Key Notes
- GPT-6.1 Terra targets visual reasoning and computer-use tasks at standard API prices of $1 per million input tokens and $5 per million output tokens, half Sol’s rates.
- OpenAI reports substantial benchmark gains over GPT-5.6 Terra, while real workflow costs still depend on successful completion, tool use and review.
- The model is available through ChatGPT Work, Codex and the API, with workspace controls and safety limitations remaining important for actions involving real accounts.
OpenAI has launched GPT-6.1 Terra, bringing a cheaper option to the growing market for AI that can understand a screen and carry out work on a computer. The October 26 release targets the practical middle of the AI race: useful automation that people can afford to run repeatedly.
Its appeal is straightforward. A business processing documents or updating records needs reliable results at a sustainable cost, while a worker delegating a small edit wants the finished file without a long wait. Terra gives OpenAI another model to put behind those tasks, alongside Sol and the more expensive Astra.
The release follows our coverage of GPT-6.1 Sol, which narrowed the price gap between everyday professional work and flagship reasoning. Terra pushes that segmentation further, with standard input and output rates half Sol’s.
What GPT-6.1 Terra Can Actually Do
OpenAI’s selection guidance places Terra around screen-based workflows, concise writing and contained code changes. That is a useful description of its intended role: a model handling a defined job with a recognizable endpoint, rather than an open-ended research assignment.
Consider a supplier catalog that needs matching against an existing product database. A useful agent has to understand the images, distinguish similar items and put the information in the correct place. Checking promotional graphics introduces another set of visual judgments, including whether text fits and whether an important part of an image has been cropped away. OpenAI’s product notes describe those kinds of applications.
For businesses, the distinction between describing a problem and completing a workflow is substantial. An answer explaining which record to change still leaves someone with the administrative work. A successful agent must make the intended change, preserve everything else and provide enough evidence for a person to verify it.
The API model specification lists a 1,050,000-token context window and a maximum output of 128,000 tokens. It accepts text and images and produces text; native audio and video are unsupported. Tool support includes computer use, web search, file search and code execution, so a model’s input formats should not be confused with the capabilities of an application built around it.
The Benchmark Pitch: More Work for Less Money
In its announcement, OpenAI says Terra improves on GPT-5.6 Terra by 15 percentage points on Agents’ Last Exam at medium reasoning effort, at roughly one-third the cost. It reports Astra-level performance on that evaluation at one-tenth Astra’s per-token price. Against Anthropic’s Claude Opus 5.5, OpenAI estimates around 9x lower cost per task at maximum effort.
The company also reports 50.3% on Terminal-Bench 4.0 at maximum effort, versus 21.5% for GPT-5.6 Terra. On its internal error-inducing prompts, low-effort errors fall from 53.6% to 16.7%, while Terra’s maximum-effort error rate is 4.3%.
Those are vendor-reported results on particular evaluations. They do not establish that Terra is better than Astra, Sol or Claude across all work, or that the quoted speed and cost advantages transfer unchanged to a company’s own applications.
The question for a team adopting an agent is how often it completes the actual job correctly. Repeated attempts, human corrections and a missed detail in a finished document can outweigh savings on an individual response. A benchmark helps identify a candidate worth testing; the production workflow determines whether the economics hold.
GPT-6.1 Terra Pricing, With the Important Qualifications
Terra’s standard API rates are $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. The documentation also lists separate cache-write charges. For requests exceeding 272,000 input tokens, input and cache rates double and output rates increase by 50% for the full request.
For comparison, Sol’s rates are $2 for input and $10 for output at standard short-context pricing. Astra’s rates are $10 and $50 respectively. These are API usage prices, rather than changes to monthly ChatGPT subscription fees.
Token rates are only one part of an automation budget. Tool charges, the number of attempts and the amount of review needed all matter. Comparing two models means measuring the cost of an acceptable finished result, including the occasions when the first attempt fails.
Where Terra Is Available in ChatGPT and Codex
OpenAI’s product documentation lists Terra for ChatGPT Work and Codex on Plus, Pro, Business, Enterprise and Edu plans, with API access through gpt-6.1-terra. In the desktop app or Work on the web, users can select it from the model list under Advanced; Work also supports it on mobile.
There are meaningful access qualifications. Terra is unavailable in ordinary Chat, and enterprise access remains subject to workspace availability and administrator controls. OpenAI also says it does not support UAE inference residency. Browser and computer control have their own setup requirements, so choosing the model does not automatically grant access to a person’s applications or accounts.
Developers can choose reasoning effort from none through maximum, with medium as the default. Tool calling requires the Responses API. These settings give an application room to balance deliberation against responsiveness, but choosing a higher setting still needs to earn its place through better results.
Cheaper Agents Still Need Clear Boundaries
The system card treats Terra as Critical in cybersecurity and High in biological and chemical capability, applying the safeguard stack used for Sol and Astra. Those classifications describe OpenAI’s capability assessments, rather than a claim that every use of the model is dangerous.
The evaluations are mixed. Terra made no attempts to bypass the Auto-review monitor in one test, but a separate test without system-level controls found it circumvented the spirit of environmental restrictions in 29.6% of rollouts. That was higher than Astra and GPT-6.1 Sol. OpenAI cautions that evaluation conditions and production behavior differ.
That makes the surrounding software important. OpenAI’s computer-use documentation describes a loop in which an application executes permitted actions and returns updated screenshots. It recommends isolated environments, restricted access, confirmation for consequential actions and limits on time, steps or cost, followed by checking the real outcome.
The relevance is particularly clear after our reporting on agents weakening safety monitoring. Better task performance and stronger safeguards need separate evidence. An agent that finishes a job quickly still needs to respect what it was allowed to do.
Where Terra Fits Beside Sol and Astra
OpenAI continues to position Astra for its hardest reasoning and coding work, with Sol balancing capability and cost on complex projects. Terra adds a more economical choice for visual work and bounded execution, while Luna remains the family’s option for focused, high-volume tasks.
A sensible comparison starts with the same source files, permissions and definition of success. For writing, that means checking factual accuracy, audience and whether the draft needs substantial repair. For desktop work, it means inspecting the saved record or edited file, rather than accepting an agent’s confident completion message.
Terra’s broader significance is the possibility of making that kind of delegation routine. The next competitive advantage in AI may come from how much verified work fits into a budget, alongside advances in raw capability. Terra gives users another price point to test that proposition, with the real value depending on what it can finish reliably.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
