The new GPT-5.6 family: Luna, Terra, Sol

OpenAI has released its new GPT-5.6 model family, which consists of three sizes: Luna, Terra, and Sol. These models feature a one-million-token context window, a 128,000-token maximum output limit, and a knowledge cutoff date of February 16, 2026. Pricing varies by model size, ranging from $1 per million input tokens for the smallest model, Luna, to $5 per million input tokens for the largest model, Sol.

The company claims significant improvements in long-running agentic performance, specifically citing the Agents’ Last Exam benchmark. In this evaluation of professional workflows, the Sol model outperformed the Claude Fable 5 model by 13.1 points. OpenAI highlights that these models offer higher efficiency and lower costs compared to competitors, with the smaller Terra and Luna models reportedly achieving better performance than Fable 5 at a fraction of the price. However, the models scored lower than Fable 5 on the SWE-Bench Pro coding benchmark, a result OpenAI attributes to flaws within that specific testing framework.

New API features accompany the release, including programmatic tool calling that allows models to execute JavaScript for orchestrating tasks and a multi-agent capability that enables the creation of sub-agents for parallel work. Developers can now utilize explicit prompt cache breakpoints and request original image resolution to avoid resizing. These updates aim to provide more granular control over model reasoning and efficiency, supporting complex workflows that require deeper integration with external tools and data.