OpenAI has announced the next generation of its model series, GPT-5.6. However, the rollout is not immediate for the general public, but will instead begin with a limited preview phase for a small group of trusted companies and partners.

The new model family consists of three different versions, each focusing on different aspects of intelligence, speed, and cost: Sol as the flagship model, Terra as a balanced version for everyday work, and Luna as a fast and cost-effective option. According to the manufacturer, the Terra variant is intended to offer performance comparable to the predecessor GPT-5.5 while costing half as much.

Strategic Release and Government Coordination

A key aspect of the launch is the controlled distribution. OpenAI stated that preview access via API and Codex is currently limited to selected partners. This step is being taken in close coordination with the U.S. government. The goal is to test the models’ capabilities before a broader release and to work with authorities on a framework for cybersecurity model safety.

Although access is currently highly limited, OpenAI plans to make the models broadly available to the general public via ChatGPT, Codex, and the API in the coming weeks.

Enhanced Capabilities in Specialized Fields

Technically, the GPT-5.6 series is characterized by improved “agentic” capabilities. In benchmark tests, it showed increased performance in the areas of programming (coding), biology, and cybersecurity.

New features include:

• Max Reasoning Effort: A function that allows the model more time to analyze complex problems deeply.
• Ultra Mode: A mode that accelerates the completion of complex tasks by utilizing sub-agents that go beyond the capacity of a single agent.
• Specialized Benchmarks: In tests such as Terminal-Bench 2.1 (for command-line workflows) and GeneBench v1 (for genomics analysis), the Sol model achieved new high scores.

Focus on Safety and Cybersecurity

Due to the increased proficiency in sensitive areas such as cybersecurity, OpenAI has implemented a multi-layered safety system. This includes protection measures integrated into the model training, real-time checks during generation, and account-level controls.

The company emphasizes that the models have been trained to refuse unauthorized assistance with cyberattacks, while continuing to support legitimate tasks such as security research or patch development. To test the robustness of the systems, over 700,000 A100 GPU hours were dedicated to automated “red teaming” to identify potential vulnerabilities and manipulation attempts (jailbreaks) at an early stage.

Pricing and Availability

The pricing structure for API usage is based on the number of processed tokens (per 1 million tokens):

• Sol: $5 (Input) / $30 (Output)
• Terra: $2.50 (Input) / $15 (Output)
• Luna: $1 (Input) / $6 (Output)

Additionally, OpenAI announced that it will launch the Sol model on Cerebras hardware starting in July, which is expected to enable speeds of up to 750 tokens per second. Access to this high-speed option will initially also be limited to select customers-.