GPT-5.6 Terra is the balanced, mid-tier model in OpenAI's GPT-5.6 family, positioned between the flagship Sol model and the fast, cost-efficient Luna model. It is designed for everyday coding, reasoning, and agentic workflows where teams need strong capability without frontier-level pricing. OpenAI describes Terra as delivering performance competitive with the prior-generation GPT-5.5 at roughly half the cost.
Architecture and Capabilities
Terra shares the same core context window as Sol and Luna, supporting roughly 1.05 million tokens of context with a maximum output of 128,000 tokens, and accepts both text and image inputs. It supports tool use, function calling, structured outputs, and prompt caching, with configurable reasoning effort levels ranging from none to a higher "max" setting. Like the rest of the GPT-5.6 family, Terra is available through the API under the model ID gpt-5.6-terra, separate from the gpt-5.6 alias, which routes to Sol.
Performance
On independent benchmark tracking, Terra has scored competitively for its tier, including strong results on GPQA Diamond, a graduate-level science reasoning benchmark, and on coding-focused indexes. OpenAI reports that Terra outperforms the prior-generation GPT-5.5 on several evaluations while using fewer tokens, and that it and Luna outperform some competing frontier models at substantially lower estimated cost. Terra is generally positioned as a practical default for production workloads such as software engineering assistance, long-document analysis, and multi-step agentic pipelines.
Safety
Terra was released alongside Sol as part of OpenAI's coordinated GPT-5.6 rollout, which included a limited preview period tied to engagement with the U.S. government over the family's cybersecurity and biological-risk capabilities. Under OpenAI's Preparedness Framework, GPT-5.6 models, including Terra, are treated as High capability in cybersecurity and biological/chemical risk domains, without reaching the framework's Critical threshold.
Exact parameter counts and training architecture details have not been publicly disclosed.