OpenAI released GPT-6.1 Terra on 26 October 2026 at $1 per million input tokens and $5 per million output tokens, according to BleepingComputer's report on the launch. OpenAI says that is 50% lower than GPT-5.6 Terra. If you run a product on the OpenAI API, re-price your workloads this week. Then keep every review step you already have. A cheaper model changes the bill, not the review.
What is GPT-6.1 Terra?
GPT-6.1 Terra is OpenAI's mid-tier model, built to be fast and cheap enough for everyday agent work, including computer use. BleepingComputer quotes OpenAI describing a model that "combines detailed visual reasoning and precise computer use alongside clearer, more natural writing." The examples are ordinary office work: filling in forms, editing videos, automating repetitive desktop tasks.
It is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and through the API as gpt-6.1-terra.
What does OpenAI claim it does better?
OpenAI's claims, as reported by BleepingComputer, are about speed, coding and errors.
- Speed. OpenAI says Terra completes tasks "over three times faster" than both GPT-6 Astra and GPT-6.1 Sol.
- Coding. Terra scored 50.3% on Terminal-Bench 4.0 at maximum reasoning effort, against 21.5% for GPT-5.6 Terra.
- Errors. On OpenAI's internal test, the error rate at low reasoning effort fell from 53.6% for GPT-5.6 Terra to 16.7%.
Two things are worth noticing. First, these are OpenAI's own numbers, mostly on OpenAI's own tests. Second, OpenAI itself still recommends GPT-6.1 Sol or GPT-6 Astra for the most complex coding work. Terra is the cheap, quick tier. OpenAI is not pretending otherwise, and neither should we.
If you need the expensive tier, we compared it earlier this month in Claude Opus 5.5 vs Grok 4.7 vs GPT-6 Astra, and I will not redo that here.
Where does a cheaper model actually save money?
In the work you run often, not the work you run once.
A price cut on a model you call twice a day changes nothing. A price cut on a model that classifies every support ticket, reads every invoice or drives an agent through a hundred screens a day changes the budget. That is where Terra is aimed. Long agent runs with computer use read a lot of screenshots and write a lot of steps. Half the token price, on that kind of volume, is real money.
The way to find out is boring and reliable. Take a week of logs. For each workload, count the input and output tokens per task, multiply by the new price, and compare with what you pay today. Do it per task, not per month, because a monthly total hides which feature is expensive. You will usually find that two or three workloads account for most of the spend. Those are the ones worth testing on Terra first.
There is a catch. Agents that are cheaper to run tend to be run more. Teams add a retry, then a second pass, then a check by another model. The unit price halves and the monthly bill stays flat. Watch the total, not the rate card.
Why does a cheaper model not mean less review?
Because the risk sits in what the model does, not in what it costs.
A model that fills in forms and clicks through desktop apps can submit the wrong form and click the wrong button. That mistake costs the same whether the tokens were cheap or expensive. If anything, a cheap model invites more autonomy. "It's only a few cents a run, let it go." That is exactly when someone stops looking.
Look at the error figure again. OpenAI says Terra's error rate at low effort fell to 16.7% on its internal test. That is a big improvement. It still means a meaningful share of answers on that test were wrong. On a computer-use task that writes to a real system, one wrong action can be enough.
What review should stay the same?
- Human approval for actions that write. Reading a screen is cheap to get wrong. Submitting a payment, sending an email or deleting a record is not. Keep a person on those steps.
- The same evaluation set. When you switch models, run the same test cases you used for the last one. A price change is not a reason to lower the bar.
- Spot checks on output. Pick a sample of runs every week and read them end to end. Automated scores miss the strange failures.
- Logging. Keep a record of what the agent saw and what it did. When something goes wrong, you will need it.
- A clear rollback. If the new model behaves worse in production, you should be able to switch back in minutes, not days.
Should you switch?
Probably, for the right workloads. If you use GPT-5.6 Terra today for high-volume, low-risk tasks, test Terra on your own evaluation set and move if it holds up. If you use a top-tier model for something that does not need it, Terra is worth a trial there too.
What I would not do is switch on price alone. Run the comparison. Read the failures. Check that the new model does not make a new kind of mistake that your old checks were never built to catch. A faster model that fails in a new way can cost more than it saves.
What would I do this week?
List every place your product calls an OpenAI model. Mark which ones are high volume and which ones write to a real system. The high-volume, read-only ones are the first candidates for Terra. The ones that write keep their review, whatever model sits behind them.
Where Alongside fits
Picking a model, running the evaluation on your own tasks, and setting cost controls is the kind of work covered by Alongside's technical advisory.