OpenAI is expanding its GPT-6 lineup with two new models designed to bring much of the technology behind its flagship GPT-6 Astra to a wider range of workloads. GPT-6 Sol is aimed at demanding coding and agentic tasks, while GPT-6 Luna is positioned as the faster, more cost-efficient option for high-volume work.

OpenAI has introduced GPT-6 Sol and GPT-6 Luna, expanding the GPT-6 family just days after the company introduced GPT-6 Astra. Rather than positioning the new models as replacements for Astra at the very top of the lineup, OpenAI is using Sol and Luna to offer different balances of intelligence, speed and cost.

The two models build on advances developed for GPT-6 Astra, including improvements to alignment and inference efficiency. OpenAI says it has also made its caching and inference systems more efficient, allowing it to pass some of those savings on through lower API pricing.

That gives the GPT-6 family a clearer three-model structure. Astra sits at the top for the most demanding end-to-end work, Sol targets complex coding and agentic workflows, and Luna is designed for focused, repeatable tasks where speed and cost matter more. OpenAI’s current model guidance describes Sol as the balance between intelligence and cost, while Luna is intended for cost-sensitive, high-volume workloads.

GPT-6 Sol is the new middle ground

GPT-6 Sol is built specifically for complex coding and agentic workflows. It supports reasoning efforts ranging from none through max, giving developers more control over how much computation the model uses for a particular task.

The model has a 1.05-million-token context window and can generate up to 128,000 output tokens. It also supports image inputs alongside text and can work with tools including web search, file search, code interpreter, computer use, hosted shell, MCP and image generation through the Responses API.

For developers, GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens under standard pricing for prompts within the shorter context range. Cached input costs $0.20 per million tokens.

That’s a substantial reduction compared with GPT-5.6 Sol’s promotional pricing, which OpenAI had previously set at $5 per million input tokens and $30 per million output tokens. OpenAI says the new Sol and Luna models have API prices that are 50% lower than the GPT-5.6 promotional pricing.

GPT-6 Luna focuses on speed and scale

GPT-6 Luna takes a different approach. OpenAI describes it as its most efficient model for focused, high-volume tasks, making it the option intended for applications that need to process large numbers of requests without the cost of using a more capable model for every interaction.

Luna has the same 1.05-million-token context window and 128,000-token maximum output as Sol. It also supports text and image inputs, reasoning, function calling, structured outputs and the same broad set of Responses API tools.

The big difference is price. GPT-6 Luna costs just $0.10 per million input tokens and $0.50 per million output tokens under standard pricing, with cached input priced at $0.01 per million tokens.

That puts Luna in a very different cost category from Sol and Astra. OpenAI’s current model catalog specifically recommends Luna for cost-sensitive, high-volume workloads.

OpenAI is bringing Sol and Luna to ChatGPT, Codex and the API

GPT-6 Sol and Luna are available through the OpenAI API, while their ChatGPT rollout is focused on Work and Codex. OpenAI says the models are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users can try GPT-6 Luna in the desktop app.

The distinction is important because these models are separate from the models offered in the regular ChatGPT chat experience. OpenAI’s release notes specifically describe GPT-6 Sol and Luna in Work and Codex as a separate model availability track.

For developers, both models are available through the Responses API and Chat Completions API, with support for batch processing as well.

The launch also continues OpenAI’s broader shift toward models that can operate as part of agentic software rather than simply answering individual prompts. GPT-6’s developer guidance adds capabilities such as asynchronous tool calling and mid-turn steering, allowing applications to keep a model working with tools while also handling new instructions or independent parts of a task.

How Sol and Luna fit into GPT-6

The result is a GPT-6 family with three distinct tiers rather than one model attempting to cover every workload.

GPT-6 Astra remains OpenAI’s most capable model for the hardest end-to-end tasks. GPT-6 Sol is designed to bring strong reasoning to demanding coding and agentic workflows at a lower cost, while GPT-6 Luna targets speed, efficiency and large-scale usage.

OpenAI’s safety evaluation also places Sol and Luna below Astra’s capability level in several areas. The company’s GPT-6 Astra system card says both models are treated as having High capability in cybersecurity and biological/chemical domains, while neither reaches the High threshold for AI self-improvement. OpenAI says it has therefore applied the same set of safeguards detailed for GPT-5.6 Sol and Luna.

Interestingly, OpenAI’s published HealthBench results show that the newer models do not simply improve every benchmark across the board. GPT-6 Sol improved slightly over GPT-5.6 Sol on HealthBench Professional, while both Sol and Luna showed regressions on some other HealthBench evaluations. OpenAI attributes some of those results to substantially shorter default answers, which can reduce coverage of details evaluated by the benchmark.

That distinction matters because the GPT-6 rollout isn’t simply about chasing higher benchmark numbers. With Sol and Luna, OpenAI is also pushing the idea that model efficiency can be just as important as raw capability when AI systems are used repeatedly, embedded into software or allowed to perform longer workflows.

For developers building those systems, the pricing may ultimately be one of the most significant parts of this release. A model that is capable enough for a particular task but dramatically cheaper to run can make previously uneconomical AI applications much easier to deploy at scale.

And with GPT-6 Astra, Sol and Luna now occupying different positions in the lineup, OpenAI is making the GPT-6 generation look less like a single flagship model and more like a platform designed to cover everything from the hardest agentic workloads to millions of routine AI requests.