OpenAI announced GPT-6.1 Terra on Oct. 26, presenting it as a fast, efficient model for professional work that combines visual reasoning, computer use and clearer writing. The company says the model is available through ChatGPT Work, Codex and its API; its internal error-test figures remain company-reported rather than independently replicated. The launch details are in OpenAI’s announcement. Independent launch coverage from 9to5Mac and Simon Willison adds pricing and practitioner context.
What OpenAI announced
GPT-6.1 Terra is OpenAI’s newest iteration of its Terra model tier. The launch announcement describes it as a fast, efficient model that combines detailed visual reasoning, precise computer use and more natural writing. OpenAI identifies workplace tasks such as drafting proposals, interpreting technical drawings, filling out forms and editing videos as examples of the work it expects the model to handle. Those examples describe the company’s intended use cases; they are not a guarantee that the model will complete every task correctly or safely.
The main product distinction is a combination of capabilities rather than a single benchmark score. OpenAI is asking customers to consider Terra where they need a model to interpret visual material or interact with computer-based workflows, while also paying attention to speed and cost. The release is a model update, not a newly announced standalone application. OpenAI says customers can access it in ChatGPT Work and Codex, and developers can use it through the API. OpenAI’s announcement is the primary source for those claims.
The phrase “computer use” refers here to the model’s ability to work with computer-based tasks, according to OpenAI’s positioning. The public materials referenced in the supplied coverage do not establish that users can delegate arbitrary, unattended computer activity or that the system is reliable in every interface. In practical deployments, a model’s ability to interpret a screen is only one part of a workflow: the available tools, permissions, safeguards and human review also shape what it can do. The launch should therefore be read as a capability and product announcement, not as evidence that complex office processes can be safely automated end to end.
Where Terra fits in the model lineup
OpenAI presents Terra as a model for high-volume professional work, between lower-cost and higher-capability choices in its family. 9to5Mac describes the lineup as Luna, Terra, Sol and Astra, from the lower-price end toward the higher-capability end. That relative positioning helps explain why the company emphasizes both useful reasoning and efficiency: Terra is meant to occupy a middle ground rather than replace every other model.
That middle position makes the launch relevant to a familiar purchasing decision. A team may prefer a less expensive model for routine classification or simple drafting, while reserving a more capable option for harder analysis. Terra’s stated appeal is that visual interpretation and computer interaction can be useful even in work that is repetitive or high-volume. But the product announcement does not provide a universal rule for choosing among model tiers. The best fit depends on the task, the quality threshold, latency needs, usage volume and actual API or subscription terms available to a customer.
Independent writer Simon Willison noted that a mid-tier model can be difficult to place because it is neither the cheapest nor the most expensive option. He relayed advice from OpenAI’s Ted Sanders, who characterized Terra as potentially stronger than Astra and Sol in technical communication, similar on vision and computer use, and weaker on many other tasks. That is an attributed view from an OpenAI employee, not a neutral comparative evaluation by Willison. It points to the intended trade-off while leaving customers to test their own workloads.
Visual reasoning and computer use
OpenAI’s most prominent capability claims concern visual reasoning and computer use. Its examples span interpreting technical drawings and filling in forms, which imply different kinds of work: one requires extracting and reasoning about information in an image, while the other requires handling a computer-oriented task. The company also mentions editing video and drafting proposals. Those use cases indicate the breadth of the pitch, but the announcement alone does not provide enough detail to compare task success rates across all of them.
The word “reasoning” in a launch description should not be mistaken for proof of dependable judgment. A system may correctly read an image in one instance and misinterpret a small label or context in another. Likewise, successful interaction with a form in a demonstration would not establish reliability across websites, document layouts or permission settings. People considering deployment should evaluate the specific materials and interfaces they use, check how errors are detected, and decide which steps require human approval. These are evaluation questions for adopters, not performance results established by the announcement.
Willison’s post adds a small hands-on observation: he generated an image of a pelican riding a bicycle at a maximum reasoning setting and described the result as good, while noting its cost and generation time. The anecdote offers a concrete glimpse of his own test, but it is not a systematic vision benchmark or evidence about business workflows. It should not be used to infer how well Terra performs on technical drawings, video editing or other image-heavy tasks at scale.
Pricing and benchmark claims
Cost is a central part of OpenAI’s story for Terra. 9to5Mac reports that the API price is 50% lower than GPT-5.6 Terra, while Simon Willison lists GPT-6.1 Terra at $1 per million input tokens and $5 per million output tokens. Willison says that is almost half the price of GPT-5.6 Terra and GPT-6.1 Sol, and about ten times the price of GPT-6 Luna. Those are reported prices at the time of publication; customers should consult current API pricing and applicable product terms before making a budget decision.
Lower per-token pricing does not by itself establish a lower total cost for a completed task. A comparison also depends on how many tokens a workflow consumes, how often it retries, the latency it needs and whether the chosen model completes the task without additional review or correction. These factors are particularly relevant for computer use, where a multi-step workflow may generate more requests than a short text exchange. The supplied sources do not provide a controlled cost-per-success comparison across representative customer workloads, so the price figures should be understood as input and output rates, not a guarantee of savings in operation.
OpenAI also reports results from an internal benchmark of “error-inducing prompts.” According to the figures quoted by 9to5Mac, the error rate at low reasoning effort fell from 53.6% for GPT-5.6 Terra to 16.7% for GPT-6.1 Terra. At maximum effort, OpenAI reported a reduction of 18 percentage points to 4.3%, which it said nearly matched GPT-6 Astra and GPT-6.1 Sol. These are the company’s benchmark results, as relayed in the independent coverage. The provided reporting does not supply a full independent replication, so the figures should not be generalized to all prompts, real-world workloads or error types.
