-5.6 Luna can expect lower list prices. Re-test outputs, since the model generation changed.
gpt-6.1-sol rather than gpt-6-sol.gpt-6-luna and re-run your evaluation set.When were GPT-6 Sol and GPT-6 Luna released?
September 22, 2026. GPT-6.1 Sol followed on September 29, 2026.
How much does GPT-6 Luna cost?
$0.10 input, $0.01 cached input, $0.125 cache writes and $0.50 output per million tokens, per OpenAI's model page (checked October 10, 2026).
Should I use GPT-6 Sol or GPT-6.1 Sol?
GPT-6.1 Sol. It is the newer model at the same input and output prices, with cheaper cached input.
Discover more content: