OpenAI launched GPT-5.6 on June 26 as a restricted preview coordinated with the US government, introducing three capability-tiered models, Sol, Terra, and Luna, with broader public availability planned within weeks pending regulatory framework development.

Key Highlights

• GPT-5.6 Sol, the flagship model, is priced at $5 per million input tokens and $30 per million output tokens, with Terra available at roughly half the cost and Luna positioned as the fastest, lowest-cost tier.

• Initial access is limited to approximately 20 trusted partner organizations at the US government's request, following a June 2 executive order establishing a framework for assessing frontier AI model capabilities.

• All three GPT-5.6 models have been classified as High capability in both cybersecurity and biological/chemical risk domains under OpenAI's Preparedness Framework, a first for smaller, lower-cost models in a family.

• OpenAI dedicated more than 700,000 A100-equivalent GPU hours to automated red-teaming ahead of launch, its most extensive pre-deployment safety effort to date.

OpenAI's GPT-5.6 release marks a structural shift in how the company packages and deploys frontier AI. Rather than a single flagship model, the series introduces three permanent capability tiers designed to serve distinct use cases. Sol is built for deep reasoning, advanced coding, agentic workflows, and cybersecurity research. Terra delivers performance broadly comparable to the prior GPT-5.5 flagship at roughly half the price, targeting everyday production workloads. Luna is the entry-level tier, optimized for high-volume, low-latency tasks at the lowest cost in the family.

Sol introduces two new inference configurations. A max reasoning effort mode grants the model extended time to work through complex problems. An ultra mode goes further, deploying specialized subagents to divide and execute multi-step, long-horizon tasks in parallel. On Terminal-Bench 2.1, which tests planning and tool use in command-line environments, GPT-5.6 Sol in ultra mode achieved a score of 91.91%, edging ahead of Anthropic's Mythos 5 at 88%.

The release was conducted under a framework shaped by a June 2 executive order directing federal agencies to develop a process for benchmarking frontier AI models before broad deployment. OpenAI previewed the models and its release plans with the US government ahead of launch and, at the government's request, began with a limited preview group of roughly 20 trusted organizations. OpenAI stated it does not consider this kind of government-approval process an appropriate long-term default, describing the current arrangement as a short-term step while a repeatable framework is developed alongside the administration.

The safety architecture deployed with GPT-5.6 is OpenAI's most layered to date. It includes model-level refusal training, newly developed activation classifiers that monitor internal model states during inference and can pause or block generation in real time, a two-tier monitoring system covering all conversations, and actor-level enforcement that can escalate or restrict accounts showing patterns of high-risk activity. The Preparedness classification of all three models at the High level for cybersecurity and biological/chemical capability, including the smaller Terra and Luna tiers, carries governance implications for enterprises deploying the models in sensitive sectors.

GPT-5.6 Sol is also planned for deployment on Cerebras hardware in July, targeting throughput of up to 750 tokens per second for enterprise applications requiring real-time frontier-grade performance.