OpenAI's GPT-6.1 Terra Wants to Be the Model That Actually Gets Work Done

OpenAI has a growing family of models, and it is getting harder to tell them apart. In September, the company released GPT-6 Sol, a frontier model aimed at the hardest reasoning tasks. Then came GPT-6 Luna, a lighter version for everyday chat. Now, on October 26, OpenAI is filling the middle with GPT-6.1 Terra, a model designed for one thing in particular: getting actual computer work done, at half the price of its predecessors.

The pitch is straightforward. Terra is built for the kind of routine tasks that fill most office days. It can read your screen, click through applications, fill out forms, enter invoices, and even edit videos. It is not trying to be the smartest model in the room. It is trying to be the most practical one.

What You Actually Get

Terra sits in an interesting spot in OpenAI's lineup. Below it is Luna, the lightweight chat model. Above it are Sol and Astra, the heavy hitters for complex reasoning and research. Terra's job is the messy middle: agentic computer use.

The benchmark numbers tell part of the story. On OpenAI's own "Agents' Last Exam" test, Terra scores 15 percentage points higher than GPT-5.6 Terra in medium reasoning mode. On the more demanding Terminal-Bench 4.0 benchmark, it solves 50.3% of tasks in maximum reasoning mode, compared to 21.5% for the previous version. That is a 134% improvement.

Perhaps more importantly for real-world use, Terra makes fewer factual errors. In low reasoning mode, its error rate drops to 16.7%, down from 53.6% in the older model. With maximum reasoning turned up, that error rate falls to just 4.3%.

Speed is another selling point. OpenAI says Terra is three times faster than both GPT-6 Astra and GPT-6.1 Sol. For tasks that involve clicking through interfaces, filling forms, and processing documents, that speed difference matters.

The context window is generous, too. Terra can handle up to 1.05 million tokens in a single conversation, with a maximum output of 128,000 tokens. That is enough to dump a small codebase, a long contract, or a sprawling spreadsheet into the model and ask it to actually do something with the information.

The Pricing Story

Here is where things get interesting. OpenAI is pricing Terra's standard API access at $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. The company claims this is 50% cheaper than both GPT-5.6 Terra and GPT-6.1 Sol.

That pricing move is not happening in a vacuum. Over the past few months, the cost of frontier LLM APIs has been in freefall. Open-weight models from Meta, Alibaba, and others have forced a race to the bottom. OpenAI's response has been to make its own models cheaper and more capable at the same time.

Whether Terra is actually half the price depends on how you use it. Cached input pricing is nearly free, which rewards workflows where the model references the same documents repeatedly. For output-heavy tasks like generating reports or long-form content, the $5 per million output tokens still adds up. But for screen-reading agents that mostly observe and occasionally click, Terra is priced aggressively.

What "Computer Use" Actually Means

Terra's headline feature is computer use, which OpenAI describes as the ability to "read screens, click through apps, fill forms, enter invoices, and edit videos." This is agentic AI in its most literal form: a model that can see what is on your screen and interact with it.

The concept is not new. Anthropic has been pushing similar capabilities with Claude's computer use feature. Meta's Muse models have agentic integrations. What is new is the pricing and speed combination. Terra is fast enough and cheap enough that using an AI agent for routine computer tasks starts to look like a real alternative to hiring a human or writing a traditional software script.

Of course, there are caveats. Computer use agents still make mistakes. They can misclick, misread, or get stuck in loops. OpenAI itself classifies Terra as having "critical" cybersecurity capabilities under its Preparedness Framework, which means the model could potentially be used for harmful purposes if not properly safeguarded. The company rates it as "high" for biological and chemical capabilities as well, though that seems more relevant to Sol and Astra than to Terra's office-work focus.

Where It Fits

Terra is available now to ChatGPT Work users on Plus, Pro, Enterprise, and Edu plans. It is also integrated into Codex, OpenAI's coding tool, and accessible via the API for developers who want to build their own agentic workflows.

The model supports functions, web search, and file search, which means it can pull in live data and reference documents as it works. That is important for real-world tasks, where the information on a screen is often not enough to complete a job.

For developers, the API ID is gpt-6.1-terra. If you are already using GPT-6.1 Sol for agentic tasks, switching to Terra could cut your costs in half while speeding things up. The tradeoff is less raw reasoning power, but for many workflows, that tradeoff is worth it.

The Bigger Picture

OpenAI's strategy is becoming clearer with each release. Rather than one model to rule them all, the company is building a tiered family. Luna for quick chat. Terra for practical computer work. Sol and Astra for the hard problems that require deep reasoning.

The risk is fragmentation. Developers now have to choose between four models in the GPT-6 family alone, each with different strengths, pricing, and context limits. The upside is specialization. A model optimized for clicking through spreadsheets and filling out forms does not need to carry the overhead of solving advanced math proofs.

Terra also signals something about where the AI market is heading. The frontier model race is not just about being the smartest. It is about being useful at a price that makes sense. OpenAI is betting that cheaper, faster agents will win more users than marginally smarter but slower and pricier models.

Whether that bet pays off depends on whether Terra's computer use capabilities are reliable enough to trust with real work. The benchmarks look promising. The pricing is competitive. But the real test will be whether office workers and developers actually reach for Terra when they have a tedious task to automate.

If they do, OpenAI may have found the model that bridges the gap between AI demos and AI that pays for itself.