OpenAI released GPT-6.1 Terra on October 26, positioned as a model for getting everyday work done at lower cost and higher speed. Using official figures and independent benchmarks, this article covers three things: what it costs, how fast it runs, and how often it gets facts wrong.
Before you start a new project, do you check how much of your AI quota is left? If so, a new model sold on efficiency is worth a few minutes of your time, if only to see what its maker actually published.
OpenAI released GPT-6.1 Terra on October 26, 2026, replacing GPT-5.6 Terra. OpenAI positions it as a model for getting everyday work done at lower cost and higher speed.
That positioning covers three things: cost, speed and accuracy. We take them one at a time: the key point first, then a plain-language explanation, then the evidence. At the end there is a summary table and four questions people commonly ask when evaluating a model.
Price: $1 in, $5 out per million tokens
Key point
The standard API price for GPT-6.1 Terra is $1 per million input tokens and $5 per million output tokens. OpenAI says this is 50% lower than both GPT-5.6 Terra and GPT-6.1 Sol.
In plain terms
A token is roughly three quarters of an English word. Everything you send to the model and every answer it returns is billed in tokens. A typical task reads far more than it writes, so both the input price and the output price matter. Take one common job size: give the model 1 million tokens of material and have it write 100,000 tokens back.
Same job: 1M tokens in, 100K tokens out
The evidence
For subscribers, OpenAI says that the same usage allowance in ChatGPT Work y Codex covers more tasks with Terra than with GPT-5.6 Terra, and substantially more than with Sol or Astra. It has not published an exact multiple. Sources: OpenAI launch post, OpenAI API model page.
Speed: OpenAI compared it with its own larger models
Key point
OpenAI says Terra completes tasks more than three times faster than GPT-6 Astra y GPT-6.1 Sol, and that it performs on par with Astra in OpenAI’s tests at one tenth of Astra’s price per token.
In plain terms
The comparison here is with the same company’s larger models, which are built to think longer and naturally take more time. Speed also depends on the reasoning effort setting: Terra offers six levels, from none to max. The higher the level, the longer the model thinks before it answers.
The evidence
Independent benchmarking firm Artificial Analysis measured the following at Terra’s highest reasoning setting.
Measured speed of GPT-6.1 Terra
Accuracy: fewer factual errors than the previous generation
Key point
On OpenAI’s internal factual-error test, GPT-6.1 Terra‘s error rate at low reasoning effort fell from 53.6% to 16.7%. At maximum effort it fell to 4.3%, close to the larger Astra and Sol models.
In plain terms
A “factual error” here means the model gave a confident answer that was wrong. These are the errors most likely to slip into a report, because nobody checks every line. The test is made of questions chosen to trip models up, so the absolute numbers look high. What is worth reading is the direction and the size of the drop.
The evidence
Factual error rate at low reasoning effort (lower is better)
Capabilities: it can read a screen and operate software
Key point
Besides text, Terra accepts images and can understand what is on a computer screen. On platforms that provide the right tools, it can click, type and switch between apps to complete multi-step work.
In plain terms
Before, you could ask a model how a product page should be designed. Now you can hand it the product photos and specs and ask it to try building a page that opens, then check the result yourself. Going from giving advice to doing the work is what “computer use” means here.
The evidence
OpenAI’s published specs show what it can take in and what it can produce. Note the limits: no audio or video input, and text-only output.
GPT-6.1 Terra at a glance
The published numbers in one place
Key point
The table below puts together the list prices of two models and the figures one independent benchmarking firm measured for both with the same method.
Published numbers for GPT-6.1 Terra y Claude Opus 5.5
| Métrico | GPT-6.1 Terra Max effort | Claude Opus 5.5 xhigh effort |
|---|---|---|
| Price per 1M input tokens | $1.00 | $4.00 |
| Price per 1M output tokens | $5.00 | $20.00 |
| Intelligence Index | 51 | 57 |
| Cost to run the full index | $972 | $4,766 |
| Output speed (tokens/s) | 78 | 107 |
| Time to first token (s) | 141.07 | 77.81 |
| Context window | 1.05M tokens | 1M tokens |
Reading the table
- Prices are per token. The total bill also depends on how many tokens a model uses to finish the job.
- The Intelligence Index summarizes many tests. It reflects average performance, not how a model handles your specific task.
- Speed and time to first token vary with reasoning setting, provider load and time of day.
For how we read another recent model launch, see our DeepSeek V4 review. For the image and video models we cover, see our Guías de modelos de IA.
Four things to weigh when evaluating a model
1 · How often does the task run?
- The more runs, the more a difference in unit price adds up
- For occasional tasks, price may not be the main factor
2 · What does one error cost?
- Is a person or script checking the output?
- Are errors easy to spot, or do they slip quietly downstream?
3 · How much reasoning does it need?
- Reasoning effort is adjustable; higher settings mean longer waits
- Everyday sorting and rewriting need less than hard multi-step problems
4 · What form are the inputs and outputs?
- Terra takes text and images and writes text
- Audio or video work needs extra tools
The most direct way to tell whether a model fits your work is to try it on a real task you already know the answer to.
Quick answers
When was GPT-6.1 Terra released?
October 26, 2026. It is available in ChatGPT Work y Codex on the Plus, Pro, Enterprise and Edu plans. The API model name is gpt-6.1-terra.
How should I read the API price?
You pay per token, with separate prices for input and output. A project’s total cost depends on how much you give the model, how much you ask it to write, and how many tokens it actually uses to finish.
Are price and answer quality the same thing?
No. Price and quality are separate dimensions, and each needs its own test results.
Can Terra work with video or audio?
No. Input is text and images only, and output is text only. For video work, extract key frames first or use tools provided by the platform.
Can I use Terra on IMA Studio?
As of publication, IMA Studio does not offer GPT-6.1 Terra.
Sources, all retrieved November 4, 2026: OpenAI: Introducing GPT-6.1 Terra · OpenAI API model page · Anthropic pricing page · Artificial Analysis comparison. Figures labeled “OpenAI internal test” are the vendor’s results on its own question set.
Acerca del autor
