GPT-6.1 Terra Released: Astra-Level Computer Operation at 1/10 Cost

OpenAI has released GPT-6.1 Terra, a new model positioned to deliver fast, powerful computer operation capabilities at a lower price. It combines fine-grained visual reasoning, precise computer control, and clearer, more natural writing.

Pricing and Availability

Standard API pricing for GPT-6.1 Terra:

Input tokens: $1 per million

Cached input tokens: $0.05 per million

Output tokens: $5 per million

This standard API price is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol , enabling developers to automate high-volume tasks such as reading invoices and entering data into accounting software at reduced cost.

Starting today, all Plus, Pro, Enterprise, and Edu users can access GPT-6.1 Terra in ChatGPT Work and Codex. Developers can call the model via the OpenAI API using model ID gpt-6.1-terra. Within ChatGPT Work and Codex, the same usage allowance completes more tasks than with GPT-5.6 Terra, and far more than with GPT-6.1 Sol or GPT-6 Astra.

From Visual Understanding to Practical Operation

OpenAI evaluated GPT-6.1 Terra's computer operation skills using the Agents' Last Exam benchmark, which covers professional tasks like creating 3D models from architectural drawings and reviewing animation films shot-by-shot to produce timestamped logs for repair teams.

At medium reasoning intensity , GPT-6.1 Terra scores 15 percentage points higher than GPT-5.6 Terra, at roughly one-third the cost .

Compared to GPT-6 Astra , performance is on par , but per-token price is only one-tenth .

At highest reasoning intensity , Terra's token-use efficiency exceeds Claude Opus 5.5 by over 3x , while per-token price is one-quarter of Opus 5.5, yielding a ~1/9 task cost .

Task completion speed is over 3x faster than GPT-6 Astra and GPT-6.1 Sol.

On the MMMU-Pro visual reasoning benchmark, GPT-6.1 Terra outperforms GPT-5.6 Terra at all reasoning levels, with scores approaching the flagship GPT-6 Astra and GPT-6.1 Sol at much lower cost.

OSWorld 2.1 tests an AI agent's ability to operate desktop applications for multi-step tasks (e.g., building a 3D model from technical drawings). At medium reasoning, GPT-6.1 Terra surpasses GPT-5.6 Terra's best score while costing far less per task.

Factual Accuracy and Coding Capabilities Significantly Improved

OpenAI invested heavily this year to reduce hallucinations and factual errors. On an internal adversarial prompt benchmark:

Low reasoning : error rate dropped from 53.6% (GPT-5.6 Terra) to 16.7% , a 37-percentage-point reduction .

Highest reasoning : error rate fell 18 percentage points to 4.3% , nearly matching GPT-6 Astra and GPT-6.1 Sol.

Software engineering ability also saw large gains. Terminal-Bench 4.0 tests command-line problem solving (e.g., fixing a page-layout shift on load). At highest reasoning, GPT-6.1 Terra solves 50.3% of tasks , more than double GPT-5.6 Terra's 21.5% . For the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choices.

Terminal-Bench-Science 0.1 evaluates scientific research tasks via code (e.g., tracking cell division in time-lapse video). At highest reasoning, GPT-6.1 Terra outperforms GPT-6.1 Sol at lower cost. Versus Opus 5.5, the score gap is under 5 percentage points while estimated cost is less than one-fifth .

Safety and Alignment

Under OpenAI's Preparedness Framework , GPT-6.1 Terra reaches the Critical capability threshold in cybersecurity while retaining the safety mitigations developed for GPT-6 Astra.

Alignment improvements over GPT-5.6 Terra include:

More likely to constrain actions to the user-specified scope.

More honest about its own limitations and potential obstacles.

Strong resistance to prompt-injection attacks that attempt to divert the model from the user's intended task.

References

Model card: https://deploymentsafety.openai.com/gpt-6-1-terra/gpt-6-1-terra.pdf

Blog post: https://openai.com/index/introducing-gpt-6-1-terra