OpenAI Releases GPT-6.1 Terra Model, API Price 50% Lower Than Previous Generation

Smart Things October 27 news: Today, OpenAI launched the new model GPT-6.1 Terra, focused on visual reasoning, precise computer operation, and more natural writing capabilities. Compared with the previous generation model GPT-5.6 Terra, the new model is further upgraded in daily communication and computer tasks, and can be used for work such as drafting proposals, interpreting technical drawings, filling out forms, and editing videos.

GPT-6.1 Terra continues to focus on the balance among speed, capability, and cost. Its standard API price is $1 per million input tokens (about 6.67 yuan), $0.05 per million cached input tokens (about 0.33 yuan), and $5 per million output tokens (about 33.35 yuan), 50% lower than both GPT-5.6 Terra and GPT-6.1 Sol, making it convenient for developers to use it for large-batch automation work such as reading invoices and entering information into accounting software.

In terms of model positioning, OpenAI's GPT-6 and GPT-6.1 series include different models such as Astra, Sol, Terra, and Luna. GPT-6.1 Terra focuses on speed, cost, and everyday task execution, and is suitable for large-scale invocation; GPT-6 Astra is OpenAI's most capable flagship model, aimed at highly difficult reasoning and complex tasks; GPT-6.1 Sol provides capabilities close to Astra at a relatively lower price, suitable for complex programming and reasoning tasks; GPT-6 Luna focuses on fast and efficient large-scale daily work.

In ChatGPT Work and Codex, using the same quota, GPT-6.1 Terra can complete more tasks than GPT-5.6 Terra, GPT-6.1 Sol, and GPT-6 Astra.

GPT-6.1 Terra is now available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex.

Currently, on the Artificial Analysis Intelligence Index leaderboard, GPT-6.1 Terra scores 48 points, 6 points higher than GPT-5.6 Terra, ahead of Grok 4.7 (xhigh) and Mimo-V2.6-Pro at 47 points, and close to GPT-6.1 Sol's score of 52 points.

1. Computer operation score rises 15 percentage points, performance on some tasks matches GPT-6 Astra

OpenAI used Agents' Last Exam to evaluate GPT-6.1 Terra's computer operation capabilities. This benchmark includes a variety of professional tasks, such as converting architectural drawings into 3D models and reviewing an animated film shot by shot to generate time-coded records for the restoration team.

At medium reasoning effort, GPT-6.1 Terra's score is 15 percentage points higher than GPT-5.6 Terra, at about one-third the cost of the latter. Its performance is comparable to GPT-6 Astra, while its price per token is only one-tenth of the latter's.

At the highest reasoning effort, GPT-6.1 Terra's token usage efficiency is more than 3 times that of Claude Opus 5.5, its price per token is one-quarter of the latter's, and the cost per single task is about one-ninth of the latter's. GPT-6.1 Terra also offers faster speed, completing tasks more than 3 times faster than GPT-6 Astra and GPT-6.1 Sol.

In the MMMU-Pro visual reasoning benchmark test, GPT-6.1 Terra performs better than GPT-5.6 Terra at all reasoning effort levels. The new model achieves results close to GPT-6 Astra and GPT-6.1 Sol at much lower cost.

OSWorld 2.1 is used to test AI Agents' ability to operate desktop applications and complete multi-step tasks, such as making 3D models based on technical drawings. In the benchmark's offline test set, GPT-6.1 Terra at medium reasoning effort surpasses GPT-5.6 Terra's highest score, with significantly lower cost per single task.

2. Factual error rate drops to 4.3% at high reasoning effort, programming task resolution rate doubles

OpenAI says it invested a great deal of work this year in reducing model hallucinations and factual errors. In an internal benchmark using prompts that easily induce errors, at low reasoning effort, GPT-6.1 Terra's error rate dropped from 53.6% for GPT-5.6 Terra to 16.7%, a decrease of nearly 37 percentage points.

At the highest reasoning effort, GPT-6.1 Terra's error rate dropped by 18 percentage points compared with the previous generation, to 4.3%, close to the level of GPT-6 Astra and GPT-6.1 Sol.

In programming, on Terminal-Bench 4.0, GPT-6.1 Terra's task resolution rate reached 50.3%, double GPT-5.6 Terra's score of 21.5%. This test examines programming Agents' ability to solve problems through the command line, such as fixing layout shifts that appear when a webpage loads.

However, OpenAI says that for the most complex programming tasks, GPT-6.1 Sol and GPT-6 Astra remain better choices.

Terminal-Bench-Science 0.1 examines Agents' ability to use code to complete scientific research tasks, such as tracking cell division in a time-lapse video. At the highest reasoning effort, GPT-6.1 Terra's score exceeds GPT-6.1 Sol's, at lower cost. The score gap between GPT-6.1 Terra and Claude Opus 5.5 is only 5 percentage points, while the estimated cost is less than one-fifth of the latter's.

3. Continuing the GPT-6 series alignment improvements, strengthening task boundaries and prompt injection protection

OpenAI says that while improving AI efficiency and accessibility, it is also continuously training, testing, and refining safeguards. Under its "Frontier Risk Prevention Framework," GPT-6.1 Terra reaches the Critical threshold for cybersecurity capability and retains the safeguards developed for GPT-6 Astra.

GPT-6.1 Terra also inherits the alignment improvements of the GPT-6 series. According to OpenAI's evaluation, compared with GPT-5.6 Terra, the new model is better able to act within the scope of tasks authorized by users and truthfully explain to users its own limitations and factors that may hinder task completion.

In prompt injection tests, GPT-6.1 Terra also shows strong resistance to malicious instructions that attempt to make it deviate from user tasks.

Conclusion: Advancing computer task automation at lower cost

While GPT-6.1 Terra focuses on improving visual understanding, computer operation, and factual accuracy, it reduces the time and cost required to complete tasks. Compared with GPT-6.1 Sol and GPT-6 Astra, this model has a higher cost-performance advantage, but there is still a certain gap in complex programming tasks.

For work that requires repeatedly reading information, operating software, and entering data, a lower cost per single task helps expand the application scope of Agent tasks such as computer operation.