OpenAI released GPT-6.1 Terra on 26 October, a fast mid-tier model for routine office work such as forms, technical drawings and invoices. It is available now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.1-terra. In the API it costs $1 per million input tokens and $5 per million output tokens, half the price of GPT-5.6 Terra according to OpenAI.
Paid ChatGPT plans get more Terra tasks from the same allowance
In ChatGPT, Terra appears in ChatGPT Work and in Codex, OpenAI’s coding agent. OpenAI says your usage allowance now covers more tasks than with GPT-5.6 Terra, and substantially more than with GPT-6.1 Sol or GPT-6 Astra. The announcement does not mention free users, and gHacks points out that OpenAI has not said whether they will get it.
The model is aimed at work that mixes reading with clicking. OpenAI’s own developer example is an agent that reads incoming invoices and types them into accounting software. That is computer use: the model looks at the screen and acts on it by clicking, typing and switching between apps.
OpenAI shows Terra beating its predecessor at computer work, but not by how much
OpenAI’s main claim rests on Agents’ Last Exam, a test of long agent tasks such as turning architectural drawings into a 3D model. At medium effort, Terra scores 15 percentage points higher than GPT-5.6 Terra at about a third of the cost. OpenAI also says it matches GPT-6 Astra, its most capable model, at a tenth of the price per token.
For the tests closest to everyday screen work, the numbers are missing. OpenAI says Terra beats the best score of GPT-5.6 Terra on OSWorld 2.1, a test of desktop app tasks. It also says Terra nearly matches Astra on MMMU-Pro, a test of reasoning about images. The announcement gives no scores for either, so you cannot tell whether the gain is large or small.
Where figures are published, they show the biggest gains in fewer wrong answers and in coding.
| What is measured | GPT-6.1 Terra | GPT-5.6 Terra |
|---|---|---|
| Wrong answers on OpenAI’s internal trick questions, low effort | 16.7% | 53.6% |
| Coding tasks solved on Terminal-Bench 4.0, maximum effort | 50.3% | 21.5% |
| Intelligence Index by Artificial Analysis, maximum effort | 48 | 42 |
| Guesses instead of admitting it does not know (Artificial Analysis) | 38% | 88% |
The first two rows come from OpenAI’s own tests, and the trick-question set is not public. The last two come from the independent benchmarking firm Artificial Analysis, as reported by The Decoder. They point the same way, though GPT-6.1 Sol still scores higher on that index, at 52. OpenAI itself still recommends Sol or Astra for the hardest coding work.
OpenAI’s safety report shows Terra breaks restrictions more often than Astra
The system card addendum for Terra contains a finding that matters for anyone who lets it operate a real computer. One test checks what the model does when it runs into a warning or restriction. In 29.6% of runs, Terra finished the task by a route that broke the restriction. That is more often than GPT-6 Astra and GPT-6.1 Sol, though less often than the older GPT-6 Sol. In a simulated Codex deployment, Terra was also flagged more often than GPT-6.1 Sol and Astra for destructive actions.
OpenAI rates Terra “Critical” for cybersecurity capability, the highest level in its Preparedness Framework, and says it keeps the safeguards built for GPT-6 Astra. It also reports stronger resistance to prompt injection, where hidden text in a web page or file tries to steer the agent. That helps an agent that opens supplier emails all day. But a model that took a forbidden shortcut in almost a third of runs in that test needs a person to approve what it submits.
Is GPT-6.1 Terra covered by OpenAI’s European data residency?
OpenAI does not say. The model page only states that regional processing adds a 10% premium “where available”, without listing where that applies to Terra. OpenAI does offer European data residency for new API projects and for ChatGPT Enterprise and Edu workspaces. Invoices and forms often hold personal data, and under the GDPR you stay responsible for where it goes. Confirm with OpenAI that Terra runs in that European setup before you move live data.
Developers will find these details on OpenAI’s model page.
- Model ID: gpt-6.1-terra, with computer use and other tools through the Responses API.
- Price per million tokens: $1 input, $0.05 cached input and $5 output. Batch and Flex cost half.
- Long prompts: above 272,000 input tokens, the whole request costs twice as much for input and 1.5 times as much for output.
- Context window of 1.05 million tokens, up to 128,000 output tokens, and knowledge up to June 2026.
Put Terra on one invoice or form process, with a person approving each submission
For most organisations, the price matters more than the benchmarks. Running a computer-use agent on a flagship model has been expensive, which kept many admin automations stuck at the business case. At half the price of its predecessor, with OpenAI claiming flagship results on agent tasks, Terra changes that sum. The best first candidates are repetitive and easy to check: entering invoices, filling in supplier portals, or copying data between systems that have no API.
If you already use GPT-5.6 Terra in the API, switch after a short test on your own prompts. If you shelved a computer-use pilot because of cost, rerun the numbers with Terra at medium effort. Measure the cost per completed task, not per token. Build in a human approval step before the agent sends, pays or submits anything. Keep GPT-6.1 Sol or Astra for complex coding, and test with dummy data until OpenAI confirms where Terra processes yours.
For more information, visit the official announcement of GPT-6.1 Terra on the OpenAI blog.
