OpenAI released GPT-6.1 Terra on October 26, 2026, and if you run a small business, this is the model launch that actually touches your workday. Terra is a fast, mid-tier AI model built to operate computers: it reads screens, clicks buttons, fills in forms, and moves between apps. Its API price is $1 per million input tokens and $5 per million output tokens, half what its predecessor cost. And here is the catch most headlines skip: you cannot just open regular ChatGPT and start using it. Terra lives in ChatGPT Work and Codex, for paid plans only, and it is rolling out gradually. This article translates the launch into plain business terms: what Terra does, what it costs per finished task, who can actually use it today, and whether you should switch your agent’s model this week.
What GPT-6.1 Terra actually is
Terra sits in the middle of OpenAI’s lineup. From cheapest to most capable, the family runs Luna, Terra, Sol, and Astra. GPT-6.1 Terra replaces GPT-5.6 Terra in that middle slot. OpenAI describes it as a model that pairs sharper visual reasoning and precise computer use with clearer, more natural writing, aimed at the communication and computer-based tasks that fill a workday.
That description matters because it tells you what Terra is for, and what it is not for. It is not a research breakthrough model. It is not the model you pick for the hardest coding problem your team has ever faced. OpenAI itself says Sol and Astra remain the better choice for the most demanding coding tasks. Terra is the workhorse: the model you hand the repetitive, screen-based jobs that eat an afternoon.
OpenAI’s own examples are deliberately unglamorous, and that is the tell. The launch materials point to reading incoming invoices and typing the details into accounting software, drafting proposals, filling out forms, reading technical drawings, and editing video. One early write-up put it well: the benchmark is not the story; the price per task is. For a small business, that is exactly right. You do not pay for benchmark scores. You pay for finished work.
A few technical facts set the frame. Terra’s context window is about 1.05 million tokens, with up to 128,000 tokens of output. It accepts text and image input and returns text. Its knowledge cutoff is June 1, 2026. Tool calling, including the computer-use tool, requires the Responses API. Through the API, the model name is gpt-6.1-terra, and it is also available through Amazon Bedrock and GitHub Copilot.
One detail deserves a clear flag, because it trips people up. Terra can operate video-editing software through its screen, but the API itself does not accept video as input. Those are two different things. If a vendor tells you Terra “handles video,” ask which one they mean: driving the app, or ingesting the footage. The first is computer use. The second is not what this model does.
The price cut, in plain numbers
Terra’s headline API price is $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. That is 50% below GPT-5.6 Terra’s July pricing on input, and about 58% lower on output. It is also half of GPT-6.1 Sol’s price, and roughly a quarter of Claude Opus 5.5’s.
Here is the dated price trail, drawn from a pricing tracker that has followed these model lines since July:
| Model line | July 2026 (input / output / cache) | October 2026 (input / output / cache) |
|---|---|---|
| Claude Opus | $15 / $75 / $1.50 | $4 / $20 / $0.20 |
| Claude Sonnet | $3 / $15 / $0.30 | $2 / $10 / $0.10 |
| OpenAI Sol | $5 / $30 / $0.50 | $2 / $10 / $0.10 |
| OpenAI Terra | $2.50 / $15 / $0.25 | $1 / $5 / $0.05 |
The last column is the one that matters most for agent work, and it is easy to miss. Cached input is what an agent re-reads on every step of a long task: the screen, the instructions, the conversation so far. Terra’s cache reads fell from $0.25 to $0.05 per million tokens, an 80% drop. For a computer-use agent that re-reads context dozens of times per task, that is where the real savings pile up.
But token prices still mislead, because a computer-use task is not one prompt and one answer. It is dozens of screenshots, clicks, and corrections, and every step re-reads context. What you actually pay for is the finished task. OpenAI published a cost-per-task chart from its Agents’ Last Exam benchmark, a set of professional tasks like turning architectural drawings into a 3D model. Read with care, since these are OpenAI’s own runs, but the shape of the numbers is instructive:
| Model | Effort | Score | Cost per task |
|---|---|---|---|
| GPT-6.1 Terra | Medium | 57.6% | $0.29 |
| GPT-6.1 Terra | Max | 60.0% | $1.01 |
| GPT-5.6 Terra | Medium | 42.6% | $0.84 |
| GPT-6.1 Sol | Medium | 53.8% | $1.31 |
| GPT-6 Astra | Medium | 57.6% | $4.82 |
| Claude Opus 5.5 | Max | 63.2% | $8.82 |
Two things stand out. First, Terra at medium effort ties Astra at medium effort (both 57.6%) for $0.29 instead of $4.82 per task. Second, Opus 5.5 still scores highest at max effort, 63.2% against Terra’s 60.0%, but costs about nine times as much per task. So the honest summary is this: Terra is the cheapest per task, not the smartest. Independent testers at Artificial Analysis scored Terra 48.1 on their Intelligence Index at $0.40 per task, below Claude Opus 5.5 and Gemini 4 Argon, while noting it pushes the cost-efficiency frontier. Cheap and best are different claims, and Terra only makes the first one.
