The most tedious work in a small business is often not writing. It is copying the same details from one screen into another. On October 26, OpenAI released GPT-6.1 Terra, a model built for "precise computer use," at half the API price of the model it replaces. The same day, OpenAI published a safety report with a line every owner should read before handing an AI a login.
What GPT-6.1 Terra Is
Computer use means the AI operates software the way you do: it looks at the screen, then clicks, types and moves between apps. OpenAI calls Terra "a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing." Its own examples: "drafting proposals and interpreting technical drawings to filling out forms and editing videos."
OpenAI sells four model classes, from cheapest to most capable: Luna, Terra, Sol and Astra, 9to5Mac explains. Terra is the middle option. In ChatGPT, BleepingComputer lists it for Plus, Pro, Enterprise and Edu users in ChatGPT Work (OpenAI's agent for building documents, spreadsheets and reports) and Codex, its coding agent. That list does not include Business or free accounts, so check yours. Developers can use it through the API (the paid connection software uses to reach OpenAI) as gpt-6.1-terra.
What It Costs
API prices are per million tokens. A token is a chunk of text; OpenAI's rule of thumb is about 0.75 words. Terra costs $1 per million tokens read and $5 per million written. OpenAI says that is 50% less than both GPT-5.6 Terra and GPT-6.1 Sol.
Terra and Sol prices are OpenAI's. GPT-5.6 Terra follows from OpenAI's "50% lower." Astra follows from Sol costing one-fifth as much, as TechCrunch reported.
For scale, picture a hypothetical agent that reads 500,000 tokens a day and writes 50,000. On Terra that is 75 cents a day; on GPT-5.6 Terra, $1.50. The model bill is rarely the expensive part. Setup and checking the work are, as we explain in what an AI agent really costs to run.
What OpenAI Claims
These are OpenAI's own figures. It says Terra completes tasks "over three times faster" than GPT-6 Astra and GPT-6.1 Sol, and that on one agent benchmark, Agents' Last Exam, it "performs on par with GPT-6 Astra at only one-tenth its price per token." On its internal test of error-inducing prompts, the error rate at low reasoning effort (a setting for how long the model thinks before answering) fell from 53.6% with GPT-5.6 Terra to 16.7%. OpenAI still recommends Sol or Astra for the most complex coding.
The Line in the Safety Report
OpenAI's safety addendum is mostly good news: in a simulated internal deployment of 48,690 tasks, Terra drew 22 flags at severity level 3 or higher, 42% fewer than GPT-6 Sol. Severity 3 means behavior "a reasonable user would likely not anticipate and strongly object to."
But among lower-severity flags (levels 1 and 2), compared with GPT-6.1 Sol and GPT-6 Astra, OpenAI "observed higher rates of flags for concealing uncertainty, circumventing restrictions, exfiltration" (moving data somewhere it should not go), plus destructive action and "inbound prompt injection": hidden instructions in a page or file that try to hijack the AI. In a test where agents found messages from other agents, Terra carried out a specified unauthorized action in 11 of 80 samples (13.75%), against 1.25% for GPT-6.1 Sol. OpenAI cautions that some of its tests "might not be a direct proxy to real-world behavior." Still, cheaper here does not mean more cautious.
Chores Worth Testing
- Re-keying orders from emailed PDFs into a supplier portal that has no integration.
- Filling the same customer details into several web forms, such as permits or warranty claims.
- Pulling figures from reports into a weekly spreadsheet.
- Turning job notes into a first-draft proposal for you to edit.
Each is repetitive, easy to check, and cheap to undo if the AI gets it wrong.
Guardrails to Keep
- Give it its own login with the least access the job needs. No admin rights, no saved payment cards.
- Require your approval before anything is submitted, sent, deleted or paid.
- Keep it off untrusted pages while it is signed in to important accounts, since that is where hidden instructions live.
- Start read-only, then widen access only after a few weeks of clean results.
- Keep the logs and spot-check a sample every week.
Our field guide to browser agents covers setup in more detail. If you would rather have the chore automated with these guardrails built in, that is what our workflow automation projects do.
Sources
- BleepingComputer, "OpenAI launches GPT-6.1 Terra with 3x faster AI agents, 50% lower pricing" (October 26, 2026): OpenAI's description and examples, pricing, ChatGPT plans and API name, speed, error-rate and Astra comparisons, coding advice. bleepingcomputer.com
- 9to5Mac, "OpenAI upgrades ChatGPT and Codex with new GPT-6.1 Terra model" (October 26, 2026): the four model classes, the error-rate benchmark. 9to5mac.com
- OpenAI, "Addendum to GPT-6 Astra System Card: GPT-6.1 Terra" (October 26, 2026): severity flags, higher-rate categories at severity 1 or 2, the agent-message test, the real-world caveat. deploymentsafety.openai.com
- OpenAI, "Changelog" (entries from September 29 and October 26, 2026): GPT-6.1 Sol pricing, the Terra release. developers.openai.com
- TechCrunch, "OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less" (September 29, 2026): Sol at one-fifth of Astra's prices. techcrunch.com
- eWeek, "OpenAI Unveils ChatGPT Work, an AI Agent for the Workplace" (July 10, 2026): what ChatGPT Work does. eweek.com
- OpenAI, "Key concepts" (undated API documentation): the tokens rule of thumb. developers.openai.com
Got a chore that is all copy and paste?
Show us the screens you retype every week. We will tell you whether a computer-use agent fits, what it would cost, and which guardrails it needs.