Direct answer: OpenAI has released GPT-6.1 Terra, a faster and roughly half-price successor to GPT-5.6 Terra built for everyday computer work. It reads what is on screen and takes action across apps — clicking, typing, navigating websites and filling in forms — and it is rolling out to ChatGPT Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, as well as to developers through the OpenAI API.
What happened
On Monday 26 October 2026, OpenAI announced GPT-6.1 Terra through a blog post and a post on X. The launch is a middle-tier move, and that positioning matters more than the version number suggests. GPT-6.1 Terra is not OpenAI’s flagship. It sits below GPT-6.1 Sol and GPT-6 Astra in the lineup, and it is pitched as the everyday model: the one you reach for when the job is routine computer-based work rather than the hardest coding problem or the deepest research project.
The headline claim is a combination that rarely arrives together: better performance and a lower price. OpenAI says standard API prices are around 50% lower than GPT-5.6 Terra, reported as $1 per million input tokens and $5 per million output tokens. At the same time, OpenAI reports that the model finishes tasks over three times faster on average than Astra and GPT-6.1 Sol. Speed and price are the story here. For a business paying per token, or for a team working through a long queue of small jobs, both numbers change what is worth automating.
There is also a quieter change that will matter to regular ChatGPT users. OpenAI says the usage allowance goes further in ChatGPT Work and Codex than it did with GPT-5.6 Terra, Sol or Astra. In plain terms, the same plan should stretch further on the kinds of agentic, computer-use tasks this model is built for. That is worth noting because the usual frustration with agent-style tools is not that they fail outright, but that they run out of steam halfway through a long job, or that using them heavily gets expensive fast. A model that is cheaper per token, faster per task, and more generous with allowance is aimed squarely at that frustration.
The timing fits a pattern that has been building all year. AI companies have been racing to turn chatbots into workers: systems that do not just answer questions, but open the right app, find the right record, and carry a task through to the end. GPT-6.1 Terra is OpenAI’s latest step in that direction, and its emphasis on computer use — reading screens and acting on them — puts it in the same conversation as the broader shift toward AI agents that small businesses keep hearing about but have found hard to use in practice. If you want the wider picture on that shift, our guide to AI agents for small businesses in the UK explains what these systems are and where they genuinely help.
It is worth being clear about what this launch is not. It is not a new ChatGPT subscription tier, and the $1 and $5 figures are API prices, not a new monthly price for Plus or Pro. It is not a replacement for OpenAI’s top models on hard tasks, either. OpenAI itself still points users toward GPT-6.1 Sol or GPT-6 Astra for complex coding and in-depth research. GPT-6.1 Terra is the workhorse: good enough for a large share of daily computer work, fast, and cheap enough that trying it on a real workflow does not feel like a gamble.
How GPT-6.1 Terra actually works
Computer use sounds simple when you say it out loud — the model looks at the screen and does things — but it is one of the harder problems in applied AI, because real software was designed for human eyes and human hands. Buttons move, layouts change, forms have quirks, and a single misclick can send a task down the wrong path. GPT-6.1 Terra is built to handle that messiness.
OpenAI describes the core capability in straightforward terms: the model understands what is on screen and takes action across apps. That covers reading screens, clicking through applications, navigating websites, drafting messages, filling out forms, interpreting technical drawings, editing videos, creating images, and automating repetitive tasks. The breadth of that list is the point. This is not a model that only writes text and stops. It is meant to carry a task across the boundary between reading and doing.
In practice, a computer-use task usually breaks down into a loop. First, the model takes in what is visible: the layout of a page, the labels on buttons and fields, the contents of a document or an image. Then it works out what to do next: which field to fill, which menu to open, which file to attach. Then it acts: a click, a keystroke, a scroll, a typed entry. Then it checks what changed on screen and decides whether the step worked or whether it needs to correct course. The quality of a computer-use model lives in that loop. A weak model misreads a label, clicks the wrong thing, and then confidently carries on as if nothing went wrong. A strong one notices the mistake, backs up, and recovers.
