GPT-5.6 arrives with three models and a bigger push into agentic work

OpenAI has released GPT-5.6 across ChatGPT, Codex and its API. This time, the company is not presenting one flagship as the answer to every job. The family comes in three tiers built around different balances of capability, speed and cost.

Three models, different jobs

Sol is the top-end model for demanding research and professional work. Terra is positioned as the everyday option, while Luna is designed for tasks where speed and price matter most. For developers, that structure should make it easier to use a stronger model only when the work actually calls for one.

The release also pushes further beyond the familiar question-and-answer format. GPT-5.6 can browse, operate software, coordinate tools and revisit its own work while producing code, documents or interfaces. OpenAI says Sol reached 62.6% on OSWorld 2.0, a benchmark for computer use, and reported large gains in several cybersecurity tests.

Those numbers were published by OpenAI, so they should be treated as early evidence rather than a final verdict. The clearer picture will come from independent testing and from people using the models on messy, multilingual work that does not resemble a benchmark.

The harder part starts after launch

Stronger cyber and biology capabilities make safeguards more important. OpenAI says GPT-5.6 remains below its critical-risk thresholds and is protected by layered monitoring. A global rollout will now test how well those controls hold up without making legitimate research unnecessarily difficult.

Sources: OpenAI's GPT-5.6 announcement and OpenAI research index, July 9, 2026.