OpenAI has started rolling out GPT-5.6, the newest generation of its large language models. It comes as three models, Sol, Terra, and Luna, that let users pick based on their needs and budget, with the aim of improving efficiency in enterprise work, coding, and scientific research. The models are available through ChatGPT, Codex, and the API, and a new tool for office tasks launched alongside them. This article walks through how the three tiers differ, their performance, and their pricing.

Three tiers chosen by use case and budget

GPT-5.6 bundles three models with distinct roles. Sol, the top tier, is the flagship built for the highest performance; Terra is the mid tier that handles everyday work with a good balance; and Luna is the lightweight tier for running large volumes of tasks at low cost. The names Sol, Terra, and Luna are Latin for the sun, the earth, and the moon, and the naming has been reorganized so that the number (5.6) marks the generation while the name marks the capability tier.

Users can choose among the three depending on whether they want to prioritize accuracy, keep costs down, or land somewhere in between. The expected pattern is Sol for work that involves difficult judgment, Terra for a broad range of daily tasks, and Luna for high-volume processing such as customer inquiries.

Coding efficiency and comparisons with rivals

What OpenAI emphasizes most this time is efficiency in coding. According to CEO Sam Altman, the flagship Sol is 54 percent more token efficient at coding than previous models. Handling the same work with fewer tokens not only speeds up responses but also makes the actual cost easier to contain.

On performance, OpenAI points to the third-party "Artificial Analysis Coding Agent Index," saying Sol scored 80, which is 2.8 points above Anthropic's top model Fable 5. The mid-tier Terra slightly edges out Fable 5, and the lightweight Luna outperforms Opus 4.8, OpenAI says. In each case, the company claims these levels were reached with fewer tokens and at a lower cost.

That said, these are comparisons presented by OpenAI itself. The announcement also lands amid intensifying competition for enterprise customers, with rivals releasing new models one after another, so the real-world experience is worth judging after independent evaluations and hands-on use in the field.

OpenAI also positions GPT-5.6 as its "strongest cybersecurity model yet." It says the model can support defensive work such as threat modeling, reviewing code, creating patches, and blue teaming, which tests defenses by simulating attacks.

A new office-task tool, "ChatGPT Work"

Alongside the refreshed lineup, OpenAI introduced a new tool for enterprise teams called ChatGPT Work. It helps with day-to-day clerical tasks such as drafting documents, spreadsheets, and presentations, and can be used on desktop, web, and mobile.

The move suggests an ambition to expand generative AI from personal research and writing toward being embedded in team-level work.

Availability and pricing

GPT-5.6 is available through ChatGPT, the coding assistant Codex, and the OpenAI API. Via the API, developers can call the three tiers separately and switch models according to how heavy a task is.

API pricing is set per 1 million tokens and varies widely by tier. The flagship Sol costs 5 USD (about 810 yen) for input and 30 USD (about 4,860 yen) for output. The mid-tier Terra costs 2.50 USD (about 400 yen) for input and 15 USD (about 2,430 yen) for output. The lightweight Luna costs 1 USD (about 160 yen) for input and 6 USD (about 970 yen) for output. By choosing a tier, users can avoid paying for more capability than they need and optimize their costs.

※1 USD = 162 JPY (as of July 11, 2026)

Summary

GPT-5.6 is a new model family split into three tiers, the flagship Sol, the mid-tier Terra, and the lightweight Luna, that lets users choose their own balance between performance and cost. With improved token efficiency in coding, performance claims said to beat rivals, and the launch of the ChatGPT Work office tool, the release is clearly built with enterprise use in mind. Because the performance figures rest on OpenAI's own comparisons, the true strength will become clear through independent evaluations to come.