What OpenAI shipped

OpenAI released GPT-6.1 Terra on 2026-10-26. The announcement, as summarised by Simon Willison, frames it as “a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing.” OpenAI calls it a significant step up from GPT-5.6 Terra and points at the tasks that fill an actual workday: drafting proposals, interpreting technical drawings, filling out forms, and editing videos.

That list is worth reading twice. It is not a benchmark chart. It is a description of desktop work - the kind of thing an agent does by looking at a screen and clicking.

Where Terra sits in the lineup

Terra is the middle child. Willison’s read is that it occupies the same slot as Sonnet in Anthropic’s collection: not the cheapest, not the most expensive, and therefore the hardest one to decide about.

An OpenAI engineer, Ted Sanders, gave a blunter comparison in a Hacker News comment quoted in the same post: Terra is “probably better than Astra and Sol at technical communication, similar on vision and computer use (but much faster!), and a bit worse on most other things.” For API developers he frames it as the option between Sol and Luna - cheaper than the flagship, much smarter than the budget model, especially for vision, computer use, and writing.

Pricing backs that up:

  • $1 per million input tokens
  • $5 per million output tokens
  • about half the price of GPT-5.6 Terra and GPT-6.1 Sol
  • about 10x the price of GPT-6 Luna

Why computer use is the part that matters

A chat turn is short. A computer-use loop is not. Every step ships a screenshot, a DOM or accessibility dump, and the accumulated history back to the model, then gets a small action back. A hundred steps is a normal afternoon for an agent, and input tokens dominate the bill.

Do the arithmetic on a modest run: 100 steps at roughly 1,500 input tokens each is 150,000 input tokens. At Terra’s input rate that is about $0.15. The same run on a model priced like Sol is about $0.30. Neither number sounds scary in isolation - but multiply by every scheduled task, every retry, and every parallel session, and the difference between “run it hourly” and “run it when I remember” is exactly this ratio.

Speed compounds the same way. Sanders’ point that Terra is much faster on computer use matters more than a benchmark delta, because in a screenshot loop latency is paid once per step. A model that answers in half the time finishes a fifty-step task in half the wall clock, which is the difference between an agent you watch and an agent you schedule.

The caveat is the one he states plainly: Terra is “a bit worse on most other things.” If your task is a single hard reasoning step - a tricky migration, an ambiguous bug - the cheap model will cost you more in re-runs than it saves per token.

What this changes for Windows desktop agents

Three practical shifts:

  1. Long loops become affordable. Form filling, repetitive navigation, and draft-then-edit cycles were previously priced like a luxury. At mid-range rates they are closer to a utility bill.
  2. Parallelism gets cheaper than sequencing. Running three candidate approaches at once is now often cheaper than one careful attempt with a frontier model - which changes how you design retries.
  3. The bottleneck moves to verification. Once the model is cheap and fast, the expensive part of your pipeline is deciding whether the result is correct. That is a human-and-tooling problem, not a model problem.

None of this makes an agent safe to leave alone. A cheaper click is still a click, and the blast radius of a wrong one is set by what the agent is allowed to touch, not by the price per token.

How to evaluate it in one afternoon

Pick one repetitive task you already do weekly, and run it end to end:

  • Give the agent the same starting state each time (a fresh profile, a known folder).
  • Cap the step count so a runaway loop fails loudly instead of quietly burning tokens.
  • Log every action with a timestamp, so you can compare a Terra run against your current model on the same task.
  • Measure three things only: steps to completion, wall-clock time, and how often you had to intervene.
  • Keep one approval gate on anything that writes outside a designated folder.

If Terra wins on all three and the interventions are the same kind you already accept, it is the cheaper default. If interventions go up, the savings are imaginary - your time is the expensive token.

Ready to try a desktop agent that drives real Windows applications rather than a browser tab? Download MOOGH and point it at the same task you used for the comparison.

FAQ

Is Terra a replacement for the flagship models?

No. OpenAI’s own framing puts it between Sol and Luna, and its engineer says it is slightly worse on most non-communication tasks. Treat it as the workhorse you use by default and escalate from, not the model you use for everything.

Does the cheaper price mean I can skip approvals?

It means the opposite. Cheaper loops encourage longer runs and more parallelism, which multiplies the number of actions taken without a human watching. Approvals, scoped permissions, and audit logs become more important, not less.