That is why the benchmark figures OpenAI reports are worth a close look, even with the caveat that they are OpenAI’s own numbers. On Agents’ Last Exam, OpenAI says GPT-6.1 Terra scores 15 percentage points higher than GPT-5.6 Terra at medium reasoning effort, at about one-third of the cost. On the same benchmark, OpenAI says it matches GPT-6 Astra at roughly one-tenth of the price per token. On an internal benchmark, OpenAI reports that the error rate fell from 53.6% for GPT-5.6 Terra at low effort to 16.7% for GPT-6.1 Terra at low effort, and down to 4.3% at maximum effort. Those are large jumps, and the direction is consistent: fewer errors, less cost, faster completion.
Coding is another area where OpenAI reports a sharp improvement. On Terminal-Bench 4.0, a coding benchmark, OpenAI says GPT-6.1 Terra reaches 50.3% at maximum effort, compared with 21.5% for GPT-5.6 Terra. OpenAI also reports strong results on OSWorld 2.1, a desktop automation benchmark, and on MMMU-Pro, a visual reasoning benchmark, describing the model as close to Astra and Sol despite the lower cost. Taken together, the picture is a model that is not just cheaper, but genuinely more capable at the screen-based, multi-step work that used to trip up this class of model.
The professional examples OpenAI cites give a sense of the intended range. One is turning architectural drawings into a 3D model — a task that requires reading a technical drawing properly, not just recognising that it is a drawing. Another is watching an animated film shot by shot to write a time-coded log for a restoration team, which is painstaking, repetitive viewing work that benefits from patience more than brilliance. A third is reading incoming invoices and typing the details into accounting software, which is exactly the kind of unglamorous, error-prone admin that eats up small-business hours every week. None of these tasks is flashy. All of them are the kind of work that, done well by a machine, frees a person for work only a person can do.
There is a useful way to think about where a model like this fits. It is not a replacement for your judgement, and it is not a member of staff. It is closer to a very fast, very literal junior assistant: excellent at following a clear process across apps, quick to tire nobody, but in need of clear instructions, sensible permissions, and someone checking the output. The businesses that get value from it will be the ones that treat it that way — giving it well-defined jobs, keeping a human in the loop, and measuring whether the result is actually right, not just fast.
What it means for UK small businesses
For a UK small business, the most important sentence in this launch is not about a benchmark. It is the price cut. Computer-use automation has been interesting for a while, but at flagship-model prices, many everyday jobs did not clear the bar: the saving was real, but the cost of running the model ate too much of it, and the error rate meant a person had to redo too much of the work anyway. A model that is around half the price of its predecessor, faster per task, and reportedly far less error-prone changes that arithmetic for a whole class of routine work.
Start with the obvious candidate: invoices and admin. Reading an incoming invoice and typing the supplier, date, amounts and VAT details into accounting software is exactly the example OpenAI gives, and it is a weekly reality for trades, retailers, agencies and professional services alike. Done by hand, it is slow and it invites typos, especially at the end of a long day. Done by a computer-use model with a person checking the entries before anything is submitted or paid, it becomes a queue that clears itself. The same logic applies to form filling: supplier onboarding forms, customer intake forms, booking details, compliance checklists, and the endless copy-and-paste between a CRM, a spreadsheet and an email inbox.
A second area is customer communication. Drafting messages, following up quotes, chasing missing information, and preparing replies from notes are all screen-based, multi-step jobs. A model that can open the right record, read the thread, and draft a sensible reply saves the fiddly part of the job. The human still sends it, or at least approves it, but the time spent hunting for context drops sharply. For a business where the owner is also the salesperson, the bookkeeper and the customer service desk, that reclaimed half-hour a day is not trivial.
A third area is content and marketing support. Editing short videos, creating images, preparing listings, and updating a website are all tasks where a computer-use model can handle the repetitive mechanics: resizing, reformatting, uploading, filling in fields, and following a checklist. The creative decisions stay human. The donkey work does not have to.
There is also a subtler benefit that matters for smaller teams: consistency. A model follows the same process every time. It does not skip a field because it is in a hurry, and it does not forget step four of a five-step process on a Friday afternoon. That consistency is only valuable if the process itself is right, which is why the pilot stage matters so much, but once a workflow is proven, a model that repeats it faithfully is a genuine asset.
