GPT-6.1 Terra at $1/$5: route back-office agents to it

OpenAI shipped GPT-6.1 Terra on Monday, and it is the model most small businesses should actually be testing. GPT-6.1 Terra lists at $1 per million input tokens and $5 per million output, half of GPT-6.1 Sol's $2/$10. OpenAI pitches it at routine computer work: invoices into accounting software, forms, proposals. That is the boring queue every operator wants off their desk.

What actually happened

OpenAI introduced GPT-6.1 Terra on October 26 as the replacement for GPT-5.6 Terra. It sits below Sol and Astra in the lineup and above Luna. Per The Next Web, it is live for ChatGPT Plus, Pro, Enterprise and Edu users and in Codex, and it is in the API as gpt-6.1-terra.

The claims OpenAI leads with, none of them independently checked yet:

• Computer use: reads the screen, then clicks, types and switches between apps. OpenAI's examples include processing incoming invoices into accounting software and filling in forms.
• Cost vs. the old model: 15 percentage points above GPT-5.6 Terra on one of OpenAI's tests at about one-third of the cost, and level with GPT-6 Astra on that test at about one-tenth of the price.
• Factual errors: on a set of trick questions at the lowest setting, the error rate fell from 53.6% to 16.7%.
• Coding: about half of the tasks on a command-line test, more than double its predecessor. OpenAI still tells you to use Sol or Astra for the hardest coding.

OpenAI also rates Terra "Critical" for cybersecurity capability under its Preparedness Framework, with the same protections as GPT-6 Astra.

Why GPT-6.1 Terra matters for your business

Most of your agent work is not hard. We look at agent logs for clients every week. The bulk of the calls are "read this PDF, put these six fields in that system." Paying Sol or Astra rates for that is a habit, not a decision. A model built for that queue at $1/$5 changes the math on automating it.

Cheap per token is not cheap per task. Computer-use runs burn tokens on screenshots, retries and tool calls. The headline rate covers text tokens. Measure cost per completed invoice or form, plus the human minutes spent fixing the misses. A 16.7% error rate on trick questions is a lot better than 53.6%. It is still not zero, so keep a review step on anything that moves money.

Route by task, not by vendor tier. Pull 50 real tasks from your back-office queue, run them through Terra and through what you use now, and keep the winner per task type. Our Sol post made the same point a month ago. The models keep moving, so the eval set is the asset.

Sources: OpenAI, The Next Web.