OpenAI's GPT-6.1 Terra can operate a computer. The reported price is the part to check first.

What happened

OpenAI introduced GPT-6.1 Terra on October 26, 2026. The pitch is three things: visual reasoning, computer use, and better writing. It's available to Plus, Pro, Enterprise and Edu subscribers through ChatGPT Work and Codex, and through the API.

The number getting passed around is the price. Reports put the standard API rate at $1 per million input tokens and $5 per million output tokens, and describe the launch as a 50% cut to API pricing. Treat both as reported, not confirmed. The available coverage does not establish whether $1/$5 applies to every Terra configuration or to one tier, and OpenAI's own pricing page is the only place that settles it.

"Computer use" is the part that matters more than the price. It means the model can drive a browser and click through an interface instead of just returning text you paste somewhere.

What it could actually do in a small shop

If it works as described, the tasks on the table are the boring ones:

• Pulling source material from a fixed list of approved sites
• Turning reporting notes into a CMS-ready draft
• Updating headlines, metadata and social copy across platforms
• Assembling research packets from documents, charts and screenshots
• Sweeping pages for broken links, missing captions, formatting errors
• Moving approved text into a publishing queue

None of that is writing. All of it is production. The question Terra puts in play is whether a two-person operation can automate repetitive production work without paying enterprise-agent prices — and the launch coverage does not answer it.

The math, and what the math leaves out

At the reported rate, a job using 1 million input tokens and 200,000 output tokens costs about $2. That is the attractive headline, and for high-volume repetitive work it is genuinely low.

It is also not the bill. A production agent adds tool calls, web access, storage, monitoring, and — the line nobody budgets — failed actions. An agent that clicks the wrong thing and has to be rerun costs twice and costs you the review time. The available reporting does not include confirmed Terra pricing for any of that.

What this is not

It is not a confirmed 50% cut on your current bill. If you are on a different model or a different tier, your invoice may not move at all. Check the pricing page against your actual usage mix before you tell anyone you saved money.

It is not a cleared path to publishing live. The coverage does not confirm reliability rates, rate limits, browser permissions, or what safeguards exist around an agent pushing to a production site. Nothing in the launch says it is safe to give an agent your CMS password.

It is not a hiring decision yet. A model that can click is not the same as a model that reliably clicks the right thing 500 times in a row unattended. Nobody has published that number.

And the speed claims floating around the launch coverage are not sourced to anything primary. Skip them.

What to do this week

1. Open OpenAI's pricing page and confirm the $1/$5 rate against the exact Terra configuration you'd call. Screenshot it with the date.
2. Pull your last 30 days of API spend and recalculate it at the new rate. If the delta is under $50, this is not a budget story for you.
3. Pick one task — link checking, metadata sweeps, social copy variants — and run it manually for a week with a stopwatch. You need the human baseline before you can claim a saving.
4. Run Terra on that same task against a staging environment, never live. Count the errors, not the successes.
5. Give any agent a read-only CMS role until you have at least 100 clean runs. Publishing permission is the last thing you grant, not the first.
6. Budget for failed runs at 20% on top of token cost until you have your own number. If that kills the economics, the project was never viable at $2 a job.