OpenAI released GPT-6.1 Terra on 26 October 2026, a model built to read screens and act inside apps. The GPT-6.1 Terra price is $1 per million input tokens and $5 per million output tokens, half what OpenAI charges for GPT-6.1 Sol. It reaches ChatGPT Work and Codex users on Plus, Pro, Enterprise and Edu plans.
What GPT-6.1 Terra Costs, per Token and per Task
The list price is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. OpenAI says that is 50 per cent below both GPT-5.6 Terra, the model Terra replaces, and GPT-6.1 Sol, its stronger sibling.
| Rate (per million tokens) | GPT-6.1 Terra | GPT-5.6 Terra |
|---|---|---|
| Input | $1.00 | $2.00 |
| Cached input | $0.05 | $0.20 |
| Output | $5.00 | $12.00 |
The GPT-5.6 Terra column comes from Artificial Analysis, the independent benchmarking firm, which published the old rates alongside the new ones. On those figures the output cut is closer to 58 per cent than to half.
Per-token rates are not the same as a bill, because a bill depends on how many tokens a job consumes. Artificial Analysis measured the cost of running its own Intelligence Index on the model at about $0.40 per task, and put the blended rate at $0.73 per million tokens on a 7:2:1 mix of cached input, fresh input and output. For agent work, OpenAI’s own figure is lower still: it says Terra beats GPT-5.6 Terra on Agents’ Last Exam at roughly one-third of the cost.
What the Model Is Built To Do
Terra is a computer-use model, meaning it is designed to look at an interface and operate it rather than only answer questions about it. OpenAI describes it as “a fast, efficient model that combines detailed visual reasoning and precise computer use”, and says it “understands what’s on screen, takes action across apps, writes naturally”.
In practice that covers filling in forms, entering invoices, reading technical drawings and moving between business systems. It accepts text and images as input, and Amazon’s own release note puts the context window at up to 1 million input tokens.
OpenAI’s Benchmark Numbers
Every figure in this section is OpenAI’s, measured on its own evaluations, and no outside party had reproduced them when this was written.
| Test | GPT-6.1 Terra | GPT-5.6 Terra |
|---|---|---|
| Agents’ Last Exam, medium effort | +15 points | baseline |
| Terminal-Bench 4.0, maximum effort | 50.3% | 21.5% |
| Factual error rate, low effort | 16.7% | 53.6% |
| Factual error rate, maximum effort | 4.3% | 18 points higher |
Agents’ Last Exam tests how well a model handles software tasks that fill an office day; one of its problems asks the model to turn architectural drawings into a 3D model. OpenAI also says Terra finishes tasks more than three times faster on average than GPT-6 Astra and GPT-6.1 Sol, and that it matches GPT-6 Astra on Agents’ Last Exam. “It even performs on par with GPT-6 Astra at only one-tenth its price per token,” the company said.
Interesting Engineering, which read the full announcement, reports two further results that other outlets did not carry: Terra beat GPT-5.6 Terra’s best OSWorld 2.1 score while costing substantially less per task, and came close to GPT-6 Astra and GPT-6.1 Sol on MMMU-Pro.
What the First Independent Test Found
Artificial Analysis scored the maximum-effort setting at 48 on its Intelligence Index, placing it 25th of 221 models against a median of 28 for comparable models. That is a strong result for the price, and it is the only independent number published so far.
The same measurements complicate OpenAI’s speed claim. Output runs at 90.9 tokens per second, 87th of 188 models and a little above the 83.4-tokens-per-second median for its price tier. But Artificial Analysis’s published measurements put time to first token at 189.33 seconds through OpenAI’s API, against a median of 3.75 seconds, which the firm calls “at the higher end”.
Those two things are not in conflict. OpenAI’s “three times faster” refers to finishing a whole task, not to answering quickly, and a model that thinks for three minutes before it types can still finish first. It does mean Terra at maximum effort is a poor fit for anything a person waits on.
One more number matters for budgets. Terra spent 130 million output tokens completing the Intelligence Index, against an 81 million median, so it is a verbose model. A halved output price on a model that writes 60 per cent more than average does not halve the invoice.
Where You Can Use It Today
As of 27 October 2026, GPT-6.1 Terra is live in four places:
- ChatGPT Work and Codex: rolling out to Plus, Pro, Enterprise and Edu accounts. ChatGPT Work is the workplace edition, the one Uber switched on for more than 30,000 staff this month. OpenAI says the usage allowance stretches further than it did with GPT-5.6 Terra, and much further than with GPT-6.1 Sol or GPT-6 Astra.
- The OpenAI API: the model identifier is gpt-6.1-terra.
- Amazon Bedrock: generally available from 26 October. AWS has not listed the supported regions in its announcement and points developers to the Bedrock documentation instead.
- GitHub Copilot: reported as generally available on 26 October by the technology site explainx.ai. Microsoft has not been quoted confirming it, so treat this one as single-sourced.
OpenAI has given no date for Free or Go tier access. The company made Terra available the same day it announced it, which is how the surfaces above came online together.
Which OpenAI Model to Pick for Which Job
OpenAI is unusually direct about Terra’s limits. For in-depth research or complex coding it tells users to choose GPT-6.1 Sol or GPT-6 Astra instead, and it repeated that advice in the launch post rather than burying it in documentation.
The practical reading is that Terra is the volume model: screen work, routine drafting, agent loops that run many steps cheaply. Reasoning effort is the dial that matters most, because the gap between settings is large. At low effort the factual error rate is 16.7 per cent, which is fine for drafting and poor for anything load-bearing; at maximum effort it falls to 4.3 per cent but the model burns roughly twice the tokens and takes minutes to start.
For comparison, Gemini 4 Argon starts at $2 per million input tokens and $10 per million output, and doubles on 1 December 2026 — which would leave Terra at a fifth of Google’s rate on input.
What Is Still Unconfirmed
Several things reported around the launch do not yet stand up twice. OpenAI’s three-times-faster claim has no independent test behind it. The model’s knowledge cutoff has been given as 1 June 2026 by a single outlet and is not in the material OpenAI published. The Bedrock region list is still unpublished.
Interesting Engineering also reports that Terra keeps GPT-6 Astra’s safeguards, resists prompt injection better than its predecessor, and sits at the Critical threshold for cybersecurity in OpenAI’s preparedness evaluations. That last point is significant, and it rests on one outlet’s reading of the announcement.
There is also a naming oddity. The announcement post in OpenAI’s own developer community is headed “GPT-6.1 Terra”, while the announcement it links to is labelled “Introducing GPT-6 Terra”. OpenAI has not explained the discrepancy, and every product surface uses the 6.1 name.
Frequently Asked Questions
How Much Does GPT-6.1 Terra Cost?
$1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens on the standard API tier. OpenAI says that is 50 per cent below GPT-5.6 Terra and GPT-6.1 Sol.
Can I Use GPT-6.1 Terra on a ChatGPT Plus Plan?
Yes, inside ChatGPT Work and Codex. Pro, Enterprise and Edu accounts get it too. OpenAI has not said whether or when Free and Go users will.
Is GPT-6.1 Terra Better Than GPT-6.1 Sol?
Not for everything. OpenAI recommends Sol or GPT-6 Astra for in-depth research and complex coding, and positions Terra for screen-based and everyday work.
How Large Is the GPT-6.1 Terra Context Window?
Up to 1 million input tokens, according to the Amazon Bedrock release note. It accepts both text and image inputs.
Is GPT-6.1 Terra Actually Faster?
It depends what you measure. OpenAI says it completes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol, but Artificial Analysis measured 189 seconds to the first token at maximum effort, so it is slow to begin answering.
