GPT-6 Sol and Luna: API Prices, ChatGPT Access and Benchmarks

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, two faster and cheaper members of the GPT-6 family that launched with GPT-6 Astra earlier this month. In the API, GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, and GPT-6 Luna costs $0.10 and $0.50, half the price of the GPT-5.6 models they replace. In ChatGPT, both are rolling out in ChatGPT Work and Codex for paid plans, with Free and Go users getting Luna in the desktop app.

What are GPT-6 Sol and GPT-6 Luna?

OpenAI now sells three GPT-6 models at different price and capability tiers. GPT-6 Astra, released on September 3, is the top model. GPT-6 Sol is the mid-tier model “for complex coding and agentic workflows,” according to its model page, and GPT-6 Luna is described as OpenAI’s “most efficient model for focused, high-volume tasks.” OpenAI’s own developer guide says to use Astra for the highest capability, Sol “for strong reasoning on demanding tasks,” and Luna “for efficient, repeatable work at scale.”

In its GPT-6 Sol and Luna announcement, OpenAI says both models were trained “with similar methods as GPT-6 Astra,” bringing Astra’s improvements in professional work, factuality, coding, computer use and alignment to cheaper models. Improvements in caching and inference, the company says, let it serve them at lower cost.

The naming continues the scheme OpenAI introduced with GPT-5.6, where Sol was the flagship tier and Luna the small, cheap tier. GPT-6 Sol and Luna are the direct successors to GPT-5.6 Sol and GPT-5.6 Luna.

How much do GPT-6 Sol and Luna cost in the API?

OpenAI’s headline is a 50% cut compared with the promotional prices of the previous generation. The full rate card, taken from the official GPT-6 Sol model page and GPT-6 Luna model page, is below alongside GPT-6 Astra for context.

A few pricing rules apply across the GPT-6 family and are easy to miss:

• Long prompts cost more. Requests with more than 272,000 input tokens are billed at twice the input and cache rates and 1.5 times the output rate for the entire request.
• Caching is cheap, writing to cache is not free. Cached input tokens cost 10% of the normal input rate; cache writes are billed at 1.25 times the uncached rate.
• Batch and Flex are half price; Fast mode is double the applicable rate.
• Regional processing adds 10% where available, and EU data residency works only with Standard processing.

OpenAI says it has also improved prompt caching for GPT-6 so that more requests hit the cache by default, and developers can now set explicit breakpoints to choose where cached prefixes end. Changing reasoning effort or switching tools on and off mid-conversation no longer breaks the cache, according to the company.

How does that compare with Claude Opus 5.5?

Anthropic launched Claude Opus 5.5 the same day at $4 input and $20 output per million tokens, with cache reads at $0.20. On list price alone, GPT-6 Sol is half the cost of Opus 5.5 per token and matches Anthropic’s Claude Sonnet 5 at $2 and $10. Per-token price is only part of the picture, though, because both companies argue their new models use fewer tokens per task; real costs depend on your workload.

GPT-6 Sol and Luna specs

Unlike GPT-6 Astra, both new models support a none reasoning effort for fast, non-reasoning replies. OpenAI notes that Chat Completions supports function calling on these models only with reasoning set to none; for built-in tools such as web search, file search, code interpreter, computer use and MCP, it points developers to the Responses API. Neither model is offered for fine-tuning, according to the endpoint tables on their model pages.

Can you use GPT-6 Sol and Luna in ChatGPT?

Yes, but not everywhere yet. OpenAI says GPT-6 Sol and Luna are available “in ChatGPT Work and Codex starting today for all Plus, Pro, Business, Enterprise, and Edu users.” Free and Go users can access GPT-6 Luna in the desktop app. The company adds that “these models are not yet available in Chat,” meaning the standard conversational mode, and that the rollout would happen gradually through launch day.

That split matters for everyday users. On the current ChatGPT pricing page, the free plan still advertises unlimited text chats with GPT-5.6 Luna, while Plus ($20 a month) promises “advanced reasoning models with GPT-6” and Pro (from $100 a month) adds Pro reasoning powered by GPT-6 Astra.

How good are they? OpenAI’s benchmark claims

All results below are as reported by OpenAI in its launch post. Competitor figures come from what OpenAI calls “publicly available reports,” and in some cases OpenAI used Claude Fable 5 scores where Fable 5.1 results were not available.

Professional work

On AutomationBench 1.0.6, which tests end-to-end business workflows using 47 tools, OpenAI says GPT-6 Sol at xhigh effort scored 33.2% at $0.27 per task. It lists GPT-6 Astra at low effort at 30.3% for 3.9 times Sol’s cost, Claude Opus 5 at max effort at 26.9% for 11.1 times the cost, and Claude Fable 5.1 with Opus 5 fallback at 31.4%. On Agents’ Last Exam, GPT-6 Sol at max effort scored 56.4%, which OpenAI says is above Claude Opus 5’s best result at 60% lower cost per task.

