A construction drawing is displayed on the screen; the model needs to understand the annotations before being transformed into a 3D model using specialized software. On another front, an animation requires meticulous inspection frame by frame, and in the end, a repair log with time stamps must be submitted. OpenAI has arranged for GPT-6.1 and Terra to not just chat with others, but to sit in front of their computers and get the work done.

The most notable change in this model is not the scale of its parameters, but rather its price. The standard pricing provided by OpenAI is $1 per million input tokens, $0.05 for caching inputs, and $5 for outputs, which is 50% lower than the standard prices of GPT-5.6, Terra, GPT-6.1, and Sol. For those who only ask questions occasionally, this price difference may not be significant; however, for teams that process invoices daily, enter data into tables, check pages, and perform hundreds or thousands of tasks, this price difference can directly affect whether an automated process can continue to operate in the long term.

The price reduction occurs at the "doing things" level.

OpenAI does not describe Terra as a scaled-down chat model. Instead, it emphasizes the combination of visual understanding, computer operations, and natural writing: the model not only needs to understand the interface but also know where to click next, and it must organize the results into deliverable text. In the past, such a combination often required piecing together image models, automated scripts, and language models, and the maintenance costs came not only from tokens but also from the need for manual intervention after failures.

The official tested professional tasks such as architectural modeling and film shot-by-shot inspection on Agents ' Last Exam. At a medium level of reasoning intensity, GPT-6.1 Terra outperformed GPT-5.6 Terra by 15 percentage points, with a cost per task that was about one-third of the latter; its performance was close to that of GPT-6 Astra, yet the price per token was only one-tenth of the latter. At the highest level of reasoning intensity, OpenAI claimed to complete tasks three times faster than Astra and Sol.

What truly changes in this approach is the division of labor among models. In the past, teams would often use the strongest model for all steps to ensure success rates, but this meant they had to reduce the number of calls to those models. Now, a more feasible approach is to let Terra handle the main processes that involve a large volume of work, multiple interfaces, and high repetition, while reserving Sol or Astra for the most complex judgments. OpenAI also maintains its own boundary: for the most complex coding tasks, Sol and Astra are still the more suitable choices.

The improvement in coding capabilities makes this division of labor more convincing. Terminal-Bench 4.0 tests whether the model can solve real engineering problems at the command line; GPT-6.1 and Terra complete 50.3% of the tasks at maximum inference intensity, while GPT-5.6 and Terra achieve a public score of 21.5%. This is not proof that "all tasks can be entrusted to it," but it does indicate that the capability gap between low-cost and high-cost models is narrowing.

Low prices do not eliminate factual errors.

OpenAI also released a set of data that is even more concerning. In internal tests designed to "induce errors," the rate of factual errors at low levels of reasoning intensity dropped from GPT-5.6 Terra's 53.6% to 16.7%; at the highest level of reasoning intensity, it dropped to 4.3%. Although the reduction is significant, 16.7% still indicates that even when costs are minimized, factual verification cannot be completely eliminated.

For business processes, the intensity of reasoning is not just about speed; it is also a form of risk budgeting. Invoices with a fixed format can be verified using rules at a lower level of intensity. However, content that involves contracts, financial judgments, or external releases requires a higher level of structured checking, or it should be reviewed by someone else. Once models become more cost-effective, a common misconception is to think that the savings in costs mean that supervision can be eliminated.

Terra also meets the "Critical" threshold for network security in the OpenAI preparation framework, and these protective measures are adopted for development in GPT-6 and Astra. However, this does not mean that the model is already secure enough to be freely integrated into production systems; on the contrary: the more it can interact with real computers and networks, the more important it becomes to have proper permission isolation, logging, and manual intervention in place.

The entry point is placed at Work and Codex.

The release path reveals the positioning of OpenAI for this model. GPT-6.1 Terra will be available to users of Plus, Pro, Enterprise, and Edu from the release date onwards, but the entry points are located at ChatGPT Work and Codex; developers can call API through `gpt-6.1-terra`. It does not appear first in the regular chat box but rather in a work environment that includes tasks, files, tools, and execution processes.

The same subscription quota can complete more tasks on Terra, which turns the "model selection" issue from a ranking concern into an operational one. In the future, a team may no longer ask which model is overall the strongest, but rather: which steps must be done with Astra, which steps can be delegated to Terra, and which results require human signature. The decrease in cost allows for more computer operations to be included in the budget, but the real new task is to clearly design the processes for authorization, review, and failure handling.

What GPT-6.1 and Terra do is to reduce the unit cost of AI. They haven't caused the strongest models to lose their position, but they have given high-frequency, routine tasks the opportunity to run on a pipeline scale for the first time. The new dividing line is no longer just "human or AI"; rather, it's about which actions can be completed automatically at low cost, and which decisions still require more expensive reasoning and human involvement.