OpenAI Launches GPT-5.6 Family and ChatGPT Work — A Three-Tier Model Strategy for the Agent Era

On July 10, 2026, OpenAI released GPT-5.6, its new flagship model family, together with a new product called ChatGPT Work. Instead of one big model, GPT-5.6 is sold in three tiers with a clear naming scheme. Sol is the flagship for the hardest tasks, Terra is a balanced mid-tier, and Luna is a low-cost, low-latency option. Each tier can be updated on its own schedule, so OpenAI no longer has to ship a new generation for every price point.

The launch is also a story about token efficiency. OpenAI says Sol finishes a typical agent task using roughly half the output tokens of the previous generation, and the new family runs about 57 percent faster on the same workload. In a coding agent benchmark, Sol scores 88.8 percent in standard mode and 91.9 percent in the new Ultra sub-agent mode, beating the closest competitor on both cost and speed.

The more interesting move is ChatGPT Work, a long-horizon agent that lives inside the ChatGPT desktop client. OpenAI reports that Codex already has more than 5 million weekly users, and over 1 million of them use it for non-coding office work. The Work agent is designed to take a single instruction, pull data, build spreadsheets, draft slides, and return a finished package, replacing workflows that used to take several days.

The most striking detail is operational. OpenAI says the smallest tier, Luna, ran its own post-training almost entirely on its own. The flagship Sol model handled GPU scheduling, picked the training configuration, wrote the launch scripts, and verified the results. For enterprise buyers, the message is clear: the next round of competition is not just about which model is smartest, but about who can run an entire model factory with the least human input.