OpenAI unveils GPT-5.6 Sol, Terra, and Luna in limited preview under agreement with US government

What happened?

OpenAI has announced the limited preview of its new GPT-5.6 model family, consisting of three variants: Sol (flagship model), Terra, and Luna. The release is exclusively through the API and Codex for a small group of partners and trusted organizations, after OpenAI presented its plans to the US government, as required by an executive order from President Donald Trump issued on June 2, 2026. That order establishes a 30-day process for federal agencies to collaborate on benchmarking the capabilities of new AI models, ensuring their safety before public release. The deadline expires on July 2, so the current deployment is provisional. This is the first time an AI company has explicitly coordinated model availability with the White House before public launch, marking a milestone in the relationship between AI developers and governments. According to VentureBeat, OpenAI stated that "we presented our plans and the models' capabilities before today's launch. At the request of the US government, we began with a limited preview for a small group of trusted partners."

Why is this important?

This launch marks a milestone in the relationship between AI developers and governments: for the first time, a company explicitly coordinates model availability with the White House before public release. Additionally, the GPT-5.6 architecture introduces significant innovations:

• Deep reasoning modes: The max mode grants extended compute time for complex problems, while the ultra mode deploys specialized sub-agents that divide and conquer long-range tasks. This multi-agent approach represents a paradigm shift from previous models that operated as a single entity. Instead of relying on instant token generation, GPT-5.6 dynamically allocates compute, allowing the model to "think" longer on complex problems.
• Superior performance: On Terminal-Bench 2.1, GPT-5.6 Sol (Ultra) achieves 91.91% in command-line automation, surpassing Claude Mythos 5 (88%). On Agent's Last Exam, it is the only model to exceed 50% (50.9% in code mode). On GeneBench v1, it outperforms GPT-5.5 using fewer tokens. These results position GPT-5.6 as the most powerful model in reasoning and agent tasks.
• Three capability tiers: Sol, Terra, and Luna offer different cost-performance ratios, with prices ranging from $5/$30 per million tokens (Sol) to more economical options for less demanding tasks. This segmentation allows businesses to choose the appropriate tier based on task complexity, optimizing costs.

The limited deployment also introduces a new compliance paradigm: enterprise buyers must navigate real-time safety interventions, mandatory compliance parameters, and structured token caching systems. This reflects an evolution in the relationship between AI providers and regulators, where oversight is integrated into the product itself.

What consequences will it have?

In the short term, only a select group of partners will be able to experiment with GPT-5.6's capabilities, potentially generating significant competitive advantages in areas like cybersecurity, development automation, and bioinformatics analysis. For example, in cybersecurity, the ultra mode could deploy sub-agents to detect vulnerabilities more efficiently than traditional methods. In the medium term, if the government evaluation process concludes favorably (expected by July 2), a broader public release is anticipated. This could accelerate the adoption of multi-agent architectures in enterprise environments, as well as set a regulatory precedent for future frontier model launches. However, dependence on government approval introduces uncertainty and could slow innovation if other countries adopt similar approaches. Moreover, the fact that OpenAI coordinated with the US government could influence policies in other nations, creating a domino effect in global AI regulation. Companies like Anthropic and Google DeepMind may feel pressured to follow a similar process, changing the dynamics of model releases.

What should readers know?

• The preview is limited and not accessible to the general public; only selected partners can access it via API and Codex.
• GPT-5.6 Sol's prices are identical to GPT-5.5 ($5/$30 per million tokens), but offer better performance on complex tasks, representing a significant improvement in cost-performance ratio.
• OpenAI has implemented a tiered reasoning system (max and ultra modes) that allows scaling computational effort according to task difficulty. The ultra mode, in particular, introduces a multi-agent approach that could redefine how complex problems are tackled in AI.
• The government review process could delay or modify the conditions of the public launch, expected after July 2, 2026. If the evaluation is unfavorable, OpenAI may be forced to implement additional restrictions or delay the launch.
• Compared to previous launches, such as GPT-4 in March 2023, which had no prior government review, this process marks a radical change. It also contrasts with Anthropic's approach of voluntary self-regulation, while OpenAI now undergoes mandatory government oversight.

"This coordination with the US government represents a seismic shift in how frontier AI models are deployed, merging technological innovation with real-time regulatory oversight." — Analyst at TheVortiq

In summary, GPT-5.6 not only offers significant technical advances but also sets a new standard in the relationship between the AI industry and governments. Companies looking to adopt this technology must prepare for a stricter compliance environment, but will also benefit from unprecedented capabilities in automation and reasoning. The outcome of the government evaluation on July 2 will be crucial in determining the immediate future of frontier AI.