What that means for a monthly bill
Let us turn those per-task figures into a monthly budget, because that is the number a business owner actually approves. Assume an agent handling routine back-office tasks, five days a week, roughly 22 working days a month:
| Volume | Terra at medium ($0.29/task) | Terra at max ($1.01/task) | Opus 5.5 at max ($8.82/task) |
|---|---|---|---|
| 20 tasks/day (440/month) | $128 | $444 | $3,881 |
| 50 tasks/day (1,100/month) | $319 | $1,111 | $9,702 |
| 100 tasks/day (2,200/month) | $638 | $2,222 | $19,404 |
Those are estimates built on OpenAI’s published chart, not a quote, and your real bill depends on task length, screenshots, and retries. But the gap is so large that the conclusion survives rough edges: at this price, an always-on agent doing boring chores in the background stops being a demo and starts being a budget decision. The flip side is equally real. A dollar a task is only cheap if the task comes back right. For anything that touches money or customers, the human review step is still part of the price, and you should budget for it.
A smaller, concrete example helps too. A short task using 10,000 input tokens and 2,000 output tokens costs about $0.02 on Terra at standard rates. That is the scale of a single form-fill or a short document pass. The costs that surprise people are the long, screenshot-heavy jobs and the retries, not the single prompt.
Five workday jobs Terra actually fits
Forget the benchmark names. Here is how Terra maps to jobs a small team already pays someone to do, or skips because nobody has time.
1. Invoice and receipt entry
This is OpenAI’s headline example, and it is the strongest fit. An agent reads an incoming invoice, pulls out the vendor, date, line items, tax, and total, and enters them into your accounting software. The work is high-volume, rule-bound, and easy to check: the numbers either match the invoice or they do not. Start here, because verification is cheap and mistakes are visible before money moves.
2. Form filling and data transfer
Permit applications, vendor onboarding forms, insurance certificates, grant portals: the same fields, typed into a different system, again. Terra can move data between apps, but every field that touches legal or financial records needs a human glance before submission. Treat the agent as a fast typist, not a signer.
3. Proposal and document drafting
Terra’s writing is pitched as clearer and more natural than its predecessor’s, and its big context window can hold a whole proposal, a price list, and a client’s brief at once. It drafts; you edit. The value is not that the first draft is perfect. It is that the first draft exists in four minutes instead of forty, and your time goes to judgment instead of formatting.
4. Catalog and drawing work
Matching product photos to catalog entries is one of the API use cases OpenAI lists, and reading technical drawings is another. If you sell products online or work with plans and schematics, Terra’s visual reasoning is the feature to test. Run a batch of fifty items, check the error rate yourself, and decide from your own numbers rather than a vendor chart.
5. Repetitive multi-app chores
Any job that is really “open this, copy that, paste it there, save, repeat” is Terra’s home turf. The rule of thumb: if you can write the job down as ten steps that never change, an agent can probably do it. If the job needs a phone call, a negotiation, or a judgment call halfway through, keep a human in the loop at that step.
The trap: Terra is not in regular ChatGPT
This is the part that will confuse the most people, so let me be blunt. If you open the ChatGPT app you already use and look for Terra, you will not find it. Terra is rolling out in ChatGPT Work and Codex, for Plus, Pro, Business, Enterprise, and Edu plans. Free and Go plans are not listed. In ChatGPT Work, you pick it from the model list; in Codex, you run it from the desktop app or the command line.
Availability also depends on your plan, your workspace settings, and how far the rollout has reached your account. OpenAI’s own release notes say access varies during rollout. So if a colleague has Terra and you do not, that is probably rollout timing, not a mistake on your end. Check your model picker in ChatGPT Work or Codex before assuming you are locked out, and check again in a week if it is not there yet.
There is a related pricing wrinkle worth knowing before you build on it. Prompts above 272,000 input tokens are charged at higher rates: double for input and cache, and 1.5 times for output. Cache writes cost $1.25 per million tokens with 30-minute retention. Batch and Flex modes run 50% lower, and a fast mode runs at double. None of this matters for a short form-fill. All of it matters if you feed an agent an entire archive on every task. Keep prompts tight, and let caching do the heavy lifting on repeated work.
Should you switch your agent’s model this week?
Here is a decision checklist, not a sales pitch. Switch a workflow to Terra this week if most of these are true:
- The job is screen-based, repetitive, and easy to verify (invoices, forms, transfers, catalog matching).
- You already run the job through an agent platform or the API, so switching models is a setting, not a rebuild.
- Your current model bill is large enough that a 50% cut shows up in your budget, or your task volume is high enough that per-task cost dominates.
- You can add a human review step for anything touching money, customers, or legal records.
- You are on a paid plan with access to ChatGPT Work or Codex, or you use the API directly.
Hold off if these are true instead:
- Your hardest work is complex coding, where OpenAI itself points you to Sol or Astra, and where independent results show Claude Sonnet 5.5 ahead on some coding benchmarks.
