OpenAI ships GPT-6.1 Terra, a cheaper middle model it says matches Astra on agent work at a tenth of the price

OpenAI released GPT-6.1 Terra on Monday, Oct. 26, 2026, for ChatGPT Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.1-terra.

The flagship gets the headlines, but the middle model is where the money moves. If OpenAI's own charts hold up, Terra does Astra-class agent work for a tenth of the per-token price, which turns "too expensive to run all day" into "fine, leave it running." The hallucination drop is the number to watch: going from wrong more than half the time on OpenAI's own test to wrong about one time in six is real progress and still not a figure you would sign a contract on. And the quiet line matters most: a mid-priced, widely available model now clears OpenAI's own top cybersecurity risk tier. Cheap, fast and capable at hacking is a combination worth more than one paragraph. All benchmarks here are OpenAI grading OpenAI; independent tests will tell.

OpenAI released GPT-6.1 Terra on Oct. 26, 2026. It is available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.1-terra. ChatGPT Work is the work product inside ChatGPT. Codex is OpenAI's coding product. The API is the paid hook developers use to call the model from their own software. Those lines are OpenAI's, from the launch post.

API pricing is $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. A token is a small chunk of text the model reads or writes. A million tokens is roughly a long book. Input is what you send. Output is what the model sends back. Cached input is text the service has already seen, so it charges less to send it again. $0.05 is one-twentieth of the $1 regular input price. OpenAI says the standard price is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. Half price, on OpenAI's comparison. Those prices are OpenAI's.

On OpenAI's Agents' Last Exam benchmark, OpenAI says GPT-6.1 Terra at medium effort scores 15 percentage points higher than GPT-5.6 Terra at about a third of the cost, and performs on par with GPT-6 Astra at one-tenth the price per token. A percentage point is one point on a 100-point scale. Agents' Last Exam is OpenAI's own test of professional tasks an agent has to finish, such as turning a drawing into a model. Medium effort is a setting for how long the model thinks. On par means about the same score. One-tenth the price means about 10 cents of Terra token cost for each dollar of Astra token cost, on OpenAI's comparison. Those figures are OpenAI's.

At maximum effort, OpenAI says Terra is more than three times as token-efficient as Anthropic's Claude Opus 5.5 at a quarter of the price per token, for roughly 9 times lower cost per task, and more than three times faster than Astra and Sol. Token-efficient means it spends fewer tokens to finish the same job. A quarter of the price is 25 cents per dollar of Opus's per-token price. Nine times lower cost per task means the finished job costs about one-ninth as much. Those comparisons are OpenAI's.

OpenAI says the hallucination error rate fell from 53.6% to 16.7% at low effort, and to 4.3% at maximum effort. A hallucination, here, is the model stating something that is not true. The rate is from OpenAI's own test of prompts that tend to draw those mistakes. 53.6% is wrong more than half the time. 16.7% is about one time in six. 4.3% is about one time in twenty-three. Those figures are OpenAI's. The launch post calls this an error rate on factual mistakes. It is not an outside lab's score.

On Terminal-Bench 4.0, OpenAI reports 50.3% for GPT-6.1 Terra versus 21.5% for GPT-5.6 Terra. Terminal-Bench is a test of whether a coding agent can solve problems from the command line, the text window programmers use to run a computer. 50.3% is a little more than half of those tasks. 21.5% is about one in five. Those scores are OpenAI's.

OpenAI says GPT-6.1 Terra meets its Critical cybersecurity threshold under its Preparedness Framework. Critical is the top tier in that framework, OpenAI's own rating of how much damage a model could do if someone used it to break into computer systems. That rating is OpenAI's.

The picture is OpenAI's own chart from the launch page, Agents' Last Exam scores versus cost per task. GPT-6.1 Terra's line sits at the low-cost end, above GPT-5.6 Terra, GPT-6.1 Sol, and GPT-6 Astra. The credit is OpenAI.

In plain terms, OpenAI on Monday shipped a middle model, GPT-6.1 Terra, to paid ChatGPT plans, Codex, and the API. The API price is $1 per million input tokens and $5 per million output tokens, which OpenAI says is half of GPT-5.6 Terra and GPT-6.1 Sol. OpenAI says Terra matches Astra on its agent test at about a tenth of the per-token price, cuts its own factual-error rate from 53.6% to 16.7% at low effort, and clears the company's top cybersecurity tier. Every score in this story is OpenAI grading OpenAI.