GPT-5.6 Sol: Max Effort and Ultra Subagents in Python

On June 26, 2026, OpenAI began a limited preview of the GPT-5.6 series and quietly changed how it names models. Instead of one monolithic release, you now pick a tier: Sol (the flagship), Terra (balanced everyday work), and Luna (fast and cheap). The number is the generation; the name is the capability tier, and tiers can move on their own schedule. Within two days the developer crowd had latched onto two new knobs in particular: a max reasoning effort that lets Sol think longer than anything before it, and an ultra mode that spins up subagents to attack a hard problem in parallel.

This guide is about using those features from Python, not admiring the benchmarks. By the end you will have made your first GPT-5.6 Sol call through the Responses API, dialed reasoning effort up and down on purpose, run a full tool-calling agent loop, orchestrated subagents the way ultra mode does, and built a small router that sends each task to the cheapest tier that can handle it. Everything is runnable, and where a feature is still rolling out behind the preview I say so plainly instead of inventing an API.