OpenAI began the general rollout of its new large language model family, GPT-5.6, on July 9. The lineup consists of three models: the flagship Sol, the everyday-work Terra, and the fast, low-cost Luna. They are rolling out in stages across ChatGPT, the business-oriented ChatGPT Work, the Codex coding assistant, and the developer API. General availability follows a limited preview that ran through late last month, and it was cleared for wider access after a review by the US government. Below is a rundown of how the three models differ, along with the key points on their published performance and pricing.

Three Models Chosen by Use Case

GPT-5.6 comes in three tiers that differ in performance and price. The top-end Sol is the flagship, aiming for state-of-the-art results across a wide range of tasks including coding, research, document creation, cybersecurity, and science. The mid-tier Terra is positioned to handle everyday work with a good balance, while the entry-level Luna suits situations where fast responses and lower costs matter most.

According to OpenAI, Sol can match or exceed the results of earlier frontier models while using fewer tokens (the smallest units used to process text), which raises the performance delivered per dollar spent. Rather than using one large, do-everything model for all tasks, the design lets users pick a model that fits the weight of the job at hand.

Performance Shown in Benchmarks

To back up its performance claims, OpenAI published several figures. On Agents' Last Exam, which measures the ability to carry out long-running professional tasks, Sol scored 53.6 percent, beating Anthropic's Claude Fable 5 by 13.1 points. The mid-tier Terra also outperforms Fable 5 at a substantially lower cost, and OpenAI says the entry-level Luna surpasses Claude Opus 4.8 on coding tasks while costing roughly a quarter as much.

That said, all of these figures were published by OpenAI itself. Real-world usefulness varies with the task and conditions, so it makes sense to weigh them alongside third-party verification.

More Tools for Developers

For developers, new features arrive to complete complex work faster. The headline addition is a setting called "ultra," which runs four agents in parallel so they can divide up a task and tackle it together.

Another is Programmatic Tool Calling, available through the API. Instead of returning each external tool call to the developer one at a time, it lets the model write and run a lightweight program on the spot to coordinate multiple tools and filter intermediate results. Because it cuts down on back-and-forth exchanges, it makes multi-step automation easier to build. A new beta for handling multiple agents was also added, letting developers assemble ultra-style coordination on their own.

Stronger Cyber Capabilities, and the Limits That Come With Them

OpenAI describes GPT-5.6 as having its strongest cybersecurity capabilities to date, citing a score of 73.5 percent on ExploitBench2, a benchmark that measures how exploitable systems are.

That strength is a double-edged sword. OpenAI plans to restrict access to its most cyber-capable models through a framework called Trusted Access for Cyber. Eligible users must enable Advanced Account Security, which uses hardware-backed passkeys, by September 1, or they will lose continued access to these models. It is a way of asking users to adopt a baseline of security settings so that highly capable models are protected from misuse or account takeover.

A Wider Range That Reaches Document Creation

Beyond coding, OpenAI highlights improved quality in document creation. GPT-5.6 can read the "design system" from a set of reference slides—layouts, typefaces, spacing, color schemes, and Slide Master rules—and apply it consistently to new material. That makes presentations, documents, and spreadsheets come out more polished.

In addition, its computer-use capabilities have improved, so the model can inspect the output it has produced and re-check it for visual or functional problems before delivering the final result.

Availability and Pricing

The GPT-5.6 family rolls out worldwide over 24 hours. API pricing is set per million tokens: Sol costs 5 USD (about 810 yen) for input and 30 USD (about 4,860 yen) for output; Terra is 2.50 USD (about 405 yen) for input and 15 USD (about 2,430 yen) for output; and Luna is 1 USD (about 162 yen) for input and 6 USD (about 970 yen) for output. What is available depends on the subscription plan across ChatGPT, ChatGPT Work, Codex, and the API.

*1 USD = 162 JPY (as of July 10, 2026)

For the flagship Sol, OpenAI also plans a high-speed offering on Cerebras hardware at up to 750 tokens per second, suggesting broader reach into use cases where response speed is paramount.

Summary

The defining trait of GPT-5.6 is a shift in how the models are offered—from using one all-purpose model to choosing among Sol, Terra, and Luna based on the weight of the job. The benchmark figures are all self-reported and should not be taken at face value, but the release puts front and center a combination of capability and price across coding, document creation, and the cyber domain. How much the developer-facing tools such as parallel agents and programmatic tool calling actually help in real work remains to be seen, and attention will turn to third-party verification and how the models are used in the field.