OpenAI launches GPT-5.6 Sol, Terra, and Luna on apps and API

GPT-5.6 is out now, and OpenAI isn't shipping one model this time, it's shipping a tiered lineup. Sol is the flagship, built for the hardest work. Terra sits in the middle as the everyday option. Luna is the cheap, fast one you'd reach for when you don't need much horsepower. All three are rolling out globally today across ChatGPT, ChatGPT Work, Codex, and the API, with full availability expected within 24 hours.

The headline feature is something OpenAI calls ultra mode, which runs four agents in parallel by default to tear through complex tasks faster, at the cost of burning more tokens. Developers get a similar trick through a multi-agent beta in the Responses API, plus something called Programmatic Tool Calling, where the model writes and executes JavaScript on the fly to juggle multiple tools, loop through steps, and check its own intermediate results before answering. It's a shift from 'answer the question' toward 'run the whole workflow.'

OpenAI is also leaning hard into GPT-5.6 as a document-and-deliverable machine. It can produce editable slides, spreadsheets, interfaces, and frontend prototypes, pulling context from files and connected work apps inside ChatGPT Work. That's consistent with what OpenAI has been signaling for months: ChatGPT isn't trying to be just a chat window anymore, it wants to be the place where the actual work output gets made.

Pricing splits cleanly by tier. Sol runs $5 per million input tokens and $30 per million output tokens. Terra comes in at $2.50 and $15. Luna is the bargain option at $1 and $6. OpenAI also tightened up prompt caching, with explicit cache breakpoints and a guaranteed 30-minute cache life, though writing to the cache now costs 1.25x the normal input rate while reads keep their steep 90% discount.

On benchmarks, OpenAI claims Sol tops the Artificial Analysis Coding Agent Index at 80.0, edging out Anthropic's Claude Fable 5 by 2.8 points, while using under half the tokens and time and costing roughly a third less. The company also points to strong results in cybersecurity testing, scientific workflow benchmarks, and browsing tasks, and says its own researchers are now using GPT-5.6 internally to debug systems and run experiments — which is either a vote of confidence or just good marketing, depending on how cynical you're feeling.