Yesterday’s model is today’s default for grunt work

OpenAI released GPT-6.1 Terra on October 26, 2026, and the headline is refreshingly unglamorous: this is not the frontier model. Terra replaces GPT-5.6 Terra as the middle of OpenAI’s current lineup — Luna at the cheap end, then Terra, then Sol, then the flagship Astra — and it’s aimed squarely at the high-volume work that fills a workday. OpenAI’s announcement describes it as “a fast and efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing.”

So what is “computer use,” exactly? Not a chatbot that describes how to fill out a form. A computer-use model looks at what’s on screen, then clicks, types, and moves between applications the way a person would — reading an invoice and entering it into accounting software, matching product photos to catalog entries, interpreting a technical drawing, drafting a proposal. PYMNTS’ launch coverage puts the same pitch in enterprise terms: automating high-volume work, not demoing a party trick.

The price is the story

Through the API, Terra costs $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens at standard rates — 50% below both GPT-5.6 Terra and the larger GPT-6.1 Sol, according to OpenAI. BleepingComputer’s launch report confirms the same figures, and SiliconANGLE notes the cut applies to standard API rates, which is what developers actually pay to build with the model.

Two practical details matter if you run it at volume. First, the context window is about 1.05 million tokens with up to 128,000 tokens of output, so very long documents fit — but OpenAI’s docs note higher rates kick in for prompts above 272,000 input tokens, applied to the whole request. Second, Terra is available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, and to developers as gpt-6.1-terra — it is not in regular Chat, so don’t go hunting for it in the model picker there.

Read the benchmarks like a grown-up

All of the performance numbers below are OpenAI’s own. They have not been independently verified, and that’s worth saying out loud before quoting any of them:

• On Agents’ Last Exam, a test of professional agent tasks, OpenAI says Terra scored 15 percentage points higher than GPT-5.6 Terra at medium reasoning effort, at about a third of the cost — and performed on par with the flagship Astra at one-tenth the price per token.
• OpenAI says Terra finishes tasks more than three times faster than Astra or Sol on its own timing runs.
• On Terminal-Bench 4.0, a coding test, Terra solved 50.3% of tasks at maximum effort, up from 21.5% for GPT-5.6 Terra.
• On OpenAI’s internal set of prompts designed to trigger factual errors, the error rate at low reasoning effort fell from 53.6% to 16.7%, and to 4.3% at maximum effort — nearly matching Astra and Sol.

If those numbers hold up under independent testing, the interesting one isn’t the coding score. It’s the combination: a model that’s good enough at computer tasks, fast, and cheap per task. That’s the profile that ends up handling most of the volume in real systems, while the flagship models stay reserved for the hardest problems. OpenAI itself says as much — for in-depth research or the most complex coding, it still points you to Sol or Astra.

About that word: “Critical”

OpenAI also says Terra meets the Critical cybersecurity threshold under its Preparedness Framework, keeping the protections developed for Astra. That word deserves a calm read. Under OpenAI’s own definition, “Critical” is a capability ceiling: with the right tools and access, the model can find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step. OpenAI’s “Path to Astra” post lays out exactly that definition, and the safeguards it says follow from it — training the model to refuse harmful cyber requests, added protections against misuse, and monitoring that can stop unauthorized activity.

“Critical” is a statement about what the model could do unsupervised, not a claim that every interaction with it is dangerous, and not a reason to avoid the model for ordinary work. But it is a reason to treat computer-use agents the way you’d treat any new employee with broad system access: least-privilege accounts, approval required before sensitive actions, separate credentials for agent work, and a human reviewing anything that moves money or changes production. The cheaper the agent gets, the more of them you’ll run — and the more those guardrails matter.

Who should actually switch

Try it now if: your agents spend their days clicking through UIs, filling forms, and processing documents; you’re running high-volume Codex or ChatGPT Work workloads and want more completed tasks per allowance; or you’re choosing a default model for everyday coding where cost and latency dominate.

Stay on Sol or Astra if: you’re doing the hardest coding, deep research, or work where a single wrong step is expensive. And whatever you run, re-run your own tasks on Terra before switching a production workload over — vendor benchmarks are a starting point, not a migration plan.

The unglamorous truth about this release is also why I like it: the AI industry spent years selling us the biggest model, and the model that will actually change how most offices work is the cheaper, faster one that just clicks the buttons. Terra is OpenAI betting that the workhorse, not the show horse, is where the volume is. That bet looks about right to me.