- Your workflow depends on regular ChatGPT chat, where Terra simply is not available.
- Your agent touches production systems with no approval step. Fix the approval step before you change the model.
- You need fine-tuning, which Terra does not support.
The smart move for most small teams is a split, not a swap. Move the boring, high-volume jobs to Terra this week. Keep the hard, high-stakes jobs on your best model. Then measure both for a month: cost per finished task, error rate, and how much human review time each one needs. The model that wins on your own numbers is the one to keep, whatever the launch charts say.
The honest limits
No launch is all upside, and Terra has limits worth naming before you commit.
First, the benchmarks are OpenAI’s. The company reports that on its internal set of error-inducing prompts, the error rate at low reasoning effort fell from 53.6% with GPT-5.6 Terra to 16.7% with Terra, and to 4.3% at max effort. On Terminal-Bench 4.0 at max effort, Terra solves 50.3% of tasks, more than double its predecessor’s 21.5%, but behind GPT-6.1 Sol and Claude Sonnet 5.5 on that test. Independent reruns had not been published when this article was written, so treat the vendor numbers as a direction, not a verdict.
Second, reasoning tokens cost money even though you never see them. Internal reasoning tokens count against context and are billed as output tokens. A task that looks short on the surface can bill longer underneath, especially at high reasoning effort. This is why cost per task, measured on your own jobs, beats token-price comparisons every time.
Third, OpenAI rates Terra at the Critical threshold for cybersecurity capability under its Preparedness Framework, with the same safeguards developed for Astra. A model that can operate a computer cheaply and quickly is useful to everyone, including people you would rather it did not help. That is not a reason to avoid it. It is a reason to sandbox it: give agents the narrowest access they need, require approval before external actions, and keep audit logs. The week’s other AI news, from agent sandboxing to lab safety incidents, points the same way.
Finally, Terra is closed-weight and OpenAI-hosted only. If your business needs a model you can run on your own hardware, or one with open weights, look at the open-weight options instead. Terra’s deal is price and convenience, not control.
A 30-minute test plan
If you want to try Terra without betting a workflow on it, here is a half-hour plan that gives you a real answer.
- Pick one job you already do weekly: ten invoices, twenty catalog items, or one recurring form. Write down the steps, exactly as you do them.
- Run five examples through Terra in ChatGPT Work or Codex, or through your agent platform with the model set to
gpt-6.1-terra. - Check every output against the source document. Record the error rate and the time it took, including your review time.
- Note the cost from your usage dashboard, not the token estimate. Divide by five finished tasks. That is your real cost per task.
- Decide: if the error rate is lower than your tolerance and the cost per task beats your current setup, move the next batch over. If not, keep Terra on the shelf and retest after the next update.
That is it. Five examples, one honest measurement, and you will know more than any benchmark chart can tell you.
Frequently Asked Questions
Is GPT-6.1 Terra available in regular ChatGPT?
No. Terra is rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu plans. It is not in standard ChatGPT chat, and Free and Go plans are not listed. Availability also depends on workspace settings and rollout progress, so it may appear for some accounts before others.
How much does GPT-6.1 Terra cost?
Through the API, standard pricing is $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. That is half the price of GPT-6.1 Sol and about 50% below GPT-5.6 Terra. In ChatGPT Work and Codex, OpenAI says the same plan allowance covers substantially more tasks with Terra than with Sol or Astra, but your actual cost depends on task length and reasoning effort.
Is Terra better than Claude Opus 5.5?
It depends on what “better” means. On OpenAI’s Agents’ Last Exam chart, Opus 5.5 scores higher at max effort (63.2% vs 60.0%) but costs about nine times as much per task ($8.82 vs $1.01). Independent testing also scores Terra below the frontier models on overall intelligence. Terra wins on cost per task; Opus wins on raw score. For routine work, cost usually decides. For the hardest work, score does.
What jobs is Terra best for?
High-volume, screen-based, easy-to-check work: invoice entry, form filling, data transfer between apps, catalog matching, and reading technical drawings. It is a workhorse for repetitive computer tasks, not the first choice for the most complex coding or for jobs that need open weights or self-hosting.
Does Terra accept video input?
No. Terra accepts text and image input and returns text. It can operate video-editing software through the screen using computer use, but that is different from accepting video files as input. Ask any vendor making video claims to say which one they mean.
Should small businesses switch to Terra now?
Move the boring, high-volume, verifiable jobs first, and keep hard or high-stakes work on your best model. Measure cost per finished task and error rate on your own jobs for a month before moving anything that touches money or customers without review. If your workflow lives in regular ChatGPT chat, Terra is not an option yet, so there is nothing to switch.
For more on how AI agent costs actually land on a freelancer’s budget, see how AI agents change freelancer costs. If you are weighing the whole model lineup rather than one launch, our flagship model pricing comparison puts Terra’s price cut in context. For the official launch details and the vendor’s own benchmark charts, see OpenAI’s Terra launch coverage, BleepingComputer’s launch report and the dated price trail for every major model line.