Agentic tasks that use a computer demand strong on-screen awareness, multimodal understanding, and speed, ideally without running up your token costs. OpenAI thinks it has cracked the code.
On Monday, OpenAI launched GPT-6.1 Terra, a model the company says is designed to deliver "fast and capable computer use at scale." In practical terms, this is OpenAI's mid-tier model that competes with Anthropic's Sonnet. Terra can deliver more precise computer use at 50% lower cost than GPT-5.6 and GPT-6.1 Sol. The company is positioning the model's selling point as getting more bang for your buck and, as a result, enabling developers to automate high-volume work.
"In ChatGPT Work and Codex, your plan goes further with GPT-6.1 Terra—the same usage allowance covers more tasks than with GPT‑5.6 Terra and substantially more than with GPT‑6.1 Sol or GPT‑6 Astra," OpenAI stated in the release.
To back up its advanced agentic performance claims, OpenAI put GPT-6.1 Terra’s computer-use capabilities to the test on Agents’ Last Exam, a benchmark designed by Berkeley RDI and over 300 industry experts to test agents' performance specifically on long-horizon, economically valuable tasks with verifiable outcomes. The company reported the results were:
- GPT-6.1 Terra scored 15 percentage points higher than GPT-5.6 Terra at about one-third the cost while on medium reasoning effort.
- Performed on par with GPT-6 Astra at only one-tenth its price per token.
- Terra can complete tasks over three times faster than GPT-6 Astra and GPT-6.1 Sol.
- Compared with Claude Opus 5.5 (a step up from Sonnet), the company reports that, at max effort, Terra is over three times as token-efficient and costs one-fourth as much per token. Of course, note that that's a previous-generation model.
Beyond benchmark performance on MMMU-Pro, OSWorld 2.1, and Terminal Bench 4, the company is also touting advances in factuality and alignment. In its internal benchmark of error-inducing prompts, OpenAI reports that the error rate at low reasoning effort fell from 53.6% with GPT-5.6 Terra to 16.7% with GPT-6.1 Terra, and at maximum effort, the error rate fell by 18 percentage points to 4.3%. In internal evaluations, OpenAI also found that, compared to GPT-5.6 Terra, GPT-6.1 Terra is more likely to "stay within scope and honestly disclose limitations and potential blockers to the user." GPT-6.1 Terra also showed strong resistance to prompt injection tests.
GPT-6.1 Terra is available to all paid users in ChatGPT Work and Codex, while developers can access it through the OpenAI API.
Our Deeper View #
As people become more familiar with AI agents, more people are willing to hand off their tasks and workflows to agents. While general demand is just beginning with tools like OpenAI Dots, Meta Muse, and Claude Duet, it's growing even faster in the enterprise. Of course, the enterprise has the most to gain by maximizing AI use, systemizing it, and then applying it to high-value workloads. To meet that need, models need to be optimized to balance lower cost, higher capabilities, and speed, which is exactly what OpenAI targeted with this release. The mid-tier Terra models are naturally well-suited to slide into that sweet spot. This likely impacted OpenAI's decision to give its mid-tier models more attention after skipping them last month when it focused on high-end Sol and the budget model Luna instead.