OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, roughly 90 minutes after Anthropic shipped Claude Opus 5.5. Sol costs 2 dollars per million input tokens and 10 dollars per million output tokens. Luna costs 0.10 and 0.50. Both are about half the price of the GPT-5.6 models they replace, and OpenAI told VentureBeat the pricing is permanent rather than a launch promotion.
Where do Sol and Luna sit in the lineup?
Beneath GPT-6 Astra, the 10 dollar / 50 dollar flagship released September 3 alongside a company statement that we are now in the AGI era. Until this launch, Astra was the only GPT-6 model, and developers doing everyday work were still on GPT-5.6. Sol is positioned for complex coding and agentic workflows; Luna is built for focused tasks that need to run cheaply at very high volume. OpenAI says both were trained with methods similar to Astra’s.
Both carry a 1.05 million-token context window with input capped at 922K, 128K max output tokens, text and image input, and OpenAI’s agent tooling including function calling, web search, file search, computer use, and MCP connections. Model IDs are gpt-6-sol and gpt-6-luna. One oddity: Luna has a newer knowledge cutoff (May 18, 2026) than either Sol (April 20) or Astra (April 30), unusual for the cheapest model in a family.
Are they actually better, or just cheaper?
Mostly cheaper, and OpenAI is fairly direct about that: its own launch page states that Astra remains its best model across the board. The pitch is cost efficiency, not a new capability ceiling.
The numbers support a mixed reading. On AutomationBench, Sol at xhigh effort scores 33.2 percent at 0.27 dollars per task, beating Claude Opus 5 at max (26.9 percent at 11.1 times the cost) and GPT-6 Astra at low (30.3 percent at 3.9 times the cost). On Agents’ Last Exam, Sol at max scores 56.4 percent against Opus 5’s best of 55.9 percent. But on DeepSWE and OSWorld 2.0, GPT-6 Sol’s best scores (68.8 percent and 64.4 percent) fall below GPT-5.6 Sol’s best (72.7 percent and 66.2 percent) and below Claude Opus 5’s best (73.7 percent and 70.2 percent). On two of six headline evaluations, the new model scores lower than the one it replaces while costing roughly 60 percent less per task.
One comparison deserves scrutiny. OpenAI says Sol can match Claude Fable 5.1 xhigh at much lower cost, and it does: 49.3 percent for 2.14 dollars against 48.7 percent for 9.27 dollars. But xhigh is Fable 5.1’s weakest setting on that chart. Fable 5.1 at low scores 49.8 percent for 2.38 dollars, which is both a higher score and a near-identical price.
What about the accuracy claim?
OpenAI says GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol, approaching Astra-level reliability at much lower cost. That figure comes from an internal factuality evaluation using de-identified real-world conversations where users flagged model errors. It is a meaningful methodology, but it is OpenAI’s own test rather than an independent benchmark, and the 50 percent price-cut headline is measured against GPT-5.6 promotional rates, so compare it with what you actually paid.
Why Luna may matter more than Sol
At 0.10 dollars per million input tokens and 0.50 per million output, Luna’s output price fell further than the 50 percent headline suggests, down from 1.20 dollars. That puts classification, extraction, routing, and other high-volume work into territory where inference cost stops being the thing that decides whether a feature can ship profitably. It is the same competitive pressure driving DeepSeek’s aggressive V4.1 Flash pricing and the enterprise shift toward cheaper models documented in our report on AT&T routing 40 percent of its AI traffic to open models.
Availability is narrower than usual at launch. Both models are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go subscribers can reach Luna through the ChatGPT desktop app. Neither is in the regular Chat interface yet.
See VentureBeat’s launch coverage for more.