Coding

On DeepSWE v1.1, OpenAI reports 68.8% for GPT-6 Sol at max effort, within 1.1 points of Claude Fable 5’s best score of 69.9%, at about 80% lower cost per task. GPT-6 Luna at max effort scored 66.6%, which OpenAI describes as comparable to Claude Opus 5 and Fable 5 at medium effort, at 93% and 96% lower cost per task respectively. On FrontierCode, OpenAI says Sol can match Claude Fable 5.1 at xhigh effort at much lower cost.

Computer use and factuality

On the offline set of OSWorld 2.0, GPT-6 Sol at xhigh effort scored 60.5% against 60.3% for Claude Opus 5 at medium effort, at about 80% lower cost per task, according to OpenAI. On an internal factuality test built from de-identified ChatGPT conversations where users flagged errors, OpenAI says GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol, and that Luna at higher effort levels matches GPT-5.6 Sol “at about a hundredth its cost.”

It is worth noting what is missing: OpenAI’s comparisons use Claude Opus 5 and Fable 5 or 5.1, not Claude Opus 5.5, which Anthropic released the same day. Anthropic’s own launch table, in turn, compares Opus 5.5 against GPT-6 Astra and GPT-5.6 Sol rather than the new Sol. Head-to-head numbers between the two new releases will have to come from independent testing.

What else is new in GPT-6 Sol and Luna

Beyond price, OpenAI highlights three changes developers will notice:

• Writing style: Sol and Luna inherit Astra’s communication style, which OpenAI says means “more clarity, less jargon, fewer odd turns of phrase” and slightly shorter answers.
• GPT-6 platform features: OpenAI’s Using GPT-6 guide lists async tool calling, mid-turn steering over WebSocket connections and the ability to change reasoning effort mid-conversation while keeping the prompt cache.
• Alignment: OpenAI says both models improve on their GPT-5.6 counterparts in its alignment evaluations, including lower rates of misleading claims about their coding work. Full results are in the system card.

OpenAI also cited GitHub, which it says has seen caching improvements over recent months cut the share of prompt tokens needing fresh processing by more than 50% across billions of requests to OpenAI models.

How to try GPT-6 Sol and Luna

1. In ChatGPT: on a Plus, Pro, Business, Enterprise or Edu plan, open ChatGPT Work or Codex and select GPT-6 Sol or GPT-6 Luna. OpenAI says that if the models do not appear yet, users should try again later because the rollout is gradual.
2. On the free or Go plan: install the ChatGPT desktop app, where OpenAI says GPT-6 Luna is available to Free and Go users.
3. In the API: set the model field to gpt-6-sol or gpt-6-luna in a Responses API request. Both default to medium reasoning effort, so set reasoning.effort explicitly if you want faster none or low replies, or deeper high, xhigh or max reasoning.
4. Watch the bill: OpenAI’s Prompt Caching Dashboard and diagnostics tool show how much of your input is being served from cache, which is where the largest savings come from at $0.20 and $0.01 per million cached tokens.

The API rollout is listed in OpenAI’s API changelog under September 22, which confirms the standard prices apply to prompts of up to 272,000 input tokens.

Should developers switch from GPT-5.6?

For most API users on GPT-5.6 Sol or Luna, the case is straightforward on price: the GPT-6 versions cost half as much per token and, by OpenAI’s account, make fewer mistakes. GPT-6 Sol also has a newer knowledge cutoff than GPT-5.6 Sol: April 20, 2026, versus February 16, 2026. The model IDs are new, so the move is a deliberate code change rather than an automatic upgrade. GPT-5.6 Sol remains available, and its model page says its promotional $4 and $20 pricing runs at least through November 21, 2026.

Teams should test before switching production traffic. Long-context workloads above 272,000 input tokens get the 2x input surcharge on every GPT-6 model, and applications that rely on Chat Completions with function calling need reasoning effort set to none or a move to the Responses API.

What happens next

OpenAI has not said when GPT-6 Sol and Luna will reach the standard Chat mode in ChatGPT, only that they are available in Work and Codex first. With Anthropic promising Claude Sonnet 5.5 and Haiku 5.5 in the coming weeks, the price cuts are unlikely to stop here. For now, GPT-6 Luna at $0.10 per million input tokens costs a tenth of Anthropic’s cheapest current model, Claude Haiku 4.5, on input, and GPT-6 Sol undercuts Claude Opus 5.5 on list price by half.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage