GPT-6.1 Terra, computer use at office prices

On 26 October OpenAI released GPT-6.1 Terra, a model that reads what is on a screen and works the software there by clicking and typing, as a person would. Computer use was already the headline feature of GPT-6 Astra in September. What is new is the price: Terra costs 1 dollar per million input tokens and 5 dollars per million output tokens, a tenth of Astra’s price per token. At that level the question is no longer whether a company can afford an agent at the screen, but where it should let one work.

What it is

Terra replaces GPT-5.6 Terra in the middle of OpenAI’s range, below Sol and Astra and above Luna. OpenAI says it excels at “the communication and computer-based tasks that fill the workday”, from drafting proposals and interpreting technical drawings to filling out forms and editing videos. It is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work, the agent mode that carries out work in other apps and browsers, and in Codex. Developers call it through the API as gpt-6.1-terra.

OpenAI says the standard API price is 50 percent lower than for GPT-5.6 Terra and GPT-6.1 Sol, with cached input at 0.05 dollars per million tokens. Its own example of what that makes practical is high volume work such as reading incoming invoices and entering their details into accounting software. In ChatGPT Work and Codex, it says, the same plan allowance now covers more tasks than with the older models.

What the numbers show

OpenAI measured computer use on Agents’ Last Exam, a set of professional tasks such as turning architectural drawings into a 3D model. At medium effort Terra scores 57.6 percent, 15 points above GPT-5.6 Terra, at 0.29 dollars per task against 0.84. That equals Astra at the same effort. At maximum effort Terra reaches 60 percent for 1.01 dollars per task, while Anthropic’s Claude Opus 5.5 scores 63.2 percent for 8.82 dollars in the same chart. OpenAI also says Terra finishes tasks more than three times faster than Astra and GPT-6.1 Sol.

The desktop test is less flattering. On OSWorld 2.1, where agents complete multistep tasks in desktop apps, Terra reaches 74.2 percent at maximum effort, up from 57.1 for its predecessor. OpenAI’s own chart puts Claude Sonnet 5.5 at 83.9 percent, and GPT-6.1 Sol at 78 percent for 1.15 dollars per task, slightly less than Terra’s 1.19. A lower price per token does not always mean a lower price per finished task.

All of these are OpenAI’s figures, not yet checked independently. OpenAI itself says Sol and Astra remain the better choice for the most complex coding work.

Where it fits, and where it does not

Computer use earns its place where there is no other way into a system: an old desktop program with no interface for other software, a supplier or government portal, a form that has to be filled in by hand, a back office step where someone copies from one window into another. Pareekh Jain of EIIRTrend told InfoWorld that Terra is “ideal for high-volume tasks such as invoice processing, document analysis, application automation and routine coding.”

Where a proper API or integration exists, it remains the better route, in our view. A call to an accounting system either succeeds or returns an error, while a click lands wherever the screen happens to be after the last update. An integration is faster, does not need a stream of screenshots, and can be limited to exactly the actions it needs. Sanchit Vir Gogia of Greyhound Research made the economic point: “Cheaper inference improves one component of agent economics.” Review and integration still cost money, and in his words “the relevant purchase is completed work that remains correct and authorized.”

What can go wrong

OpenAI treats Terra as Critical in cybersecurity under its Preparedness Framework and gives it the same safeguards as Astra. It says Terra showed “strong resistance to malicious instructions designed to divert it from the user’s task” in its prompt injection tests, and that it stays within scope more often than its predecessor. In a simulation of internal Codex use, it drew about 42 percent fewer serious flags than GPT-6 Sol.

The system card also has less comfortable figures. When Terra met a barrier such as an access denied error, it finished the task by another route that “violated the spirit of the restriction” in 29.6 percent of runs at maximum effort, more often than Astra, Luna or GPT-6.1 Sol, though less often than GPT-6 Sol. OpenAI notes that this test ran without system level controls, “many of which would plausibly stop the circumventions”. Faced with messages from apparent other agents, Terra carried out the unauthorized action in 11 of 80 samples, against 1 for GPT-6.1 Sol.

For a company the risks are concrete. Text on a page or in a document can try to give the agent orders, a wrong click can change a live record, and an agent that logs in needs credentials that someone has to look after. Approvals and a complete log matter more than any benchmark.

How to start

Pick one narrow task that can be undone, such as entering supplier invoices into a test company or downloading statements from a portal into a folder. Give the agent its own account with the least access the task needs, never a person’s login, and keep passwords out of what it can see. Run it supervised, with a person approving every step that sends, pays, deletes or submits, and keep a record of every action.

Then measure it against the manual way, along the lines Jain suggests: tasks completed, accuracy, time, cost per task and how often a person had to step in. Only when those numbers hold up does it earn a second task.

Where we stand

We would use it in that order: integrations first, the screen where there is no other way in. Twin, the AI brain we build for companies, works in Claude, ChatGPT and Codex and connects to a company’s systems through integrations and MCP servers we build, with approvals and rights by role and every step logged. Helm, still in the making, is our vision of a company run by agents, with the founder approving at the key points. A cheaper model at the screen makes that more affordable. It does not change who answers for the click.