OpenAI shipped GPT-6.1 Terra on 26 October 2026, and the headline is simple: computer-use AI agents — the kind that read a screen and click through apps on their own — just got cheaper and noticeably more accurate. If you're already running or piloting agent-driven automation, this changes your cost math this week. If you're not, skip to the section that tells you so.
What Did OpenAI Actually Change With GPT-6.1 Terra?#
GPT-6.1 Terra is OpenAI's new model built specifically for "computer use" — understanding what's on a screen and taking action across apps, rather than just answering questions in a chat window. It replaces GPT-5.6 Terra as the workhorse behind agent tasks in ChatGPT Work and Codex.
The pricing moved from $2 input / $12 output per million tokens to $1 / $5, with cached input dropping from $0.20 to $0.05. On OpenAI's own Agents' Last Exam benchmark, per-task cost fell to roughly a third of the predecessor's while the score rose 15 points. Separate analysis puts Terminal-Bench 4.0 resolution at 50.3%, up from 21.5%, and factual error rate down from 53.6% to 16.7% at low reasoning effort. It's rolling out now to Plus, Pro, Enterprise and Edu users, with API access under the name gpt-6.1-terra.
| GPT-5.6 Terra | GPT-6.1 Terra | |
|---|---|---|
| Input / output per 1M tokens | $2 / $12 | $1 / $5 |
| Terminal-Bench 4.0 | 21.5% | 50.3% |
| Factual error rate (low effort) | 53.6% | 16.7% |
Does This Change Anything If You're Not Running AI Agents Yet?#
If your automation is WhatsApp flows, lead routing, or a chatbot answering FAQs, no. None of that touches a screen the way Terra does, so this announcement changes nothing for you today.
It matters if you're the kind of business that ruled out automating something — a form that lives in an old internal tool, a process spread across three logins with no API — because the cost didn't justify it. That's exactly the category Terra is aimed at. We build WhatsApp and voice automation for GCC, European and Indian clients, and the tasks we previously quoted as "technically possible but not worth the per-run cost" are the ones worth re-pricing now, not the ones already running cleanly on APIs.
We covered the adjacent story — ChatGPT's parallel agents — a few days ago. That update was about running more agents at once. This one is about each agent run costing less and erring less.
The Honest Trade-offs#
The gains are real on OpenAI's own numbers, and a 58% drop in output pricing is not a rounding error. Faster too — Terra finishes tasks over 3x quicker than Astra and Sol in testing.
What's unproven: these are OpenAI's benchmarks, not independent, third-party-audited results, and "fewer factual errors" still isn't "no factual errors" — a 16.7% error rate at low reasoning effort means roughly one in six low-effort runs still gets something wrong. Cheaper agents also means more of them get deployed faster, and access governance hasn't caught up to that pace industry-wide — we wrote about 76% of enabled AI agents sitting dormant with live credentials just this week. A cheaper agent that nobody's watching is still a liability, just a cheaper one.
How We're Already Building Around This#
We reviewed the Terra pricing the morning it landed and re-ran the cost estimate on two client automation specs that had been sitting at "maybe next quarter" because the per-task math didn't work at the old rates. At the new $1/$5 pricing, one of them now clears our usual ROI bar; the other still doesn't, because its bottleneck was never model cost, it was a login wall with no API behind it.
At WebEpex we're wiring Terra into computer-use tasks for a GCC retail client's back-office reconciliation flow this week, built on our usual n8n and self-hosted VPS stack rather than a fully managed agent platform — mostly because we want our own logging on what the agent touched, not just OpenAI's. That's the same instinct behind the containment gaps we flagged in UAE AI agent deployments: cheap and capable doesn't mean unsupervised.
What I'd Tell a Client Asking About This Week#
If you've got an automation spec sitting on a shelf because the cost didn't work last quarter, pull it back out and re-run the numbers against $1/$5 pricing — it might clear now. If you don't have anything like that, there's genuinely nothing to do here. And if you're already running agents on OpenAI's computer-use models, the move this week isn't just adopting Terra — it's pairing the switch with an access audit, since the price cut doesn't come bundled with better guardrails.
That's it. Short post, same as the news.
If you want a straight read on whether this actually moves your numbers, send me what you're running and I'll tell you — two minutes, no pitch. cal.com/webepex/growth-review
