The computer that operates itself just got cheaper
OpenAI confirmed on Monday, in an official post on its own X account and in a blog announcement, the launch of GPT-6.1 Terra. It's a mid-sized model built for the most repetitive part of work: reading what's on screen, acting inside apps, writing natural text and finishing whole tasks without a person clicking through every step. The company calls this computer use, the ability for the model to see a graphical interface the way a human would and operate the mouse and keyboard on its own.
Why price is the real story
OpenAI's headline isn't just capability, it's cost. According to the official announcement, GPT-6.1 Terra's standard API price dropped 50% compared to its predecessor GPT-5.6 Terra and also to GPT-6.1 Sol: $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. On the internal Agents' Last Exam benchmark, which measures professional tasks like turning an architectural drawing into a 3D model or reviewing a film shot by shot, OpenAI says Terra at max effort is over three times as token-efficient as Claude Opus 5.5 and costs one-fourth the price per token, resulting in roughly a 9x lower cost per task. The company also claims the model completes tasks over three times faster than both GPT-6 Astra and GPT-6.1 Sol.
Where it's already running
Availability is real, not a future promise. GPT-6.1 Terra is live in ChatGPT Work and Codex for all Plus, Pro, Business, Enterprise and Edu plans, and can also be accessed via the API under the identifier gpt-6.1-terra. OpenAI's public changelog, which documents real-time Codex changes, already lists the model as available starting today. OpenAI president Greg Brockman separately posted on X that computer use with Terra crossed a qualitative threshold, describing the experience as faster than what he could do unassisted, reinforcing that the company's internal bet is that the model already changed how teams use their own tooling day to day.
What improved beyond cost and speed
OpenAI also reports reliability gains. On its internal benchmark of error-inducing prompts, the failure rate at low reasoning effort dropped from 53.6% with GPT-5.6 Terra to 16.7% with GPT-6.1 Terra. At maximum effort, the rate fell to 4.3%, close to the numbers for GPT-6 Astra and GPT-6.1 Sol. On Terminal-Bench 4.0, which tests coding agents solving command-line problems, Terra went from 21.5% to 50.3% of tasks solved. For the most complex coding tasks, OpenAI itself recommends sticking with Sol or Astra, making clear Terra is born as a mid-tier model, not a flagship.
The limit nobody should ignore
A model that sees the screen and acts on its own inside apps widens the risk surface for misconfigured automations, and OpenAI itself acknowledges this: GPT-6.1 Terra meets the Critical threshold for cybersecurity under the company's Preparedness Framework, its internal capability risk classification, and inherits the same safeguards applied to GPT-6 Astra. The company also says it strengthened the model's resistance to malicious instructions hidden in third-party content, known as prompt injection, precisely because an agent operating a computer on its own is a natural target for that kind of attack. For anyone automating tasks with Terra, the practical takeaway is to configure permissions carefully and review exactly what the agent is allowed to touch.
Sources
OpenAI: Introducing GPT-6.1 Terra; OpenAI on X: official GPT-6.1 Terra announcement; Greg Brockman on X: computer use with Terra; OpenAI: Codex/ChatGPT changelog, October 26-30, 2026.