OpenAI Launches GPT-5.6 Sol Under U.S. Government Restrictions

OpenAI Unveils GPT-5.6 Amid Regulatory Tensions

OpenAI has released the GPT-5.6 model suite, comprising Sol, Terra, and Luna, in a limited preview that follows a direct request from the Trump administration. The move, announced on June 26, 2026, marks a significant step in the ongoing tension between frontier AI development and national security concerns. The company is rolling out the models to a small group of trusted partners, with government approval required for each participant.

CEO Sam Altman informed staff in a memo that the government requested a staggered release, approving customers one by one during the preview period. This echoes the recent launch of Anthropic's Mythos, which faced similar restrictions. OpenAI has made clear that this arrangement is not its preferred long-term model, calling it a "short-term step" toward broader availability in the coming weeks.

Three Tiers of Intelligence: Sol, Terra, and Luna

The GPT-5.6 family introduces a new naming system where the number denotes the generation, and the tier names—Sol, Terra, and Luna—indicate durable capability levels. Sol is the flagship model, designed for the most demanding tasks. Terra offers a balanced option for everyday work, and Luna is optimized for speed and affordability.

OpenAI claims Sol is its strongest model yet, with improved agentic capabilities in coding, biology, and cybersecurity. It introduces a max reasoning effort mode for deep analysis and an ultra mode that coordinates multiple sub-agents to tackle complex tasks. This architecture allows Sol to outperform previous models on benchmarks like Terminal-Bench 2.1 for coding workflows and GeneBench v1 for biology analyses.

Benchmark Performance and Competitive Edge

GPT-5.6 Sol sets a new state of the art on Terminal-Bench 2.1, which tests command-line workflows requiring planning and tool coordination. On ExploitBench, Sol is competitive with Anthropic's Claude Mythos Preview while using only about one-third of the output tokens. This efficiency gain is a key differentiator in the cybersecurity domain.

The model also shows strong performance on ExploitGym, a benchmark created by UC Berkeley researchers in collaboration with OpenAI. Terra and Luna also demonstrate improvements in cyber capabilities as reasoning increases. OpenAI emphasizes that Sol is better at finding and fixing vulnerabilities than executing end-to-end attacks, positioning it as a defensive tool rather than an offensive weapon.

Safety Stack: Layered Defenses Against Misuse

OpenAI has implemented its most robust safety stack to date with GPT-5.6. The system includes model-level training to refuse prohibited cyber assistance, real-time classifiers that evaluate outputs during generation, and account-level review for persistent malicious behavior. For high-risk cases, a larger reasoning model can pause generation and assess context before releasing output.

The company dedicated over 700,000 A100-equivalent GPU hours to automated red-teaming, focusing on finding universal jailbreaks that work across many prompts. This effort is complemented by human expert red-teaming, which continues during the preview. OpenAI maintains a rapid-response process to address newly discovered vulnerabilities and update safeguards.

Despite these measures, GPT-5.6 Sol does not cross the Cyber Critical threshold under OpenAI's Preparedness Framework. In evaluations involving Chromium and Firefox, it identified bugs but did not autonomously produce a functional full-chain exploit. However, the company acknowledges that benchmarks cannot capture every potential misuse scenario.

Government Access and Industry Implications

The preview is limited to approximately 20 companies whose participation has been shared with the U.S. government. OpenAI has been previewing GPT-5.6 with the administration for the past month, including meetings between Altman and White House officials in early June. The company expects to expand access to more organizations next week, with a broader release targeted in the coming weeks.

OpenAI's stance on the restrictions is clear: "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." The company is working with the administration to develop a cyber Executive Order framework and a repeatable process for future model releases.

By August, as part of the Executive Order, the administration must establish a classified process to assess AI models' cyber capabilities and determine which qualify as "covered frontier models." This regulatory backdrop is shaping the release strategies of major AI labs, including OpenAI and Anthropic.

Pricing and Availability

GPT-5.6 is priced per 1 million tokens across the three tiers: Sol at $5 input / $30 output, Terra at $2.50 input / $15 output, and Luna at $1 input / $6 output. The models introduce more predictable prompt caching with explicit cache breakpoints and a 30-minute minimum cache life. Cache writes are billed at 1.25x the uncached input rate, while reads receive a 90% discount.

OpenAI also plans to launch GPT-5.6 Sol on Cerebras hardware in July, offering speeds up to 750 tokens per second. Access will initially be limited to select customers as capacity expands. During the preview, models will be available through the API and Codex, with broader ChatGPT integration expected soon.

The Road Ahead

The preview period is designed to test both model capabilities and safeguard effectiveness. OpenAI wants to understand whether legitimate users can complete normal work reliably while the system constrains misuse. Feedback will inform adjustments to reduce unnecessary blocks and delays before wide release.

Enterprise customers are being offered longer-term approaches, including privacy-preserving detection and customer-operated safety controls. OpenAI is also working on access calibrated to the risk profile of each customer, user, or workload. The company's goal is a broad release in the coming weeks, pending successful testing and government approval.

This release represents a critical test case for the relationship between frontier AI development and national security. As models become more capable, the balance between innovation and regulation will continue to shape the industry. OpenAI's experience with GPT-5.6 may set a precedent for how future models are launched in a regulated environment.