OpenAI Releases GPT-6.1 Terra, Halving Computer-Use API Prices

OpenAI has released GPT-6.1 Terra, a mid-range model built for office work and computer control, priced at $1 per million input tokens and $5 per million output tokens, half the API price of its predecessor.

OpenAI has released GPT-6.1 Terra, a mid-range model built for everyday office work and automated computer control, at half the API price of its predecessor. The model, announced on October 26, 2026, costs $1 per million input tokens and $5 per million output tokens, with cached inputs at just $0.05 per million tokens, a quarter of the previous rate.

Terra is not OpenAI’s most powerful model, and it is not trying to be. It is the company’s closest answer to Anthropic’s mid-tier Sonnet series, aimed at the high-volume work that fills a normal workday: drafting proposals, interpreting technical drawings, filling out forms, and operating software on a user’s behalf. It is available now for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, as well as through the API.

What the benchmarks say

The headline claim is about value, not raw capability. On Artificial Analysis’s Intelligence Index, GPT-6.1 Terra scores 48, between GPT-6.1 Sol at 52 and its predecessor GPT-5.6 Terra at 42. The more striking numbers concern cost per task. At maximum reasoning effort, Artificial Analysis says the model gains 6 points on the Intelligence Index over GPT-5.6 Terra while costing less than a third as much per task: $0.40 versus $1.40. At lower effort levels, the gains are larger, 7 to 13 points, placing the model on what analysts call the Pareto frontier, the point where cost and performance can no longer both improve at once.

On computer-use benchmarks, the improvement is sharper still. On Terminal-Bench 4.0 at maximum effort, GPT-6.1 Terra solved 50.3% of tasks, against 21.5% for GPT-5.6 Terra. Low-effort factual errors fell from 53.6% to 16.7%. Those are the numbers that matter for the use case OpenAI is targeting: an agent that drives a computer for hours must be both cheap enough to run continuously and reliable enough that its mistakes do not cost more than its labor saves.

OpenAI says the model combines fast visual reasoning with precise computer control and a more natural writing style. That combination is the point of the Terra line. A model for office work does not need to win every reasoning benchmark. It needs to read a screen correctly, click the right button, fill the right field, and write the result in language a colleague would accept. Terra’s pitch is that it does those ordinary things well, at a price that makes running it all day affordable.

Why price is the story

The AI industry’s frontier race usually gets measured in benchmark scores. Terra’s release is a reminder that the other race, the one over cost, may matter more for adoption. A model that costs half as much per token and a third as much per task changes what businesses can afford to automate. Tasks that were too expensive to hand to an agent at GPT-5.6 Terra prices, like continuously monitoring an inbox, triaging documents, or working through a long form queue, become plausible at Terra prices.

That is also why the comparison with Anthropic matters. The mid-tier market, the models companies actually run at scale, is where the volume is. Frontier models win headlines; mid-range models win workloads. By halving the price of its mid-range computer-use model while improving its scores, OpenAI is bidding directly for that volume, and putting pressure on rivals to answer on price rather than on capability alone.

There are caveats, as with any vendor benchmark. Artificial Analysis’s index is one yardstick, and Terminal-Bench measures a specific kind of agentic task. Real-world office software is messier than any benchmark: unusual interfaces, unexpected dialogs, and tasks that require judgment about what not to do. Terra’s error rate has improved, but 16.7% low-effort factual errors is still a number that demands human oversight for anything consequential.

Cached input pricing deserves a note of its own. At $0.05 per million tokens, the cost of keeping long documents, conversation histories and screen context in front of the model collapses. For agents that work on the same files all day, that is often where the real bill accumulates, and it is where Terra’s pricing bites hardest. A model that is cheap to prompt is one thing; a model that is cheap to keep informed is what makes always-on agents economically plausible.

What it means for users

For users, the practical change is simple. ChatGPT Work and Codex users on paid plans now have a cheaper, more capable model for computer-use tasks, and developers can build agents on it through the API at $1 and $5 per million tokens. For businesses weighing agent deployments, the cost math has shifted: the same budget now buys roughly twice the tokens, or three times the completed tasks at maximum effort, compared with the previous generation.

The release also fits a pattern in OpenAI’s 2026 lineup. The GPT-6.1 family now spans from the higher-end Sol to the mid-range Terra, with each tier priced against a different job. Terra’s job is the unglamorous one: sit at the computer, do the paperwork, and do it cheaply. If it does that job reliably, it will be the model most people actually use, even if it is never the model anyone writes about first.