OpenAI releases the GPT-5.6 family with three models and an ultra mode

OpenAI released the GPT-5.6 family with three models and ultra mode

On July 9, 2026, OpenAI released the GPT-5.6 family to general availability. The lineup includes the flagship model GPT-5.6 Sol, the balanced GPT-5.6 Terra, and the most affordable GPT-5.6 Luna; the models became available in ChatGPT, Codex, and the OpenAI API with gradual global rollout over the next 24 hours.

Which models are in the lineup

The family consists of three performance tiers. Sol is designed for the most complex tasks: programming, professional analysis, scientific research, cybersecurity, and long-running agent scenarios. Terra offers a balance of capabilities and cost, while Luna is aimed at fast and large-scale workflows.

• GPT-5.6 Sol costs $5 per million input tokens and $30 per million output tokens
• GPT-5.6 Terra costs $2.50 per million input tokens and $15 per million output tokens
• GPT-5.6 Luna costs $1 per million input tokens and $6 per million output tokens
• Ultra mode by default coordinates four agents in parallel
• Programmatic Tool Calling and a beta version of multi-agent mode are available in the API

The core idea of the update is to get more useful work out of each token. According to OpenAI, GPT-5.6 Sol achieves comparable or higher results than previous and competing models while using fewer tokens and less time. Terra and Luna should make a similar level of automation accessible for everyday tasks.

What changed for developers

GPT-5.6 Sol has become OpenAI's most powerful model for programming. In the Artificial Analysis Coding Agent Index test, it scored 80 points, and in Terminal-Bench 2.1 — 88.8%. Ultra mode raised the Terminal-Bench result to 91.9%, coordinating multiple parallel workflows.

The model also received new capabilities for working with tools. In the Responses API, the Programmatic Tool Calling function allows GPT-5.6 to write and run small programs in memory, process intermediate data, and pass only important results back to the model. This reduces the number of model calls and the amount of context in tasks with many tools.

The max mode gives the model more time for reasoning, checking alternatives, and correcting errors. Ultra goes further: multiple agents execute parts of a complex task in parallel, after which GPT-5.6 combines their results. For developers, this means the ability to build more autonomous systems without manual scripting for each step.

Where the model shows improvement

Beyond code, OpenAI highlights work with documents, presentations, and spreadsheets. GPT-5.6 is better at following templates, slide structure, typography, indentation, and formatting rules. The model analyzes the rendered result and can fix visual and functional shortcomings, not just generate source code.

In the BrowseComp test, GPT-5.6 Sol with maximum reasoning level achieved 92.2%, and in OSWorld 2.0 — 62.6%. On ExploitBench 2, the model scored 73.5% compared to 47.9% for GPT-5.5 at a comparable output budget. In scientific tasks, OpenAI reports improvements in biology, natural sciences, and chemistry.

For cybersecurity, the company simultaneously strengthened restrictions. GPT-5.6 supports defensive scenarios — vulnerability analysis, code fixing, threat modeling, and blue teaming. Access to the most sensitive capabilities is available to verified users and organizations through the Trusted Access for Cyber program.

What this means

GPT-5.6 is not just an update to response quality, but an attempt by OpenAI to turn the model into a more independent executor of complex work. Three price points, tool calls from programs, and parallel agents should reduce the cost of long-running workflows, although maximum modes will require more tokens and computation.