What to pilot first
Do not start with your most complex workflow, and do not start with anything involving payments, legal commitments, or irreversible actions. Start with one repetitive, low-risk task that you can measure. Good first pilots share three traits: the task happens often, the inputs are structured enough to check, and a mistake is easy to spot and cheap to fix.
Invoice entry is the classic first pilot. Take a batch of invoices you have already processed, run them through the new workflow with a person verifying every field, and compare the model’s output against the known-correct entries. Measure two things: how long it takes per invoice including the checking time, and how many fields needed correcting. If both numbers beat the manual process, you have a real result. If the checking time eats the saving, the workflow needs tighter instructions or a narrower scope, not more enthusiasm.
Form filling and data transfer are equally good pilots. Moving details from an enquiry form into a CRM, or from a booking email into a scheduling sheet, is measurable in the same way: time per record, errors per record, and how often a human had to intervene. Quote follow-ups are another strong candidate, provided a human approves every message before it goes out. The rule for every pilot is the same: the model drafts and prepares, the human approves and sends.
Keep the pilot small on purpose. One task, one person owning it, two to four weeks, a simple log of time saved and errors caught. That log is what turns a demo into a decision. Without it, you are relying on a feeling that things got faster, and feelings are a poor basis for changing how a business handles its money or its customers.
Cheaper automation, honestly assessed
The price cut deserves a clear-eyed reading. API prices of $1 per million input tokens and $5 per million output tokens are genuinely low by flagship standards, and for a developer building an automation, that changes what can be offered profitably. But most small businesses will not buy tokens directly. They will meet this model inside ChatGPT Work or Codex on a Plus, Pro, Enterprise or Edu plan, or inside a tool built by someone else. What matters to them is not the per-token figure, but what their plan allows, what the tool charges, and how far the allowance stretches on their actual workload. OpenAI says the allowance goes further in ChatGPT Work and Codex than with the previous models, which is encouraging, but the only allowance that counts is the one on your plan, doing your jobs.
There is also a hidden cost that price-per-token figures never show: supervision. A model that is right 95% of the time still needs a checking step, and that checking step is human time. The honest calculation is total cost per completed, verified task: model cost plus checking time plus the cost of fixing the errors that slip through. On that measure, GPT-6.1 Terra looks promising precisely because OpenAI reports both lower cost and a much lower error rate. Fewer errors mean less checking and less rework. But you should still verify that on your own tasks before you scale anything up, because your forms, your suppliers and your software will not match OpenAI’s benchmarks exactly.
Device and tooling notes
You do not need to buy new hardware for this. GPT-6.1 Terra reaches users through ChatGPT Work and Codex and through the OpenAI API, which means it runs in the environments where those products already run: the ChatGPT app and web experience, and developer-built tools. What you do need is the right access: a Plus, Pro, Enterprise or Edu plan for the ChatGPT route, or a developer setup for the API route. Check which route your team actually uses before assuming anything. A business that works entirely in ChatGPT does not need the API, and a business that wants automation wired into its own systems will need a developer to set that up properly, with permissions kept tight.
Permissions are the practical detail most small businesses overlook. A computer-use model is only as safe as the access it is given. Give it read access where reading is enough. Give it write access only to the specific app or folder the task needs. Never hand it broad admin rights, shared passwords, or access to banking and payment systems during a pilot. If a task genuinely needs to move money or submit something official, that step stays human until the workflow has earned trust over time — and even then, keep an approval step. The technology is new; the principle is old: nobody gets the keys to the whole business on their first week.
Training is light but not zero. The person running the pilot needs to know how to write a clear instruction, how to spot a plausible-looking wrong answer, and how to stop a run that is going off track. That is an afternoon of learning, not a course. The bigger shift is cultural: getting comfortable with checking machine output properly instead of skimming it. A quick skim defeats the purpose. The check has to be real, at least until the error rate on your own tasks proves it can be lighter.
