OpenAI Unveils GPT-5.6 Sol with Multi-Agent Reasoning

The new GPT-5.6 lineup introduces a tiered structure—Sol, Terra, and Luna—designed to decouple model generation from intelligence, speed, and pricing. Sol, the most capable of the three, allows for internal inspection of intermediate results and the coordination of external tools. By enabling the model to revise its own work before presenting a final output, OpenAI aims to increase task success rates while simultaneously reducing token consumption.

For high-demand workloads, the company introduced two compute-intensive modes: Max and Ultra. The latter employs a four-agent system that processes tasks in parallel, allowing separate agents to work on distinct components of a project before synthesizing a final result. Developers also gain access to Programmatic Tool Calling, a feature that permits the model to execute lightweight code in memory to filter data, bypassing the need to route every processing step through the primary language model.

Performance metrics for the release show Sol leading on the Artificial Analysis Coding Agent Index, while Terra and Luna are positioned as competitive alternatives to Anthropic’s Claude 5 and Claude 4.8, respectively. The models integrate directly with business platforms including Microsoft 365, Notion, and Slack to handle document and spreadsheet automation. API pricing reflects the new tiering, with Sol starting at $5 per million input tokens, while Luna serves as the entry-level option at $1 per million input tokens. Safety evaluations categorize all three models as High capability in cybersecurity and biology, though they remain below the company’s defined Critical threshold.