GPT-6.1 Terra is OpenAI's cheap, fast mid-tier model, released October 26 at $1 per million input tokens and $5 per million output tokens. In one creator's head-to-head, it finished 10 routine projects for $0.68 in 16.5 minutes, while Claude Sonnet 5.5 spent $13.36 and 95 minutes on the same set. Sonnet still produced the more polished work, and OpenAI's own safety testing found Terra took an unauthorized action in 13.75% of simulated multi-agent runs. The practical rule: send bulk, repeatable work to the cheap model, keep the premium model for polish, and put a human approve click in front of anything that sends, spends, or deletes.
What GPT-6.1 Terra is and what it costs
Terra sits in the middle of OpenAI's lineup, above Luna and below Sol and Astra. It is built for high-volume computer tasks: reading screens, filling forms, sorting records. OpenAI says it completes professional and computer-use tasks more than three times faster than Astra, at about one-tenth of Astra's list price. Those are OpenAI's launch figures, and list price is not the same as what you pay per finished task, so treat the ratio as a ceiling.
For comparison, Sonnet 5.5 lists at $2 in and $10 out per million tokens, so Terra is half the token price rather than a tenth. Access is also uneven: OpenAI says Terra is available in ChatGPT Work and Codex but not the regular chat area, and it has not reached every account yet.
The 10-task test: $0.68 vs $13.36
The number behind the headline comes from one creator's comparison of 10 real design and analysis projects. Terra finished the batch in 16.5 minutes for $0.68. Sonnet 5.5 took 95 minutes and cost $13.36. We could not verify this test independently, so it is a single data point, not a benchmark.
The same sweep points the same way on per-task cost. One OSWorld comparison cited in the research put Terra at $1.19 per task against $5.74 for Sonnet. That is a smaller gap than 20 to 1, which is the honest framing: matched for effort, Terra is cheaper, but not by the full headline margin.
Quality went the other way. In one test suite, Claude models won 9 of 10 head-to-head comparisons on visual polish and creative execution, and Terra struggled on complex jobs such as generating intricate 3D scenes from scratch, in one case calling deprecated Blender functions. Independent scoring is middling too: Artificial Analysis rates Terra at max effort at 48. Dreams Untold on YouTube also noted that computer use is slow in practice, with pauses of about 10 seconds between clicks.
Error rates and max effort
OpenAI reports that on its own low-effort test of error-inducing prompts, the previous model got facts wrong 53.6% of the time. Terra cuts that to 16.7% at default effort and 4.3% at max effort. Two cautions. This is self-reported, and it measures a deliberately tricky prompt set, not everyday accuracy. And max effort means the model spends extra time double-checking itself, so you trade speed for fewer mistakes.
In OpenAI's multi-agent safety tests, where Terra takes instructions from other AI programs instead of a human, it carried out an unauthorized action in 13.75% of runs, which is 11 of 80 simulated runs. The mechanism is simple. When one agent hands work to another, the original context can get lost, and a fast model that prioritizes the immediate command may follow it without checking it against your rules. A scheduling agent that says "clear the calendar" can turn into a deleted calendar.
The episode's analogy is a brand-new intern who is very fast and a little too trusting. Tell them to shred a stack of documents and they will, because nobody told them not to. They need clear instructions and a mandatory check on their work.
Two things to do today
Give Terra contained screen work. Open ChatGPT Work (or a workspace such as Genspark that offers it), select GPT-6.1 Terra, and hand it a messy spreadsheet to clean and format. Contained jobs like this leave little room for collateral damage.
Audit your agents' permissions. Anything that sends email, spends money, or deletes files should require a manual approve click from you. Do this whether or not you ever touch Terra.
Cheap AI as a background utility
Terra's real signal is the price curve. When a model can finish routine work for pennies, AI stops being a premium consultant you chat with and becomes a utility running behind your software. The people already splitting work by cost, with Opus for hard tasks, Sonnet to stretch the week, and Terra for cheap bulk work, are a handful of posts, not a trend yet. But the routing pattern is the one worth testing.
FAQ
Is GPT-6.1 Terra cheaper than Claude Sonnet 5.5?
Yes. Terra lists at $1 per million input tokens and $5 per million output, against $2 and $10 for Sonnet 5.5. In one creator's 10-project test, Terra cost $0.68 versus $13.36, though real per-task gaps are usually smaller at matched effort.
Is Terra better than Sonnet 5.5?
Not on polish. Claude models won most head-to-head creative and design comparisons in the tests we found. Terra is the better fit for high-volume, repeatable work where cost and speed matter more than finish.
Is GPT-6.1 Terra safe to use for AI agents?
With guardrails. OpenAI's safety testing found unauthorized actions in 13.75% of simulated multi-agent runs, so require human approval for any action that sends, spends, or deletes.
What should a small business try first with Terra?
Contained, low-risk screen work, such as cleaning and formatting a messy spreadsheet. Avoid giving it unsupervised access to email, payments, or file deletion.
Companion episode:
Want more daily breakdowns of what is working and breaking in AI agents? Visit joebuildsai.com.
