01What GPT-6.1 Terra Benchmarks Tell You About the Model
GPT-6.1 Terra benchmarks are the test results OpenAI shared for its new fast, efficient model, and they show it may be one of the best models for AI agents right now. Terra can look at your screen, click, type and move between apps at half the price of the last version and over three times faster than OpenAI's top models.
Terra sits below GPT-6.1 Sol and GPT-6 Astra, which are OpenAI's big, expensive frontier models. But it was built for one job above everything else: computer use. That means it can:
- Look at what is on your screen and understand it.
- Click a button or fill out a form.
- Copy something from one app and paste it into another.
OpenAI's own example is reading incoming invoices and typing the details into software. That is boring, repetitive work that eats your whole day, and it is exactly the work you want an agent doing.
In this episode, Julian Goldie walks through what Terra is, why the numbers make it so good for agents, and how he plugs it into a real agent setup. The benchmarks matter because they tell you whether a model is cheap, fast and accurate enough to trust on its own.
02GPT-6.1 Terra Pricing and Why Cost per Task Matters
To see why the price matters, you need to know how an agent works. An agent does not answer one question and stop. It works in a loop.
It looks at the screen, thinks, clicks, looks again and thinks again. Every single step costs tokens.
So with agents, the price per task matters far more than the price per question. A model that is brilliant but expensive gets painful fast when your agent runs 50 steps, 100 times a day. Terra is built to fix that.
Here are the standard API prices:
- Input: $1 per million tokens.
- Output: $5 per million tokens.
- Cached input: just $0.05 per million tokens.
- Compared to before: half the price of GPT-5.6 Terra.
- Compared to Sol: OpenAI says it is 50% cheaper than GPT-6.1 Sol too.
Cheap cached input is a big deal for agents, because they reread the same instructions many times while they work. When every loop costs less, you can afford to let an agent run more steps and take on more jobs each day.
03Agents Last Exam: GPT-6.1 Terra Benchmark Results
The main test OpenAI used is called Agents Last Exam. It is a test of real professional tasks, such as turning architectural drawings into a 3D model. It checks whether a model can actually finish work, not just answer questions.
Here is what OpenAI reports:
- Against the old Terra: at medium effort, the new Terra scores 15 points higher at about one third of the cost.
- Against GPT-6 Astra: it performs on par with Astra, OpenAI's most powerful model, at one tenth of the price per token.
- Against Claude Opus: at max effort, OpenAI says it is over three times as token efficient as Claude Opus at a quarter of the price per token.
- Cost per task: OpenAI's estimate is about nine times lower cost per task.
Julian points out that these are OpenAI's own numbers, so you should treat them as OpenAI's claims. Even so, if you knock a little off each one, the gap is still massive.
For a business owner, the takeaway is simple. You can get close to top-model results on everyday agent work while paying a small part of the price. That changes which jobs are worth automating.
04Speed and Fewer Mistakes: Two Benchmarks That Matter Most
The second big result is speed. OpenAI says Terra finishes tasks over three times faster than GPT-6 Astra and GPT-6.1 Sol. For an agent, speed is everything.
If your agent takes 10 minutes to fill out a form, you will just do it yourself. If it takes three, you will hand it off and forget about it. Faster agents are agents you actually use.
The third result is fewer mistakes. OpenAI ran an internal test of prompts designed to trip the model up. Here are the error rates:
- Old Terra at low effort: wrong 53.6% of the time.
- New Terra at low effort: wrong 16.7% of the time.
- That is a drop of nearly 37 points.
- New Terra at max effort: wrong just 4.3% of the time, close to Sol and Astra.
This matters because an agent runs while you are not watching. One wrong number on step three gets copied into every step after it. A model that makes fewer mistakes is a model you can actually trust to run on its own.
Lower error rates mean less checking, and less checking means the agent really saves you time.
05Coding, Staying on Task and Writing More Naturally
Coding also got a lot better. There is a test called Terminal Bench 4.0 that checks how well coding agents fix real problems from the command line, such as a web page whose layout jumps around while it loads. At max effort, Terra solves 50.3% of those tasks.
The old Terra solved 21.5%, so that is more than double.
OpenAI is upfront here. For the hardest coding jobs, it still recommends Sol or Astra. Think of Terra as your everyday builder.
Terra is also better at behaving like a good agent:
- Stays on task: OpenAI says it is more likely to stay inside the job you gave it.
- Honest when blocked: it is more likely to tell you when it cannot do something, instead of saying it finished when it did not.
- Resists hidden instructions: OpenAI tested it against instructions hidden in web pages and documents that tried to pull it off task, and it held up well.
- Writes more naturally: OpenAI says its writing is clearer and more human, which helps with proposals and client email summaries.
So the agent that does the work can also write up what it did in a way that sounds like a person wrote it.
06How to Get GPT-6.1 Terra and Plug It Into Your Agents
Terra is already live, and you can start using it in a few minutes. Here is how Julian explains getting access and moving work over safely:
- 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 your own AI agents, use it in the API as GPT-6.1 Terra.
- You can also plug the CLI into Hermes or whatever agent tool you already use.
- Keep your best model for the hard stuff.
- Move your everyday tasks to Terra first, such as forms, reports and data entry.
- Watch it for a week, and if it handles them well, move more over.
OpenAI also says your ChatGPT plan goes further with Terra, because the same usage covers more tasks than it would with Sol or Astra.
You do not need to switch everything today. Start small, watch the results, and let the model earn your trust. Every workflow you build this week keeps paying you back.
07Real Agent Examples You Can Set Up This Week
Inside his agentic operating system, Julian has lots of agents doing different jobs. Some jobs are hard, like planning a build or fixing a tricky bug. But most jobs are simple and repetitive, like pulling numbers from Search Console, dropping them into a report, filling in a form or checking a page.
That simple part is where Terra shines.
So the setup is this: Terra runs the everyday work, fast and cheap. When something is really hard, the agent hands it to Sol or Astra. Here are three examples from the episode:
- SEO reporting agent: every Monday it opens Search Console, reads the charts on screen, notes what went up and down, and writes you a short report. Terra's visual reasoning nearly matches OpenAI's best models on a test called MMMU Pro, so reading charts and dashboards is right in its sweet spot.
- Lead handling agent: when a new lead comes in, it reads the form, adds the details to your CRM and drafts a personal reply for you to approve.
- Admin agent: it opens each email in your inbox, reads it, and types the details in wherever they need to go or organizes them.
None of these need the most powerful model on earth. They need one that is fast, cheap, accurate and good at using a computer.
08Old Way vs New Way: Who Should Use GPT-6.1 Terra
Julian puts the change side by side. In the old way, you did all the clicking yourself, or ran agents on the most expensive model because the cheap one made too many mistakes. That led to a few problems:
- Every task cost a lot, so you only automated a couple of things.
- Your agents were slow, so you sat there watching them all day.
- You checked every result because you did not trust them.
In the new way, Terra handles the everyday clicking and typing:
- It costs a fraction per task, so you can automate far more of your day.
- It is over three times faster, so you can hand a job off and walk away.
- It makes far fewer mistakes, so you can let it run.
- When something is really hard, it hands it up to a bigger model.
Terra is for business owners and builders who spend hours on forms, reports, data entry and inbox admin. If your agents cost too much or need too much watching, these benchmarks suggest Terra is worth testing now. The people who learn to run agents cheaply today will be well ahead when everyone else catches up.