OpenAI｜GPT-5.6 series models officially launch—get a clear look in one article at the performance of three versions: Sol, Terra, and Luna

OpenAI latest announced the release of its newest GPT-5.6 series models. The three GPT-5.6 models—Sol, Terra, and Luna—are marked with different capability levels, helping users and developers choose based on intelligence, speed, and cost more clearly. All three GPT-5.6 models are available for users to download starting today.

OpenAI | New real-time voice model GPT-Live is here, supporting real-time translation and enabling a real-time conversational experience

AI frenzy | Meta plans to produce AI chips in September, doubling compute capability

The three versions Sol, Terra, and Luna roll out simultaneously. The flagship Sol fully refreshes records across intelligence, efficiency, reasoning depth, tool collaboration, and security protection; mid-tier Terra and entry-level Luna, meanwhile, beat competing models in their class with extremely low cost, further driving the mass adoption of AI.

OpenAI emphasized that GPT‑5.6 Sol is the strongest general-purpose model available today, and for the first time in multiple professional benchmarks, it achieves leading results with fewer Tokens and lower cost:

• Agents’ Last Exam: 53.6 points, 13.1 points higher than Claude Fable 5 (adaptive reasoning); even in the mid-tier reasoning mode, it still secures an 11.4-point advantage at about one-quarter the cost.

• Artificial Analysis Intelligence Index: only 1 point behind Fable 5, but it shortens task completion time by 61%, with costs about half of the competitor.

• Coding Agent Index: 80 points (max reasoning), 2.8 points higher than Fable 5, while halving output Tokens, halving time, and cutting cost by one-third.

Sol’s core advantage is that efficiency is performance: higher output per Token, deeper reasoning, and more precise tool collaboration, enabling professionals’ work such as programming, scientific research, analysis, browsing, and more to move faster across the board.

The mid-tier GPT-5.6 Terra targets everyday work workflows. Its performance can directly match the previous-generation GPT-5.5 models, but it costs only half as much. Lastly, the lightweight GPT-5.6 Luna focuses on extreme cost-effectiveness and processing speed, making it a suitable option for low-cost, high-volume data processing.

In performance design, GPT‑5.6 adds two key new capabilities: max reasoning strength (max reasoning) and an ultra multi-agent mode, addressing two major AI pain points—deep reasoning and breaking down large tasks.

First, max reasoning gives the model more compute time, enabling deeper logical thinking, exploring more solution paths, and repeatedly checking and correcting along the way. This extends reasoning time in exchange for higher-quality design, making it suitable for high-value scenarios such as compliance reviews, long-chain analysis, engineering workflows, and scientific research tasks, allowing the model to produce more stable and accurate results on complex problems.

On the other hand, ultra mode allows a single AI to simultaneously activate multiple “sub-agents” for parallel collaboration. Since real-world tasks are often large and multi-step, a single model can’t complete everything in one go. With sub agents, GPT‑5.6 can autonomously break down tasks, plan workflows, coordinate tools, integrate results, and iterate repeatedly until reaching the best outcome.

Starting today, GPT‑5.6 will roll out to globally eligible users. Plus, Pro, Business, and Enterprise users can use Sol; Free/Go users use Terra; Work and Codex users can choose Sol/Terra/Luna, and can enable max reasoning strength and ultra mode (depending on the plan).

《ET贏商有計》Episode by episode invites Hong Kong corporate brands and shares business operation tips.