Term
GPT Terra
GPT Terra is the middle tier of OpenAIs GPT-5.6 family (July 2026), sitting between the efficient Luna and the flagship Sol — built for a strong price-performance balance with a 1M-token context.
GPT Terra — explained in more detail
GPT Terra is not a standalone model but the middle tier of the GPT-5.6 family that OpenAI released on 9 July 2026. This generation is, for the first time, split into three variants ranked by capability: Luna (the most efficient), Terra (the balanced middle) and Sol (the flagship). Terra is therefore the model that offers a strong balance of capability and cost — comparable to the role a single default model used to play in the line-up.
Technically, Terra shares the familys core traits: a context window of roughly 1 million tokens and up to 128,000 output tokens per response. Its knowledge base runs to 16 February 2026. In benchmarks Terra reached, among others, 88.4 percent on GPQA Diamond (scientific reasoning) and 72.0 percent on TAU-Bench (agentic tasks). On price it sits in the middle of the family at roughly 2 to 2.50 US dollars per million input and 12 to 15 US dollars per million output tokens; the figures vary by provider and routing.
Access is proprietary: Terra runs through the OpenAI API and connected platforms, and the weights are not public. OpenAI positions the family for enterprise work, coding, scientific research and cybersecurity.
Example / Practical context
In practice, Terra is the obvious pick within the GPT-5.6 family for most day-to-day tasks: coding, writing, data analysis and agentic workflows where a good balance of quality, speed and cost matters. Those who need maximum performance for especially hard reasoning or research reach for Sol; those who want to process very large volumes of simple requests as cheaply as possible take Luna. Terra is the default choice in between — for example as the engine behind a coding agent making many tool calls over a long context.
Distinction from related terms
Terra is one of three tiers of a model family, not a model of fixed size — the tier only says where in the capability and price spectrum the model sits. Within GPT-5.6, Terra lies between Luna (cheaper, weaker) and Sol (stronger, pricier). Open-weight models (such as Llama, Qwen or DeepSeek) differ fundamentally from the API-only GPT models: their weights can be downloaded and run locally, whereas GPT Terra is available solely as a hosted service. Against Anthropics Claude models of the same period, Terra appears mainly as the more cost-efficient alternative, while Claude leads on individual coding benchmarks such as SWE-Bench Pro.