Where this fits alongside your existing tools
GPT-6.1 Terra does not replace your accounting software, your CRM, or your booking system. It works across them. That is the distinction that makes computer-use models interesting for small businesses that cannot afford to rebuild their systems. Instead of buying a new platform or paying for a custom integration, a model that can read one screen and act in another can bridge tools that were never designed to talk to each other. That is also why the supervision point keeps coming back: a bridge between systems is powerful, and a bridge needs inspecting.
If you already use AI agents in some form, this launch is a chance to revisit the cost side of that setup. A cheaper, faster model for routine steps, with Sol or Astra reserved for the hard steps, is a sensible division of labour, and it mirrors how businesses already split work between junior and senior staff. Our earlier coverage of OpenAI’s Dots framework looks at a related part of OpenAI’s agent push, and it is worth reading alongside this launch if you are trying to see where the pieces fit together.
Honest limits / what to watch
Every launch announcement deserves a careful read, and this one is no exception. The strongest claims about GPT-6.1 Terra come from OpenAI itself, and they should be treated as vendor-reported results until independent testing confirms them. Benchmarks are useful, but they are run under conditions OpenAI chose, on tasks OpenAI selected, and the headline figures — a 15 percentage point gain here, a 50.3% coding score there — are not a guarantee of how the model will perform on your supplier’s invoice format or your council’s planning portal. Treat them as a reason to pilot, not as proof.
Errors still happen, and the numbers themselves say so. An error rate of 4.3% at maximum effort on OpenAI’s internal benchmark is a big improvement on 53.6% or even 16.7%, but it is not zero. In a batch of a hundred invoices, that kind of rate would still mean several entries needing correction if it transferred directly to your workload — and it may not transfer directly at all. The practical lesson is unchanged: keep a human checking step, keep permissions narrow, and keep irreversible actions human for now. A model that is usually right is more dangerous than one that is obviously wrong, because its mistakes are easier to miss.
Computer-use actions need supervision and sensible permissions by design. A model that can click and type can also click and type the wrong thing, and the more access it has, the further a wrong action can travel. This is not a reason to avoid the technology. It is a reason to roll it out the way you would onboard any new helper: limited access, clear tasks, close supervision at first, and wider trust only after a track record. If a vendor or a tool offers to let an agent loose across your systems with full access from day one, that is a red flag, not a feature.
There is also a real risk of confusing two different prices. The $1 and $5 figures are API prices for developers buying tokens. They are not ChatGPT subscription prices, and they do not mean that Plus or Pro gets cheaper, or that a tool built on the API will cost you $1 and $5 per million tokens. Your cost will be whatever your plan or your tool charges, plus the human time spent checking. Anyone quoting the API figures as if they were the price you will pay is either confused or selling something.
For the hardest work, OpenAI’s own guidance still applies: complex coding and in-depth research belong with GPT-6.1 Sol or GPT-6 Astra. Using the everyday model for a job it was not positioned for, and then blaming the model, helps nobody. The smart setup is tiered: Terra for routine computer work, the stronger models for the tasks that genuinely need them. That is also the setup that keeps costs down, because you are not paying flagship prices for invoice entry.
Finally, watch the rollout itself. Availability is described as rolling out, which means not every user on an eligible plan will see it at the same moment, and the exact experience in ChatGPT Work and Codex may evolve as OpenAI tunes it. If you cannot see it yet, that is not necessarily a problem with your account. Check your plan, check for updates, and judge the model on a real task once it reaches you rather than on launch-day headlines.
The bottom line is a balanced one. GPT-6.1 Terra looks like a genuinely useful step: faster, cheaper, and reportedly much more reliable at the routine computer work that fills small-business days. It is not magic, it is not error-free, and it is not a substitute for judgement. Piloted carefully, with a human checking the output and permissions kept tight, it is one of the more practical AI launches of the year for ordinary businesses — not because it does something spectacular, but because it does something ordinary, well, and affordably.
