GPT 5.6 Terra

The GPT-5.6 model that balances intelligence against cost, sitting between the flagship Sol and the cost-efficient Luna and corresponding to the mini tier of earlier generations. OpenAI positions it as a lower-cost option whose performance is competitive with GPT-5.5. It suits everyday coding, reasoning and agentic work where Sol is more model than the task needs.

Terra is the middle tier, and OpenAI describes it in one line: a lower-cost model whose performance is competitive with GPT-5.5, the previous frontier generation. That is the whole proposition. Anything last year's flagship could do, Terra is meant to do for a fraction of the money, which makes it the default for everyday coding, analysis and agent work where Sol is more model than the task needs. Three weeks after launch the lab cut its price by a further 20%, to two dollars per million input tokens and twelve per million output.

The lab's own tables put Terra ahead of GPT-5.5 on most of what it measured: 50.4% against 46.9% on Agents' Last Exam, 63.4% against 59.4% on SWE-Bench Pro, 87.4% against 85.6% on Terminal-Bench 2.1, and 50.2% against 47.5% on OSWorld 2.0. On the coding-agent index OpenAI cites, it reports Terra landing just above the strongest competing model, and says the family reaches those results in roughly a third of the time, with about half as many output tokens, and at approximately a quarter of the estimated cost.

Long context is where Terra separates itself from the cheaper tier rather than from the expensive one. On the lab's multi-round retrieval evaluation it reports 89.6% at 256K to 512K tokens and 72.5% at 512K to 1M - within a couple of points of Sol, and roughly double what Luna manages at the same depths. If your work involves feeding a whole repository, a long filing or a stack of documents into the window, Terra is the cheapest tier that actually holds it.

The gaps to Sol are real and worth knowing before you commit. On the hardest mathematics tier OpenAI reports 68.3% against Sol's 83%, and below GPT-5.5's 72.5%. On its genomics evaluation Terra scores 23.3% against Sol's 28.7%, and across cybersecurity the spread is wider still - 52.9% against 73.5% on ExploitBench. Terra is the tier for volume and for ordinary difficulty; the moment a task is genuinely at the frontier of maths, biology or security research, the lab's own numbers say to escalate.