OpenAI launches new GPT-6.1 Terra for workplace tasks
OpenAI has launched GPT-6.1 Terra. It is now available in ChatGPT Work, Codex and the OpenAI API.
OpenAI says GPT-6.1 Terra is made for work tasks that combine writing, understanding images and other visual information, and using computers. It replaces GPT-5.6 Terra. The company says the new model costs less and performs better on several tests, both inside and outside OpenAI.
The standard API price is USD $1 for every million input tokens, USD $0.05 for every million cached input tokens and USD $5 for every million output tokens. OpenAI says these prices are 50% lower than the standard API prices for GPT-5.6 Terra and GPT-6.1 Sol.
The model is designed for tasks that businesses do often, including office administration and software work. These tasks include reading invoices and entering their details into accounting software, writing proposals, understanding technical drawings, filling in forms and editing videos. People with Plus, Pro, Enterprise and Edu plans can also use it in ChatGPT Work and Codex.
Test results
OpenAI shared results from several tests comparing GPT-6.1 Terra with older models and competing systems. On Agents' Last Exam, a test of professional tasks done on computers, GPT-6.1 Terra scored 15 percentage points higher than GPT-5.6 Terra when both used medium reasoning effort. OpenAI says the newer model cost about one-third as much to run.
OpenAI also says GPT-6.1 Terra performed about as well as GPT-6 Astra on this test while costing one-tenth as much per token. At the highest reasoning setting, the model used more than three times fewer tokens than Claude Opus 5.5 to do the work. Each token also cost about one-quarter as much. OpenAI says this made the cost of each task about nine times lower.
Speed was another area OpenAI compared. The company says GPT-6.1 Terra finished tasks more than three times faster than both GPT-6 Astra and GPT-6.1 Sol.
On MMMU-Pro, a test of how well AI understands visual information and reasons about it, GPT-6.1 Terra scored higher than GPT-5.6 Terra at every reasoning setting, according to OpenAI. The company says its scores were almost as high as those of GPT-6 Astra and GPT-6.1 Sol, but it cost less to run.
Tests on OSWorld 2.1 also showed improvements. This test measures how well AI agents use desktop apps to complete tasks with several steps. OpenAI says GPT-6.1 Terra beat the best score previously achieved by GPT-5.6 Terra, using medium reasoning effort and costing much less for each task.
Accuracy and coding
OpenAI says it has worked this year to reduce made-up answers and factual errors. In the company's own test, which uses difficult questions designed to make models get things wrong, GPT-6.1 Terra had a factual error rate of 16.7% at low reasoning effort. GPT-5.6 Terra's error rate was 53.6% on the same test.
At the highest reasoning setting, GPT-6.1 Terra's error rate dropped to 4.3%, according to OpenAI. The company says this was close to the error rates for GPT-6 Astra and GPT-6.1 Sol.
The new model also did better on software engineering tests. On Terminal-Bench 4.0, which tests how well coding agents solve tasks using command-line tools, GPT-6.1 Terra completed 50.3% of tasks at the highest reasoning setting. GPT-5.6 Terra's published score was 21.5%, OpenAI says.
OpenAI says GPT-6.1 Sol and GPT-6 Astra are still better choices for the hardest coding tasks. On Terminal-Bench-Science 0.1, a test of how agents use code for scientific research, GPT-6.1 Terra scored higher than GPT-6.1 Sol while costing less. It also scored within five percentage points of Claude Opus 5.5 at less than one-fifth of the estimated cost, according to OpenAI.
Safety measures
OpenAI says GPT-6.1 Terra meets the Critical level for cybersecurity under its Preparedness Framework. The model also keeps the safety protections first introduced for GPT-6 Astra.
OpenAI's tests suggest that GPT-6.1 Terra is more likely than GPT-5.6 Terra to stick to what the user asked it to do. It is also more likely to be honest about its limits and anything that stops it from completing a task. OpenAI says tests of prompt injection also showed that the model was good at resisting harmful instructions designed to distract it from the user's task.
"We're introducing GPT-6.1 Terra, a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing."