OpenAI released GPT-6.1 Terra on October 26. The model pairs visual reasoning with computer use, and it is built to be one thing above all: cheap enough to automate work in volume. Standard API pricing is one dollar per million input tokens, five cents per million cached input tokens and five dollars per million output tokens — 50 percent below GPT-5.6 Terra and GPT-6.1 Sol.
Terra is available now to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and through the API as gpt-6.1-terra.
The numbers that concern Anthropic
OpenAI compares Terra’s computer use directly against Claude Opus 5.5. On the Agents’ Last Exam benchmark at maximum reasoning effort, OpenAI reports Terra is more than three times as token-efficient as Opus 5.5 at a quarter of the price per token, which puts cost per task at roughly one ninth. At medium effort, Terra scores 15 points above GPT-5.6 Terra at about a third of the cost. It is also fast: more than three times faster than GPT-6 Astra and GPT-6.1 Sol.
On Terminal-Bench-Science 0.1, Terra lands within five points of Opus 5.5 at less than a fifth of the estimated cost. For the hardest coding tasks, OpenAI says GPT-6.1 Sol and GPT-6 Astra remain the better pick.
Fewer wrong answers
On an internal benchmark of prompts designed to provoke mistakes, the error rate at low reasoning effort fell from 53.6 percent with GPT-5.6 Terra to 16.7 percent. At maximum effort it sits at 4.3 percent, 18 points below the predecessor.
In cybersecurity, Terra crosses the “Critical” threshold of OpenAI’s Preparedness Framework and therefore keeps the safeguards that were built for GPT-6 Astra. On prompt injection, OpenAI’s own tests report strong resistance.
The price is the argument, not the benchmark
Terra is not OpenAI’s strongest model, and OpenAI says so. It is the model that drags computer use into the range where reading invoices and filling out forms actually pays — at a dollar per million input tokens. Anyone running agents on Opus 5.5 today and counting cost per task now has a number on the table they have to argue against. That the comparison comes from OpenAI itself does not make it wrong, but it only becomes checkable once somebody else runs the same benchmarks.