OpenAI released GPT-6.1 Terra on October 26, a mid-tier model built for one job in particular: using a computer the way a person does. It reads what is on a screen, then clicks, types and moves between applications to get work done. The pitch is less about raw intelligence and more about price. Terra runs at a dollar per million input tokens and five dollars per million output tokens, roughly half what its predecessor cost.
That predecessor is GPT-6.1 Sol, which OpenAI pitched three weeks ago at legacy code migration. Terra is the cheaper sibling, aimed at the long, repetitive agent tasks that burn through tokens: filling forms, pulling figures out of documents, clicking through a web app to finish an order.
What the numbers say
OpenAI's own figures put Terra at about 57.6 percent on Agents' Last Exam at medium effort, for roughly 29 cents a task, rising to 60 percent at maximum effort for about a dollar a task. For comparison, the company's chart lists Anthropic's Claude Opus 5.5 at 63.2 percent, but at 8.82 dollars a task. OpenAI also claims Terra matches its flagship GPT-6 Astra on some agent work at a tenth of the per-token price, and runs more than three times faster.
Those are vendor numbers, and they deserve the usual caution. One independent tester, NerdsChalk, measured a smaller speed gain, closer to 1.8 times on a handful of short tasks. An OpenAI engineer quoted in early coverage was blunter still, describing Terra as a way to stretch usage limits rather than a real step up in capability. Cheaper is the headline. Smarter is not the claim.
Where to find it
Terra is live now in ChatGPT for Plus, Pro, Enterprise and Edu subscribers, inside Codex, and through the API as gpt-6.1-terra. Cached input drops to five cents per million tokens, batch and flex runs cost half the standard rate, and a fast mode costs double. Prompts above 272,000 tokens are billed higher, which matters given the million-token context window.
The release fits a pattern. The frontier labs have largely stopped competing on who has the single smartest model and started competing on who can make capable agents cheap enough to run all day. OpenAI spent last week putting its models on Google's cloud, and it has been candid about the oversight problems that come with agents left to work on their own. A model priced to run unattended, for hours, makes both of those stories more pressing.
Commentarii · 0