OpenAI Launches GPT-6.1 Terra AI Model for Professional Work

OpenAI has released its new midsize GPT-6.1 Terra AI model. The model works with ChatGPT Work, Codex and the API. It is designed for computer tasks, analyzing images and other visual data, programming and professional work. According to the company, GPT-6.1 Terra is more accurate than the previous GPT-5.6 Terra and costs 50% less to use via the API. The model is intended to automate long-running tasks at an affordable price.

OpenAI divides its core models into Luna, Terra, Sol and Astra classes based on price and capabilities. The new Terra can fill out forms in a browser, work with spreadsheets and explain technical drawings. It can also edit videos, draft proposals and produce more natural writing. The company claims that the model completes computer tasks more than 3 times faster than GPT-6 Astra and GPT-6.1 Sol.

In OpenAI’s internal test of prompts that cause errors, the error rate in low reasoning mode fell from 53.6% to 16.7%. This is an improvement of approximately 37 percentage points over the previous GPT-5.6 Terra. In the highest reasoning mode, errors fell by 18 percentage points to 4.3%. The result is close to those of GPT-6 Astra and GPT-6.1 Sol.

In the Agents’ Last Exam benchmark, which measures professional computer work, the new model outperformed the previous Terra by 15 percentage points in medium reasoning mode. The cost per task also fell by approximately 3 times. GPT-6.1 Terra achieved a result close to GPT-6 Astra on the same benchmark. In addition, its price per token is 10 times lower than Astra’s. The model also came close to Astra and GPT-6.1 Sol in the MMMU-Pro benchmark, which measures the analysis of images and other visual data.

Significant gains were also recorded in programming benchmarks. GPT-6.1 Terra scored 50.3% on the Terminal-Bench 4.0 test in the highest reasoning mode. The previous GPT-5.6 Terra scored 21.5%. Nevertheless, OpenAI recommends GPT-6.1 Sol or GPT-6 Astra for the most difficult programming, research and professional tasks.

Under the API pricing, 1 million input tokens cost $1 (approximately 2 manats). The price for 1 million output tokens generated by the model is $5 (approximately 9 manats). A token is a small unit of text processed by AI. One million previously processed and reused input tokens cost $0.05 (approximately 0.09 manats). By comparison, the same volume of input tokens costs $10 (approximately 17 manats) with GPT-6 Astra. The price for output tokens reaches $50 (approximately 85 manats). GPT-6.1 Terra is available to users of eligible paid plans, and developers can also access the model through the API.