On this page7 sections
The notes fit on a few lines. The customer notice still needs writing, and somebody has to put it in the right place. Imagine a teammate handing over a maintenance window, a screenshot and one last request: save the notice as a draft for the correct service. It sounds like a small job until you count the changes of attention. Read the notes. Read the screen. Write for someone who was not in the meeting. Find the right record.
That is the territory OpenAI described for GPT-6.1 Terra in its October 26, 2026 release: writing, visual reasoning and computer use, with lower standard token prices than GPT-6.1 Sol. The launch announcement and dated developer announcement introduced API access and rollout in ChatGPT Work and Codex. This is a source-based explanation, not a hands-on Terra test. Our notice and the other tasks below are illustrative examples.
The notes need a model. The status page needs an application.
Hand Terra the notes and it can produce wording. Getting that wording into a status page takes something more. ChatGPT Work, Codex or a developer's own application supplies the information and permitted tools around the model. The model requests an operation; that application performs it. Selecting Terra does not, by itself, connect an inbox or grant control of a computer. The function-calling guide describes this division, and our inference guide follows one request through it.
The notes can arrive as text, the screenshot as an image. According to the model page, Terra accepts both and produces text. It has room for a context of 1,050,000 tokens, the units such as word pieces that a model processes, and can produce up to 128,000 output tokens. That large context window is capacity, not a promise to recall every detail equally well. Audio and video are not native inputs; an application offering those needs another component to handle them.
Twenty minutes, one feature and a time zone
Suppose the maintenance notes read: “Storage migration, Thursday 22:00–22:20 UTC. Reporting may pause. Purchases unaffected. Status page will carry progress.” There is enough here for a short notice. There is not enough to promise that every part of the service will be back by 22:20, or to announce that the whole service will go offline.
This is where a smooth sentence can quietly change the job. Replace “reporting” with “the service” and customers may expect their purchases to fail. Drop “UTC” and the same twenty minutes lands at different times for different readers. The useful draft keeps those small details intact while making the message easy to read: when the work happens, what the customer might notice and where to find updates. A planned window stays a plan, not a guarantee.
OpenAI's selection guidance points to this kind of concise writing from supplied material. Proposals and handovers have the same opportunity: you bring the facts, the audience and the reason for writing; Terra helps reshape the wording. The notes still get the deciding vote when an elegant sentence disagrees with them.
The tiny label that changes the comparison
Now look at the screenshot beside the notes. Two dashboard charts may look almost identical until you notice that one measures seconds and the other milliseconds. Before describing a change, identify the series, time range and units. And keep the time of the screenshot in view: it shows what was on that screen then, not necessarily what the service is doing now.
The same attention to a small visible difference applies beyond dashboards. OpenAI's GPT-6 guidance names product-photo matching and technical drawings as intended uses. Picture two catalog parts with similar outlines but different mounting holes. Finding those holes is useful. Guessing a blurred part number adds confidence without adding information.
Small text, rotated images, graph styles and precise spatial interpretation all appear in the documented vision limitations. If the label is unreadable, a clearer crop or the underlying numbers may do more for the comparison than another round of reasoning. The task is to recover the detail that changes the answer, not to make the model sound certain about a detail nobody can see.
The notice has to survive the Save button
With the notice written, the work moves into the status-page application. In computer use, the model receives a view of the screen, requests an action and uses the resulting view to choose what comes next. The software running the task performs the clicks and typing. The computer-use documentation describes that back-and-forth.
Our draft now has several places to go wrong without a single spelling mistake. The wording may be right but attached to the wrong service. The maintenance time may be entered without saving. Reopening the draft lets you check the service, window and text together. A completion message is useful communication; the saved record is where you find out what actually happened.
There may be no reason to click through the screen at all. If an existing connector or direct API exposes the exact operation and returns a structured result, it may be a better route. Computer use extends the range of applications the assistant can work with. It does not make a dependable existing integration obsolete.
Saving a draft also stops short of publishing it. Keep the organization's approval process in place, with access limits enforced by the application and its environment. The execution guidance puts controls there as well as in the model's instructions. “Be careful” is a reasonable request, but it is not an access-control system.
