OpenAI launched GPT-6.1 Terra on October 26, 2026, a model aimed at agents that operate software the way a person does: clicking through dashboards, filling forms, and pulling numbers out of tools that have no decent API. As of that date, standard API pricing is $1 per million input tokens and $5 per million output tokens, half the rate of GPT-5.6 Terra and GPT-6.1 Sol. For anyone running agency workflows on a budget, that is the headline.
What happened and when
OpenAI announced Terra on October 26 in its GPT-6.1 Terra announcement. Launch coverage from gHacks and DIY AI reports the same $1 and $5 rates. SBS adds the $0.05 cached-input rate. The API model identifier is gpt-6.1-terra. Inside ChatGPT, Terra is available to eligible paid subscribers, including Plus, Pro, Enterprise, and Edu, through ChatGPT Work and Codex.
The numbers that matter
According to OpenAI's own testing, as relayed by gHacks, Terra beats GPT-5.6 Terra by 15 percentage points on Agents' Last Exam, a benchmark of professional computer-use tasks, at medium reasoning effort and about one-third of the cost. On Terminal-Bench 4.0 at maximum reasoning effort, it solves 50.3% of tasks, compared with 21.5% for GPT-5.6 Terra. Those are vendor-run results, so read them as a direction, not a guarantee. The same coverage says OpenAI claims Terra performs on par with GPT-6 Astra at one-tenth the per-token price. We have not verified that comparison independently.
What it means for solo operators and marketers
Computer-use agents have been expensive enough that most small shops ran them as a pilot and then quietly stopped. A $1 input rate changes the math. A reporting agent that logs into four ad platforms, exports the numbers, and drafts a weekly summary can now run every day without a line item your accountant will question.
Caching deserves more attention than it gets. Agents re-read the same system prompt, brand rules, and client notes on every step, and the cached-input rate of $0.05 is where the savings compound. If your workflows reuse long instructions, the effective cost per run will sit well below the headline rate.
Two cautions. A cheaper model that clicks the wrong button is still a liability. Any agent with write access needs scoped credentials, a read-only default, and a human approval step before it spends money or sends anything to a client. Second, price per token is not price per outcome. The claim that Terra finishes tasks faster and with fewer retries is the number to test against your own workflow before you move a single job over.
Where it fits
Terra sits below GPT-6.1 Sol on price and is positioned for computer use, coding, and office work. If you already run agents on GPT-5.6 Terra, the migration case is simple on paper. Run the same ten real tasks through both models, log cost and failure rate per completed job, and let the spreadsheet make the call.
Our take
This is the most useful kind of AI announcement: a price cut that makes a category of work boring enough to automate. After 25 years in digital marketing, we have watched plenty of tools promise to replace a junior analyst. The ones that stuck were the ones cheap enough to run every morning without a budget meeting. OpenAI appears to understand that the fastest way to get an agent into every operator's workflow is to make it cost less than a lunch. We think that is good for solo operators and bad for anyone still selling per-seat software that does the same job by hand. The risk is that cheap agents get deployed carelessly, so the winners will be the operators who pair the savings with boring things like permissions, logs, and a review step. Serious about the technology, slightly amused by the pricing war, and firmly convinced that the approval button still belongs to a human.