OpenAI launches GPT-6.1 Terra, a faster and cheaper AI model for computer work

OpenAI has launched GPT-6.1 Terra, a new artificial intelligence model designed to do everyday computer tasks faster and at a lower cost than earlier versions. It is now available to ChatGPT Plus, Pro, Enterprise, and Edu users. People can also use it through OpenAI’s Codex and API platforms.

What is new in GPT-6.1 Terra?

GPT-6.1 Terra is designed for what OpenAI calls “computer use.” This means the model can look at a computer screen, understand what it sees, and use the computer by clicking, typing, and moving between apps, much like a person would. It can automate tasks such as filling out forms, recording invoices, writing proposals, reading technical drawings, and editing videos. OpenAI says the model is better at understanding images, writing in a clear and natural way, and using information from pictures and documents.

OpenAI says its own tests showed that GPT-6.1 Terra could turn architectural drawings into 3D models and create logs with time codes for animated films.

How does GPT-6.1 Terra compare to previous GPT models?

GPT-6.1 Terra replaces GPT-5.6 Terra. It sits in the middle of OpenAI’s range of models, below Sol and Astra but above Luna. Compared with GPT-5.6 Terra, the new model is much faster and more accurate. OpenAI says GPT-6.1 Terra finishes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol.

According to The Next Web and BleepingComputer, GPT-6.1 Terra scored 15 percentage points higher than GPT-5.6 Terra on a test called Agents’ Last Exam. This test measures how well AI models handle professional computer tasks. The improvement was measured at medium reasoning effort, and the new model cost about one-third as much. Under these conditions, it also performed as well as GPT-6 Astra while costing one-tenth as much per token.

GPT-6.1 Terra also makes fewer factual mistakes. In a set of questions designed to challenge AI models, its error rate at the lowest reasoning setting fell from 53.6% with GPT-5.6 Terra to 16.7%. At the highest reasoning setting, its error rate dropped by 18 percentage points to 4.3%. This brings it close to the error rates of more powerful models such as GPT-6 Astra and GPT-6.1 Sol.

For coding tasks, GPT-6.1 Terra scored 50.3% on the Terminal-Bench 4.0 test at the highest reasoning setting. This was more than twice the 21.5% score of the previous model. However, OpenAI still recommends Sol or Astra for the hardest programming tasks.

What are its uses and limitations?

GPT-6.1 Terra is made for everyday office work and tasks done on a computer. OpenAI gives examples such as reading invoices and entering their details into accounting software, filling out forms, writing business documents, understanding technical drawings, and editing videos. Because it can understand and work with what appears on a screen, it can help automate repeated computer tasks and business processes.

Despite these improvements, OpenAI says GPT-6.1 Terra is not designed for the most difficult programming tasks. For those jobs, it recommends more powerful models such as Sol and Astra. The performance figures come from OpenAI’s own tests, and results may differ depending on the tasks people need to do and the computer environment they use.

How does pricing work?

OpenAI has cut the standard API prices for GPT-6.1 Terra by 50% compared with GPT-5.6 Terra and GPT-6.1 Sol. Developers pay $1 for every million input tokens and $5 for every million output tokens. Cached input tokens cost $0.05 per million. These prices let users complete more tasks within the same usage allowance than with earlier models. The model is available through the API under the name “gpt-6.1-terra.” ChatGPT Plus, Pro, Enterprise, and Edu users can also access it through ChatGPT and Codex.

How does it perform on tests?

OpenAI’s own tests show that GPT-6.1 Terra performs much better than the model it replaces. On Agents’ Last Exam, it scored 15 percentage points higher than GPT-5.6 Terra at medium reasoning effort and matched GPT-6 Astra’s performance at a much lower cost. Its error rate at low reasoning effort fell from 53.6% to 16.7%. At the highest reasoning setting, the error rate dropped to 4.3%. On command-line coding tasks, GPT-6.1 Terra achieved a success rate of 50.3%, more than twice that of GPT-5.6 Terra.

OpenAI also compared GPT-6.1 Terra with Anthropic’s Claude Opus 5.5. The company said Terra completed tasks at about one-ninth of the cost, but these results have not been independently checked.

What about safety and reliability?

OpenAI says GPT-6.1 Terra reaches the “Critical” level for cybersecurity skills under its Preparedness Framework, which it uses to rate abilities that could pose risks. The model has the same safety protections as GPT-6 Astra. Compared with its predecessor, it is more likely to stay focused on the task it was given and tell users about its limitations. OpenAI also says it is better at resisting prompt injections, which are attempts to make the model ignore the user’s intended task and follow different instructions instead.

Who can use GPT-6.1 Terra?

GPT-6.1 Terra is now available to ChatGPT Plus, Pro, Enterprise, and Edu users. It can also be used through Codex and the API. The announcement does not mention access for users on the free plan. The model is also available on Amazon Bedrock, Amazon’s cloud-based AI service.