OpenAI introduces a new family of GPT-5.6 models: Sol, Terra, and Luna — an architectural revolution or a marketing move?
OpenAI has taken a significant step in the development of its language models by announcing the GPT-5.6 family, which includes three specialized models: Sol, Terra, and Luna. Currently, access is only available to a limited circle of trusted partners via API and Codex, and this preview mode has been coordinated with the U.S. government.
Lineup Architecture and Pricing Policy
The new lineup clearly shows segmentation by performance and cost. The flagship model Sol is positioned as the most powerful in the company's history. Terra represents a balanced solution for everyday tasks, while Luna is a budget-friendly option with high speed. Notably, according to the developers, Terra delivers performance on par with GPT-5.5 but costs half as much. Pricing is as follows: per 1 million input tokens for Sol — $5, for Terra — $2.50, for Luna — $1. Generation costs (output tokens) are higher: $30, $15, and $6, respectively.
Technical Innovations and Benchmarks
The most intriguing aspect is the introduction of new reasoning modes. The Sol model features a max mode, which allocates additional time for deep task analysis, and an ultra mode, which uses sub-agents to accelerate complex computations. This points to an evolution toward more flexible architectures capable of adapting computational resources to query complexity.
Test results are impressive. Sol set a new record in Terminal-Bench 2.1 for command-line tasks. On GeneBench v1, the model outperformed GPT-5.5 while consuming fewer tokens. Particularly noteworthy is ExploitBench, where Sol proved competitive with Mythos Preview while using only a third of the output tokens. This indicates a significant improvement in efficiency.
Safety: A New Standard or Precautionary Measures?
OpenAI claims that the GPT-5.6 family features the most robust safety stack. The company conducted extensive red-teaming, utilizing over 700,000 GPU-hours (in A100 equivalents) to identify vulnerabilities. Results show that Sol does not cross the Cyber Critical threshold under the Preparedness Framework. While the model can find bugs and exploitation primitives in tests with Chromium and Firefox, it was unable to autonomously create a fully functional exploit. During the preview phase, multi-layered restrictions have been implemented, including real-time checks and account-level monitoring.
Plans and Strategic Context
In July, OpenAI plans to launch GPT-5.6 Sol on the Cerebras platform with a potential speed of up to 750 tokens per second, though access will again be limited. It is worth noting that the company recently filed a confidential IPO application, which may explain the heightened focus on safety and regulation.
My Expert Analysis: The division into three models is not just marketing but a deliberate strategy to create an ecosystem. Sol is a tool for research and complex tasks, Terra is a corporate standard, and Luna is a mass-market product. However, the restricted access and coordination with U.S. authorities hint that we are on the cusp of an era where AI capabilities will be regulated as strictly as nuclear technology. Investors should closely monitor how these measures impact adoption and monetization rates.