OpenAI Launches GPT-5.6 with Sol, Terra, and Luna Models

On June 26, 2026, OpenAI announced the limited preview release of its next-generation large language model suite, GPT-5.6. This release came less than 24 hours after reports surfaced that the Trump administration had requested a phased approach to the public rollout of next-generation AI models, marking a highly unusual sequence of events.

Rather than a single model, GPT-5.6 is a suite comprising three distinct models: the flagship “Sol,” the mid-range “Terra,” and the fast and low-cost “Luna.” OpenAI has stated that these models excel particularly in coding, cybersecurity, and biology, and are designed to maintain focus over extended periods in agent-based AI tasks.

Pricing Structure and Competitor Comparison

The Sol model is priced at $5 per million input tokens and $30 per million output tokens. This cost is approximately half of that for Anthropic’s flagship model, “Claude Fable 5,” which charges $10 per million input tokens and $50 per million output tokens. Meanwhile, the Terra model is priced at half of Sol’s cost, and Luna is priced at less than half of Terra’s cost, offering a tiered pricing structure.

This pricing strategy clearly aims to provide a competitive edge. The significant reduction in inference costs lowers the barriers to entry for enterprises and developers alike. This is particularly beneficial for operations requiring high-frequency API calls, directly reducing their overall expenses.

New Modes: Max and Ultra

Sol introduces two groundbreaking inference modes: “Max” and “Ultra.” Max mode enables deeper reasoning capabilities, while Ultra mode leverages sub-agents for task execution. The Ultra mode is reminiscent of Anthropic’s “OpenClaw” technology, a similarity attributed to Peter Steinberger, the creator of OpenClaw, who recently joined OpenAI.

By utilizing sub-agents, complex tasks can be broken into smaller processes and executed simultaneously. This innovation is expected to improve efficiency in long-term inference tasks such as analyzing large codebases or conducting multi-stage security audits.

A Strong Emphasis on Safety

A significant portion of OpenAI’s announcement blog focused on safety and the prevention of misuse. Amid rising political tensions over cybersecurity in Washington, the company stated, “GPT-5.6 is trained to reject prohibited cyber assistance requests, even when users attempt to disguise their intentions or bypass safeguards.”

OpenAI also noted that Sol is better suited for identifying and fixing vulnerabilities rather than executing end-to-end cyberattacks. The company clarified that Sol does not exceed the critical cybersecurity thresholds outlined in OpenAI’s internal preparedness framework. However, this framework was revised in April 2026, removing certain previously included research domains.

For safety verification, OpenAI invested approximately 700,000 A100e GPU hours in automated red-teaming and collaborated with third-party testers. The third-party testing phase is set to continue for another two weeks.

Careful Management of the Preview Period

OpenAI has adopted a cautious approach for the preview period, particularly considering the close scrutiny of this release by the Trump administration. The company explained, “In dual-use areas where defensive and offensive activities initially appear similar, safeguards may sometimes interfere with legitimate operations. Addressing this balance is part of the purpose of the preview testing.”

This statement acknowledges the trade-off between safety and usability in advance. While excessive safeguards could harm user experience, they represent a realistic compromise to avoid straining relations with the administration.

Editorial Opinion

In the short term, GPT-5.6’s competitive pricing is likely to significantly impact the enterprise AI market. By setting Sol’s output costs at 60% of Claude Fable 5’s, mid-sized businesses and startups wary of high costs may find it easier to adopt large language models. On the other hand, the fact that the release timing was directly influenced by coordination with the Trump administration hints at the potential direction of future U.S. AI regulations. The creation of a precedent for federal intervention in AI companies’ release schedules could lead to a reassessment of risk management strategies across the industry.

From a long-term perspective, the broader framework for evaluating AI safety is entering a critical stage. OpenAI’s decision to revise its preparedness framework in April to relax certain safety standards indicates that internal benchmarks can shift under political or competitive pressure. Furthermore, the two-week duration set for third-party testing raises questions about whether sufficient scrutiny will be applied. While the introduction of the Ultra mode, which leverages sub-agents, represents a leap forward in AI autonomy, debates around its controllability remain unresolved.

As a publication, we believe the following questions should be addressed:

1. How will the industry adapt to increasing government involvement in AI development timelines?
2. Can OpenAI’s safety measures, including its updated framework, withstand external pressures and ensure robust risk management?
3. Will Ultra mode’s enhanced autonomy introduce new risks, and how will these be mitigated?

References

• The Verge — Published on 2026-06-26