What the launch trials can tell our teammate
Would Terra finish this little chain of work faster? The launch results give a reason to try it, not the answer for this particular notice. OpenAI reported that Terra at medium effort scored 15 percentage points above GPT-5.6 Terra on Agents' Last Exam at about one-third the estimated cost. In that evaluation it also reported completion more than three times faster than GPT-6 Astra and GPT-6.1 Sol. Those are vendor results for particular professional-software tasks, not a promised speedup for your notice, presentation or browser workflow.
The launch-era system-card addendum supplies a different view. It reported fewer factual errors than GPT-6 Sol and performance comparable with GPT-6.1 Sol on conversations selected because users had flagged factual problems. That is a deliberately difficult collection, not an ordinary slice of production traffic.
It also recorded cases in which Terra found an alternative route that violated an environmental warning in a test without system-level controls. That observation does not tell us a production failure rate. It does explain why the permission to save our draft, and the separate permission to publish it, belong in actual controls rather than in a hopeful sentence at the top of a prompt.
Two cents in, one cent out
The bill is easier to picture with a specific request. Start with standard API text pricing for no more than 272,000 input tokens:
| Token category | GPT-6.1 Terra, per million tokens |
|---|---|
| Uncached input | $1.00 |
| Cached input read | $0.05 |
| Output | $5.00 |
The pricing page lists those input and output rates at half the corresponding GPT-6.1 Sol rates. That comparison is per token, not per finished notice. A model that uses more tokens or needs another attempt can change the saving.
For an illustrative request, let the notes and other supplied material total 20,000 uncached input tokens. Allow 2,000 total billed output tokens. The arithmetic is:
(20,000 ÷ 1,000,000 × $1) + (2,000 ÷ 1,000,000 × $5) = $0.03.
Repeat that identical request a thousand times and the model-token cost is $30 under these assumptions, before tools and other charges. The word “total” in the output allowance matters: billed output includes reasoning as well as visible text. A 2,000-token answer with additional reasoning already exceeds our 2,000-token allowance. The reasoning-token explanation describes that accounting.
Reusing material introduces another distinction. The cheap cached-input rate applies to eligible cache reads, not every prompt that happens to look familiar. Standard cache writes cost $1.25 per million tokens for 30-minute retention or $2.50 for 12 hours. And if input exceeds 272,000 tokens, the entire request moves to double input and cache rates and 1.5 times output rates. These conditions are worth checking before turning our short set of notes into a very large document bundle.
Pick a task whose finish you recognize
For the person with the notes, the first question may simply be where Terra appears. The launch named Plus, Pro, Enterprise and Edu access in ChatGPT Work and Codex, not an identical selector in every ordinary Chat mode or client. The dated availability announcement and current model controls need to be read with your plan, workspace and rollout in mind.
Developers select gpt-6.1-terra; tool calling requires the Responses API. Its effort settings are none, low, medium, high, xhigh and max. Medium is the API default. The model guide recommends starting there and raising effort when the improvement earns the extra time and tokens. A subscription has separate accounting from API charges, so our three-cent calculation cannot tell you how many tasks a subscription will allow.
The choice need not become a permanent allegiance. OpenAI's selection guidance places Terra with scoped fixes, interface work, visual material and writing. Sol is a candidate for more complex projects where cost matters; Astra is the strongest option for demanding reasoning and end-to-end work; Luna sits at the lower-cost end for focused, repeatable tasks. Those are places to begin comparing, not guarantees about your particular job.
Return to the maintenance notice, or choose a chart you can read yourself and a record you can reopen after saving. Compare the usable result and the time spent reviewing it with your usual approach. For our imagined teammate, the helpful finish is quite ordinary: the right facts, a readable notice and the correct saved draft. A lower token bill becomes interesting when it buys that finish with less work left for you.
Sources & context
Sources linked in this article. Read alongside the author’s analysis; a citation does not independently verify a publisher’s claims.
- Launch resultsopenai.com
- Launch availabilitycommunity.openai.com
- function-calling guidedevelopers.openai.com
- Terra pricing conditionsdevelopers.openai.com
- Model-selection guidancedevelopers.openai.com
- GPT-6 guidancedevelopers.openai.com
- Vision limitationsdevelopers.openai.com
- computer-use documentationdevelopers.openai.com
- Computer-use safety guidancedevelopers.openai.com
- system-card addendumdeploymentsafety.openai.com
- OpenAI API pricingdevelopers.openai.com
- Reasoning-token accountingdevelopers.openai.com
- model controlslearn.chatgpt.com
- Plan usage and API pricinglearn.chatgpt.com