Key facts
| Fact | Detail |
|---|---|
| Model | GPT-6.1 Terra, successor to GPT-5.6 Terra |
| Announced | Monday 26 October 2026, via OpenAI blog post and X post |
| Position | Everyday / middle-tier model, below GPT-6.1 Sol and GPT-6 Astra |
| Core capability | Computer use: reads screens and takes action across apps |
| API pricing (reported) | Around 50% lower than GPT-5.6 Terra; $1 per million input tokens, $5 per million output tokens |
| ChatGPT availability | Rolling out to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex |
| API name | gpt-6.1-terra |
| Usage allowance | Goes further in ChatGPT Work and Codex than with GPT-5.6 Terra, Sol or Astra (OpenAI says) |
| Agents’ Last Exam | 15 percentage points higher than GPT-5.6 Terra at medium reasoning effort, at about one-third the cost (OpenAI-reported) |
| Agents’ Last Exam vs Astra | Matches GPT-6 Astra at roughly one-tenth the price per token (OpenAI-reported) |
| Speed | Finishes tasks over 3x faster on average than Astra and GPT-6.1 Sol (OpenAI-reported) |
| Internal benchmark error rate | Fell from 53.6% (GPT-5.6 Terra, low effort) to 16.7% (GPT-6.1 Terra, low effort) and 4.3% at maximum effort (OpenAI-reported) |
| Terminal-Bench 4.0 coding | 50.3% at maximum effort vs 21.5% for GPT-5.6 Terra (OpenAI-reported) |
| Other benchmarks | Strong on OSWorld 2.1 desktop automation and MMMU-Pro visual reasoning, close to Astra/Sol despite lower cost (OpenAI-reported) |
| Example tasks cited by OpenAI | Turning architectural drawings into a 3D model; shot-by-shot time-coded log of an animated film for a restoration team; reading invoices into accounting software |
| Still use Sol or Astra for | Complex coding and in-depth research |
FAQs
What is GPT-6.1 Terra?
GPT-6.1 Terra is OpenAI’s everyday, middle-tier model for computer-based work. It understands what is on screen and takes action across apps, such as reading screens, navigating websites, drafting messages and filling out forms, and it sits below GPT-6.1 Sol and GPT-6 Astra in OpenAI’s lineup.
How much cheaper is GPT-6.1 Terra than GPT-5.6 Terra?
OpenAI’s standard API prices for GPT-6.1 Terra are around 50% lower than GPT-5.6 Terra, reported as $1 per million input tokens and $5 per million output tokens. Those are API prices for developers, not ChatGPT subscription prices, so Plus, Pro, Enterprise and Edu users should check what their own plan includes.
Who can use GPT-6.1 Terra?
It is rolling out to ChatGPT Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and developers can use it through the OpenAI API as gpt-6.1-terra. OpenAI says the usage allowance goes further in ChatGPT Work and Codex than it did with GPT-5.6 Terra, Sol or Astra.
What can GPT-6.1 Terra actually do?
OpenAI describes it as a computer-use model: it reads what is on screen, clicks through applications, navigates websites, drafts messages, fills out forms, interprets technical drawings, edits video, creates images and automates repetitive tasks. Cited examples include turning architectural drawings into a 3D model and reading invoices into accounting software.
Is GPT-6.1 Terra better than GPT-6.1 Sol or GPT-6 Astra?
Not for the hardest work. OpenAI says GPT-6.1 Terra matches Astra on some agent benchmarks at roughly a tenth of the price per token and finishes tasks over three times faster on average, but for complex coding or in-depth research OpenAI still points users to GPT-6.1 Sol or GPT-6 Astra.
Should a UK small business switch to GPT-6.1 Terra straight away?
Start with a small supervised pilot, not a wholesale switch. Pick one repetitive, low-risk task such as invoice entry or form filling, run it with a person checking the output, measure the time saved and the error rate, and only widen access once the results hold up in your own tools and workflows.
Sources
- SiliconANGLE: OpenAI launches GPT-6.1 Terra for faster, better-performing AI agents at half the price
- Digit: OpenAI introduces GPT-6.1 Terra AI model, says it excels at communication and computer-based tasks
- Interesting Engineering: OpenAI GPT-6.1 Terra, 50 percent cheaper
- OpenAI Developer Community: GPT-6.1 Terra: delivering fast and capable computer-use at a lower price