The numbers may still be useful as a signal of what OpenAI chose to emphasize: fewer errors on a test specifically designed to induce errors, alongside price and speed. But judging a model requires more than a single internal measure. Readers would need to know how prompts were selected, what counted as an error, how many trials were run, whether outputs were judged blind and how the test relates to the tasks a buyer cares about. The sources available for this article do not establish those methodological details sufficiently to assess the benchmark independently.
What independent coverage adds
9to5Mac provides a concise account of the launch, including the prior GPT-5.6 Terra baseline, the availability statement and the figures OpenAI shared about pricing and its internal error test. Its article is useful as independent reporting on the announcement, but much of the performance evidence it describes originates with OpenAI. The existence of independent coverage does not convert company-supplied benchmark figures into independently measured results.
Simon Willison’s October 26 post takes a more evaluative, practitioner-oriented approach. It discusses the model’s place between cheaper and more expensive tiers, relays an OpenAI employee’s characterization of relative strengths, and describes Willison’s own pelican image test. He also reports token prices and timing for that generation. His post is a distinct independent perspective, but it is a short link post rather than a broad comparative study. It offers context and an example, not a comprehensive assessment.
Together, these sources support the core news: OpenAI announced GPT-6.1 Terra on October 26 and positioned it around fast professional work, visual reasoning, computer use and lower cost. They do not establish that Terra universally outperforms its predecessor, that every advertised task is production-ready, or that customers will realize a specific savings rate. Those stronger conclusions would require evidence beyond the launch materials and coverage available here.
Availability and open questions
OpenAI’s announcement and 9to5Mac’s report say GPT-6.1 Terra is rolling out through ChatGPT Work, Codex and the API. The sources do not spell out all eligibility rules, regional differences, account-level timing or the precise limits of access. Simon Willison’s discussion also distinguishes between subscribers and API developers when describing who might value the model, but it does not establish a complete availability matrix. Users should check the model selector or API documentation associated with their account for current access details.
The announcement leaves important questions for prospective users. It does not, in the supplied source material, provide an independent benchmark suite covering visual reasoning and computer use, a broad head-to-head evaluation across common business tasks, or a public estimate of typical workflow costs. Nor does a named use case reveal what safeguards are available when the model interacts with a form, edits a file or works with sensitive content. These gaps are not evidence of failure; they describe what the launch information alone cannot answer.
OpenAI has published a system-card addendum dated October 26 for GPT-6.1 Terra. It describes the document as an update to baseline safety metrics and evaluation results and says the model is treated as Critical in cybersecurity capability and High for biological and chemical capability under OpenAI’s Preparedness Framework. That classification is OpenAI’s own assessment, not an independent finding. The addendum also cautions that evaluations performed in a research environment or through the API may differ somewhat from production ChatGPT because system prompts, tools and effort settings can vary. This is a useful reminder that model behavior depends in part on the environment in which it is evaluated and used.
What this means for users
For businesses, developers and individual subscribers, the release creates another option to test rather than an automatic reason to migrate. Teams with image-heavy or computer-based tasks may want to evaluate Terra against their present workflow, using representative examples and a clear definition of success. A useful pilot would track accuracy, completion time, token use, correction effort and cases where a human must intervene. The sources do not report such customer trials, so any expected benefit remains a hypothesis until measured in the relevant setting.
It is also sensible to separate a model’s apparent ability from an organization’s authority to let it act. A system that can interpret a screen may still need constrained permissions, confirmation before consequential actions and a way to recover from mistakes. The announcement’s examples are framed as workplace tasks, but it does not claim that every task should be delegated without supervision. Organizations remain responsible for deciding which actions the model may take and which require review.
For developers, the price figures could make Terra worth evaluating in workloads where the alternatives are either too costly or less capable on relevant image and computer-use tasks. Yet the meaningful comparison is the cost of an acceptable result, not simply the listed price per million tokens. For subscribers, Willison’s reported perspective suggests the potential appeal may be faster computer use or clearer communication, rather than a dramatic leap for every interaction. Both are attributed assessments, and individual experience can vary with task, settings and account access.
The clearest conclusion is therefore narrow. GPT-6.1 Terra is a new OpenAI model release, publicly announced on October 26, with positioning and reported access across ChatGPT Work, Codex and the API. OpenAI emphasizes visual reasoning, computer use and efficiency, while independent coverage adds pricing context and limited hands-on perspective. Whether those claims translate into reliable, cost-effective work for a particular user is still something that must be tested rather than assumed.
FAQs
What is GPT-6.1 Terra?
It is OpenAI’s new Terra-tier model, which the company describes as a fast, efficient model for professional work with visual reasoning and computer-use capabilities. OpenAI also says it offers clearer, more natural writing.
When was GPT-6.1 Terra announced?
OpenAI’s announcement and the independent articles reviewed here are dated October 26, 2026.
Where does OpenAI say it is available?
The company says the model is rolling out in ChatGPT Work and Codex and is available through its API. The supplied sources do not provide a detailed account of regional or account-level availability.
How much does the API model cost?
Simon Willison reports a price of $1 per million input tokens and $5 per million output tokens. Prices and terms can change, so developers should confirm the current rate in OpenAI’s API materials before use.
Are the reported benchmark improvements independently verified?
No independent replication is established by the sources reviewed here. The error-rate figures are attributed to OpenAI’s internal benchmark and should be treated as company-reported results.
Does the launch prove that Terra is better for every work task?
No. The announcement describes intended strengths and selected company tests. Performance depends on the task and deployment, and the available independent coverage is not a comprehensive evaluation of all workflows.