OpenAI has taken a significant step in the development of its AI family by opening limited access to three new models: GPT-5.6 Sol, Terra, and Luna. At this stage, access is granted to select partners via API and Codex. The company is coordinating with US authorities, who requested a preliminary preview before a full-scale launch. This is an unconventional but telling move, demonstrating the growing attention regulators are paying to advanced AI systems.

Lineup and Positioning

Within the family, roles are clearly defined. Sol is positioned as the flagship—the most powerful OpenAI model to date. Terra is the "workhorse" for everyday tasks, which the company claims is comparable to GPT-5.5 in performance but half the cost. Luna is a fast and most affordable option. This segmentation allows covering a wide range of scenarios—from complex analytical tasks to budget-friendly ones.

Technical Breakthroughs and Benchmarks

GPT-5.6 Sol demonstrates impressive results. For the first time in the series, two new reasoning modes have appeared: max, giving the model more time for deep processing of complex tasks, and ultra, using sub-agents to accelerate multi-step work. On benchmarks, Sol set a new record in Terminal-Bench 2.1 for command line, surpassed GPT-5.5 on GeneBench v1 with fewer tokens, and showed competitiveness with Mythos Preview on ExploitBench, using only a third of the output tokens. In ExploitGym tests, all three models improved results as reasoning depth increased—a key signal for the cybersecurity market.

Security as a Priority

OpenAI claims that the family has received the strongest security stack in the lineup. According to the company's assessment, Sol does not cross the Cyber Critical threshold within the Preparedness Framework. In tests with Chromium and Firefox, the model found bugs and exploitation primitives but could not autonomously create a fully functional exploit. Over 700,000 GPU-hours in A100 equivalent were used for automated red teaming to verify protections. During the preview phase, multi-layered measures are in place: model-level restrictions, real-time checks, account-level signals, and monitoring. Some requests may be blocked or take longer to process due to additional checks.

Pricing and Plans

Cost per 1 million tokens: Sol — $5 input and $30 output, Terra — $2.50 and $15, Luna — $1 and $6. In July, OpenAI plans to launch GPT-5.6 Sol on the Cerebras platform with speeds of up to 750 tokens per second, but access at this stage will also be limited. The company expects a full public launch in the coming weeks, but emphasizes that the current order of interaction with authorities will not become a long-term norm.

My comment: The emergence of three clearly segmented models and the introduction of max/ultra reasoning modes is not just an evolution, but an attempt by OpenAI to redefine performance and security standards. Particularly noteworthy is that the company voluntarily shares partner data with the government. This could set a precedent for the entire industry, where the balance between innovation and control is becoming increasingly delicate.