GPT-5.6: A Revolution in AI with Sol, Terra, and Luna

A New Era for GPT-5.6 with Sol, Terra, and Luna

The GPT-5.6 model family is now available to the general public after a limited preview phase. This new series includes the flagship model Sol, accompanied by Terra, designed for balanced everyday use, and Luna, the most economical model in the range.

GPT-5.6 Sol redefines the standards of artificial intelligence, standing out for its efficiency and ability to excel in various fields such as coding, cybersecurity, and sciences. It surpasses not only its predecessors but also its competitors, while requiring fewer tokens and reducing costs. This translates into better performance per dollar invested, allowing for more refined results within the same budget or reducing overall expenses while maintaining quality. The model also introduces the ultra parameter, which coordinates multiple agents to accelerate complex tasks, optimizing the use of computing resources and design judgment.

Unmatched Performance and Increased Efficiency

GPT-5.6 has been trained to maximize the utility of each token. In the Agents’ Last Exam, an assessment of professional workflows across 55 domains, GPT-5.6 Sol achieved an impressive score of 53.6, surpassing Claude Fable 5 by 13.1 points. Even in medium reasoning scenarios, it outperforms Fable 5 by 11.4 points, while being approximately four times less expensive. This efficiency is also reflected in the smaller models, GPT-5.6 Terra and GPT-5.6 Luna, which outperform Fable 5 at one-sixteenth of the cost. On the Artificial Analysis Intelligence Index, which evaluates intelligence across various domains, GPT-5.6 Sol approaches Fable 5 while completing tasks in 61% less time and at about half the cost.

Enhanced Security and Advanced Protections

The launch of GPT-5.6 comes with the most robust protections ever integrated, designed to withstand abuse while allowing legitimate use. Before its general availability, the model underwent rigorous evaluation, combining large-scale human and automated testing. During the preview phase, close collaborations with experts and partners helped to test and strengthen defenses, resulting in a system that integrates model-level protections with real-time checks and access adjusted according to trust and risk levels.

Default Efficiency, Maximum Performance on Demand

GPT-5.6 Sol stands out as the best coding model to date. On the Artificial Analysis Coding Agent Index, it sets a new record with a score of 80, surpassing Fable 5 by 2.8 points while using less than half the tokens and reducing time and cost by about a third. This advantage extends across the entire range, with Terra and Luna outperforming Fable 5 and Opus 4.8, respectively, each achieving these performances in one-third of the time, with half the tokens and at a quarter of the cost.

GPT-5.6 is capable of creating and executing lightweight programs that coordinate tools, process intermediate results, monitor progress, and choose subsequent actions. This helps reduce the number of tokens needed and optimize workflow. For tasks requiring a greater investment of time and computation, GPT-5.6 can exceed its default efficiency through max mode, which allows it more time to explore alternatives and revise its approach. The ultra mode goes even further by coordinating four agents in parallel, trading higher token usage for more robust and faster results.

A Significant Advancement in Design

GPT-5.6 marks a turning point in design judgment. With high-level guidelines, it is capable of creating sleek and functional interfaces, thanks to its enhanced computing capabilities that allow it to inspect and refine the final output, detecting visual and functional issues and applying necessary finishes.

GPT-5.6 excels in professional tasks, transforming complex contexts from daily documents and workflows, such as Slack, Notion, Microsoft 365, and Google Drive, into shareable expert-level artifacts. Its strength in knowledge work is evident in assessments covering long-term professional analysis, navigation, tool usage, and computer utilization. GPT-5.6 Sol achieves top results on BrowseComp with 92.2% and OSWorld 2.0 with 62.6%, surpassing Opus 4.8 while using 85% fewer output tokens.