OpenAI launches GPT-6.1 Terra for high-speed computer tasks

A new balance of cost and speed

OpenAI has introduced GPT-6.1 Terra, a new model designed to handle professional work and complex computer interactions. This release comes nearly a month after the company debuted the GPT-6.1 Sol model. The new model is specifically built to manage tasks like filling out browser forms, working within spreadsheets, and interpreting technical drawings.

One of the primary goals of Terra is to provide high-quality results without the high price tag of the company's flagship models. As reported by Yahoo Tech, Terra achieves results similar to the Astra model but at one-tenth of the price per token. It is also designed to be significantly faster, completing tasks more than three times faster than both GPT-6 Astra and GPT-6.1 Sol.

The model is now available to paid subscribers who use ChatGPT Work and Codex. Software engineers and developers can also begin integrating the model into their own applications through the OpenAI API. This rollout marks a shift toward making specialized AI capabilities more accessible for everyday professional use cases.

Performance on professional tasks

To measure how well the model handles software in a work environment, OpenAI used the Agents' Last Exam. This benchmark tests how well AI agents can complete professional computer tasks. On this test, GPT-6.1 Terra scored 15 percentage points higher than the older GPT-5.6 Terra model while costing about one-third as much per task.

The results show that Terra is highly effective at navigating software and completing multi-step instructions. It even matches the performance of GPT-6 Astra on this specific benchmark. This suggests that for many standard business workflows, users may not need to pay for the most expensive model available.

Beyond general computer use, the model also shows strength in visual reasoning. On the MMMU-Pro benchmark, Terra performed on par with both Astra and GPT-6.1 Sol. These capabilities allow it to interpret visual data and complex layouts that were previously difficult for smaller models to process accurately.

Significant gains in coding

The new model also shows a substantial improvement in its ability to write and debug code. On the Terminal-Bench 4.0, which measures coding performance at maximum reasoning effort, Terra scored 50.3%. This is a major leap from the 21.5% score achieved by the previous GPT-5.6 Terra model.

While these gains are significant, OpenAI clarifies that Terra is not intended to replace the highest-tier models for every situation. The company still recommends using GPT-6.1 Sol or Astra for the most demanding and complex coding projects. This distinction helps users choose the right tool based on the difficulty of the specific programming task they are facing.

By offering a high-performing coding option at a lower price point, OpenAI is targeting a broader range of developers. It allows for more experimentation and integration into smaller projects where the extreme reasoning of Astra might be unnecessary. This tiered approach helps balance the need for power with the reality of operational costs.

API pricing and accessibility

The pricing structure for GPT-6.1 Terra is designed to be much more economical than the Astra model. For every million input tokens, Terra costs $1, while the output tokens are priced at $5. For developers who use cached input, the cost drops even further to $0.05 per million tokens.

In comparison, the GPT-6 Astra model is significantly more expensive at standard API rates. Astra requires $10 for every million input tokens and $50 for every million output tokens. This makes Terra exactly one-tenth as expensive as Astra for standard operations, providing a clear path for cost-conscious scaling.

The pricing reflects a strategy to provide a more affordable way to run complex tasks. By lowering the barrier to entry for high-speed computer use, OpenAI is positioning Terra as a primary tool for companies that need to automate high volumes of digital work without the premium costs of their most advanced reasoning models.