01What Is GPT-6.1 Terra and How Does It Work?
What is GPT-6.1 Terra? It is OpenAI's new fast and low-cost model that can actually use your computer for you. It is not the smartest model OpenAI makes, but it is fast, cheap and accurate enough to make AI agents worth using for normal, everyday work.
OpenAI now has a family of models. At the top sit GPT-6.1 Sol and GPT-6 Astra. Julian Goldie calls them the senior experts.
They are brilliant, but slow and expensive, so you save them for your hardest problems.
Terra is the fast, skilled worker. Its job is not to solve the hardest problem in the room. Its job is to get through the pile of everyday work quickly and cheaply without making a mess.
Here is what it can do on a screen:
- Look at your screen and understand what it sees.
- Click buttons and type into boxes.
- Scroll and open the next app.
- Keep going until the job is done.
In this episode, Julian explains in plain English what Terra is, why it is useful, and the five big things it changes about how AI agents work for everyone.
02GPT-6.1 Terra Price, Speed and Test Results
Most of the work in a business is not hard. It is just boring. Think of logging into a supplier portal to check stock, copying new bookings into a spreadsheet, or setting up a new client in four different tools.
None of that needs a genius. It needs someone fast, cheap and careful, and that is the gap Terra fills.
OpenAI's numbers back this up:
- Price: $1 per million tokens in and $5 per million tokens out.
- Savings: that is 50% lower than both the old Terra and GPT-6.1 Sol.
- Cached input: just $0.05 per million tokens.
- Speed: over three times faster than Sol and Astra.
- Agents Last Exam: on this test of real professional agent tasks, OpenAI says Terra performs on par with Astra for a fraction of the price per token.
- Gain over the old model: the new Terra scores 15 points higher than the old one on Agents Last Exam at about a third of the cost.
Put simply, you get results close to the top models on everyday work for a small part of the cost. That is why Julian sees Terra as the model that makes agents make sense for normal jobs.
03How to Start Using GPT-6.1 Terra Today
Getting started is simple. OpenAI has put Terra inside ChatGPT for paying users and in the API for people who build their own agents. Here is the path Julian lays out in this episode:
- If you are on ChatGPT Plus, Pro, Enterprise or Edu, open ChatGPT Work or Codex.
- Pick GPT-6.1 Terra from the model list.
- If you build agents yourself, use GPT-6.1 Terra in the API instead.
- Write down every task you did on your computer this week that felt boring.
- Pick the one task you hate the most.
- Hand that one task to Terra and watch how it does.
- When you trust it, hand over the next task.
That is basically it. One boring task at a time. You do not need to rebuild your whole business on day one.
You just need to clear one job, check the result, and then move on to the next.
Julian's point is that the people who start handing work to agents now, while models like Terra keep getting faster and cheaper, will be miles ahead of the people who wait.
04Shifts 1 and 2: Automate the Boring Work in Any App
The first big shift is that automation goes from a few tasks to most of the boring work. In the past, you would set up an agent on a top model, and every run cost real money. Agents work in a loop.
They think, click, loop again and think again, and every step costs tokens. So you would automate two or three things that were really worth it and leave everything else manual.
With Terra, the cost per task drops a lot. Cached input at $0.05 per million tokens matters because agents reread the same instructions over and over. Now the maths works for the small stuff too:
- The 10-minute jobs you do every day.
- The repeat tasks that add up to hours each week.
- The small admin you used to skip automating.
The second shift is that any app with a screen can now be used by an agent. Before computer use, connecting an agent to a tool usually meant some kind of integration. If your software was old, you were stuck.
Terra uses the app the same way you do, by looking at the screen and clicking.
So an ancient booking system, a supplier portal from 2009 or an old internal tool can all be part of the workflow. OpenAI tested this on OS World 2.1, a benchmark about finishing multi-step tasks in real desktop apps. At medium effort, the new Terra beats the old Terra's best score.
05Shift 3: Fewer Mistakes Means Less Babysitting
The third shift is a big one. You stop babysitting your agents. The main reason most people do not trust agents is mistakes.
You let an agent run, it gets one thing wrong on step three, and that mistake gets copied through everything after it. So you end up checking every result, and at that point you may as well do it yourself.
OpenAI ran an internal test full of prompts designed to trip models up. Here is how the error rates compare:
- Old Terra at low effort: wrong 53.6% of the time.
- New Terra at low effort: wrong 16.7% of the time.
- New Terra at max effort: wrong just 4.3% of the time.
Terra is also better behaved as an AI agent. OpenAI says it is more likely to stay inside the job you gave it. It is also more likely to tell you honestly when it is blocked, instead of pretending it finished.
On top of that, it held up well against hidden instructions inside web pages that tried to pull it off task. That is exactly what you need from an agent you are not watching. Fewer errors and more honest reporting mean you can let it run and check less often.
06Shifts 4 and 5: A Model Team and a Plan That Goes Further
The fourth shift is that you stop using one model for everything. For a long time, the default was to use the best model you could afford for every job. Terra changes that.
The smart setup now is a team with a router:
- Terra does the everyday work.
- When a job is really hard, it gets handed to Sol or Astra.
- You get speed on most of the work and the big brain only when you need it.
Julian says this is how he thinks about it inside his own agent system. Most tasks are simple and repetitive, so they go to the fast, cheaper worker. The few hard ones go to the expert.
OpenAI is clear about this too. Terra's coding got much better on Terminal Bench 4.0, going from 21.5% of tasks solved to 50.3% at max effort. But for the hardest coding jobs, OpenAI still recommends Sol or Astra.
The fifth shift is that your plan goes further. If you are on ChatGPT, OpenAI says your usage goes further with Terra than with Sol or Astra. Speed matters here too.
If an agent takes 10 minutes to do a job, you will just do it yourself. When it is much faster, you hand it off. So you get more tasks you can actually give away.
07What GPT-6.1 Terra Does Not Change
Julian is straight about what Terra does not change. It is not the smartest or most advanced model OpenAI makes. For the really hard thinking, Sol and Astra are still the ones to use.
Keep that in mind so you do not expect expert-level results on your toughest problems.
Here are the limits and habits to keep:
- Hard problems still go up a level. Send the toughest thinking and coding to Sol or Astra.
- Set up your agents properly. A good model still needs a good setup.
- Write clear instructions. Tell the agent exactly what the job is and when it is done.
- Check anything important. For things like money going out, do a quick check before it is final.
A common mistake is to hand Terra everything at once. A better approach is to move one boring task over, watch it, and only then add the next. That way you learn where it shines and where it needs help.
For the everyday pile of work, though, the trade-off has completely changed. Low cost, high speed and far fewer mistakes mean it is now worth automating jobs that were never worth it before.
08Old Way vs New Way: Who GPT-6.1 Terra Is For
Julian puts the change side by side. The old way looked like this:
- You did the boring computer work yourself, or paid top model prices for every little task.
- You only automated a couple of things.
- Tools without an integration stayed manual.
- You checked every result because the agent made too many mistakes.
- You sat there waiting because it was slow.
The new way looks very different:
- Terra does the everyday clicking and typing.
- The cost per task is low enough to automate the small stuff.
- Any app with a screen can be part of the workflow.
- It makes far fewer mistakes and tells you when it is stuck.
- It is over three times faster, so you can hand work off and walk away.
- Hard problems go to the big models.
Terra is for anyone who spends hours on portals, bookings, weekly admin and spreadsheets. If you run a business and most of your computer time is boring rather than hard, this model is built for you. Pick the most boring task on your list and let Terra do it this week.