GPT-6.1 Terra: OpenAI's fast, low-cost model for computer work, tested and compared

Picture a small architecture studio where half the day is not design at all. Someone reads incoming invoices and types them into the accounting software, someone else checks a contractor's drawing against the 3D model, and everyone writes proposals and status emails. None of it is hard. All of it takes time. That is the work OpenAI built its newest model for.

OpenAI released GPT-6.1 Terra on October 26, 2026. It describes Terra as a fast, efficient model that combines detailed visual reasoning and precise computer use with clearer, more natural writing, and calls it a significant step up from GPT-5.6 Terra. Its standard API prices are 50% lower than GPT-5.6 Terra's and GPT-6.1 Sol's, and in ChatGPT Work and Codex the same plan allowance covers more tasks than with Sol or GPT-6 Astra. ¹

Where Terra fits in OpenAI's lineup

OpenAI now sells four GPT-6 models at different prices. Astra is the flagship for the hardest work, Sol is the capable all-rounder, Luna is the cheapest, and Terra sits between Luna and Sol: priced closer to the bottom, but built specifically for seeing what is on a screen and acting on it. OpenAI says that for the most complex coding tasks, Sol and Astra remain the better choice. ¹

Cached input, meaning material the model has already seen recently, such as a long set of instructions, costs $0.05 per million tokens on Terra. For automation that sends the same instructions with every invoice or form, that discount does most of the work. ¹

What OpenAI's benchmarks show

OpenAI's headline test is Agents' Last Exam, a set of professional computer tasks such as turning architectural drawings into a 3D model or reviewing an animated film shot by shot to build a time-coded log. At medium reasoning effort Terra scores 15 points higher than GPT-5.6 Terra at about a third of the cost. At max effort it scores 60%, level with GPT-6 Astra's 59.3%, at about $1.01 per task against Astra's $7.24. ¹

Claude Opus 5.5 still scores highest on the test, at 63.2%, but OpenAI puts it at about $8.82 per task. That is where its claim of a roughly ninefold lower cost per task comes from: Terra at max effort uses about a third as many tokens as Opus, at a quarter of the price per token. OpenAI also says Terra finishes tasks more than three times faster than Astra and Sol. ¹

The pattern is clear. On visual and computer-use tests, Terra lands within a few points of Sol and Astra. On OSWorld 2.1, which tests using desktop apps for multistep tasks, it scores 74.2% at max effort against Sol's 78% at a similar cost, and OpenAI notes that at medium effort it already beats GPT-5.6 Terra's best score. On Terminal-Bench 4.0 for coding, it more than doubles GPT-5.6 Terra's 21.5% to 50.3%, but stays behind Sol and Astra. ¹

Far fewer factual errors

OpenAI says reducing made-up answers has been a priority this year, and Terra shows it. On OpenAI's internal set of prompts designed to trip models into factual errors, Terra's error rate at low reasoning effort is 16.7%, down from 53.6% for GPT-5.6 Terra. At max effort it falls to 4.3%, close to Astra's 4.2% and Sol's 4.4%. ¹

Independent testing points the same way. On Artificial Analysis's AA-Omniscience test, which penalizes confident wrong answers, Terra's hallucination rate is 34.5%, lower than Sol's 51.9%, Astra's 49.8% and Claude Opus 5.5's 50.3%. Gemini 4 Argon remains far ahead at 4.3%. ³

What independent testing says

Artificial Analysis moved to version 5.0 of its Intelligence Index this week, a new mix of nine evaluations, so its scores are not comparable with the version 4.3 figures in our earlier model articles. On version 5.0, Terra at max effort scores 51, behind Sol at 54.5 and Astra at 56.9, and level with Gemini 4 Argon at 50.4. Its strength is cost: $0.96 per index task, less than Sol's $1.27 and a fraction of Astra's $6.25 or Claude's roughly $10. ³

One claim the independent numbers do not echo is speed. OpenAI says Terra completes Agents' Last Exam tasks more than three times faster than Astra and Sol. Artificial Analysis measures Terra's output at 91 tokens per second, a little faster than Sol and Astra, and its average time per index task at about 11 minutes, roughly the same as Astra and about a quarter quicker than Sol. Faster, yes; three times faster on general work, not in this data. ¹³

Terra against Claude and Gemini

Two comparisons matter for most businesses. Against Anthropic's Claude Sonnet 5.5, OpenAI's own chart data shows Sonnet ahead on computer use, 83.9% to 74.2% on OSWorld 2.1, and well ahead on coding, 70.6% to 50.3% on Terminal-Bench 4.0, but at several times the cost per task: about $5.74 and $10.44 against Terra's $1.19 and $1.84. Against Claude Haiku 5.5, which costs a tenth of Terra per token, Terra scores much higher on independent testing, 51 to 36. ¹³

Gemini 4 Argon scores about the same as Terra on Artificial Analysis's index and hallucinates far less, but costs about four times as much per index task. For work where a wrong fact is costly, such as legal or financial research, Argon still has the edge. For high-volume computer work where speed and cost matter most, Terra is the better buy. ³

What a business can use it for

OpenAI's own examples are refreshingly ordinary: drafting proposals, interpreting technical drawings, filling out forms, editing videos and reading incoming invoices into accounting software. Here is how those map to small business work. ¹

Safety, availability and access

OpenAI says Terra meets the Critical threshold for cybersecurity under its Preparedness Framework and carries the same protections developed for GPT-6 Astra. In OpenAI's evaluations it is more likely than GPT-5.6 Terra to stay within the scope of a task and to tell the user honestly about limits or blockers, and it showed strong resistance to prompt injection, meaning hidden instructions in web pages or documents that try to redirect it. ¹²

Terra is available now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and to developers through the OpenAI API as gpt-6.1-terra. Because it uses less of your plan's allowance per task, OpenAI says the same plan covers more work with Terra than with Sol or Astra. ¹

How to test it on your own work

• Pick one repetitive screen task, such as entering ten real invoices or filling one form you complete every week.
• Run it on your current model and on Terra at medium effort, with the same instructions.
• Count errors and corrections, not just speed.
• Try max effort only on the cases medium gets wrong.
• Keep a person approving anything that pays, submits or signs.