GPT-6.1 Terra computer use capabilities turn screen reading and app actions into everyday tools for anyone handling forms, invoices, drawings, or multi-step desktop work.
Here’s the quick take:
- GPT-6.1 Terra is OpenAI’s mid-tier model built for visual understanding plus precise computer actions—clicking, typing, and moving between apps the way a person would.
- It scores higher on professional computer-use benchmarks than its predecessor while costing far less per task and finishing work faster.
- Best for high-volume office workflows: reading screens, filling forms, logging data, interpreting technical drawings, and drafting clear text.
- Available now in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, plus the API as
gpt-6.1-terra. - Pairs detailed visual reasoning with natural writing so the output feels usable right away.
That combination is the real story. Most people don’t need the absolute top-end model for every task. They need something reliable, fast, and cheap enough to run repeatedly.
What GPT-6.1 Terra computer use capabilities actually deliver
OpenAI positions GPT-6.1 Terra as the model that “understands what’s on screen, takes action across apps, writes naturally.” In practice this means the model can see a desktop or browser view, parse the UI, and then perform the next steps—clicking buttons, entering data, switching windows—without constant human guidance.
The standout evaluation is Agents’ Last Exam, a set of professional multi-step tasks. One example: turning architectural drawings into a 3D model. Another: reviewing an animated film shot by shot and producing a time-coded log. At medium reasoning effort, GPT-6.1 Terra scored 15 percentage points higher than GPT-5.6 Terra at roughly one-third the cost per task. It matched GPT-6 Astra’s score at about one-tenth the price per token and finished tasks more than three times faster than both Astra and GPT-6.1 Sol.
OSWorld 2.1, which tests agents inside real desktop applications, shows the same pattern. Medium-effort Terra already beat the previous Terra’s best score while using far fewer resources. Visual reasoning on MMMU-Pro improved across every effort setting and closed much of the gap to the frontier models at lower cost.
Pricing is straightforward. Standard API rates sit at $1 per million input tokens and $5 per million output tokens—half the rates of the prior Terra and of GPT-6.1 Sol. Cached input is even cheaper. Context window reaches roughly 1.05 million tokens. Reasoning effort can be set from none to max, so you control the speed-cost trade-off.
The computer_use tool lives in the Responses API alongside web search, file search, code interpreter, hosted shell, and others. Input is text plus images; output is text. Fine-tuning and certain other features are not supported.
How GPT-6.1 Terra computer use capabilities compare
Here’s a clean side-by-side drawn from OpenAI’s reported numbers on Agents’ Last Exam (medium and max effort where available):
| Model | Medium Effort Score | Medium Cost/Task | Max Effort Score | Max Cost/Task | Relative Speed vs Terra |
|---|---|---|---|---|---|
| GPT-6.1 Terra | 57.6% | ~$0.29 | 60.0% | ~$1.01 | Baseline |
| GPT-5.6 Terra | 42.6% | ~$0.84 | 50.7% | ~$3.58 | Slower |
| GPT-6 Astra | 57.6% | ~$4.82 | 59.3% | ~$7.24 | ~3× slower |
| GPT-6.1 Sol | 53.8% | ~$1.31 | 58.6% | ~$2.00 | ~3× slower |
| Claude Opus 5.5 | — | — | 63.2% | ~$8.82 | Much slower / costlier |
Numbers come directly from OpenAI’s launch materials. Terra sits in the sweet spot for volume work: close enough to the top models on computer-use and vision tasks, dramatically cheaper and faster for repeated runs.
In my experience, the models that win daily adoption are the ones teams can afford to leave running. Terra fits that description.
Step-by-step action plan for beginners
Want to start using GPT-6.1 Terra computer use capabilities without overcomplicating things? Follow this sequence.
- Confirm access. Log into ChatGPT with a Plus, Pro, Enterprise, or Edu plan and switch to Work or Codex mode. Developers head to the API and select
gpt-6.1-terra. - Start with a single, clear computer task. Example: “Open this invoice image, extract the vendor, amount, and date, then enter those fields into the accounting form on screen.” Keep the first prompt short and specific.
- Set reasoning effort to medium for most work. Drop to low or none when speed matters more than perfection. Raise to high or max only when the stakes justify the extra tokens and time.
- Use the Responses API if you’re building agents. Enable the computer_use tool explicitly. Feed screenshots or live screen context as images so the model can “see” the current state.
- Add guardrails early. Tell the model exactly which apps or fields it may touch and what it must never change. Review the first ten runs manually before letting it run unattended.
- Measure real cost and success rate on your own data. OpenAI’s benchmarks are useful orientation; your actual forms, drawings, and software will differ. Track tokens spent and error rate for a week, then decide whether to stay on medium effort or adjust.
- Scale only after the loop is reliable. Batch similar tasks, cache stable prompts and tool schemas, and monitor for drift.
GPT-6.1 Terra computer use capabilities What I’d do if I were setting this up for a small team tomorrow: pick one painful, repetitive process (invoice entry or form filling), run it under human review for two days, then expand. Skip the temptation to automate everything on day one.
Common mistakes and how to fix them
Mistake 1: Treating Terra like a frontier research model.
People load it with open-ended, high-stakes coding or deep analysis and then complain about gaps. Fix: Keep Terra on visual + computer-use + clear writing tasks. Route the hardest coding or research to Sol or Astra.
Mistake 2: Running everything at max effort.
That burns tokens and time for little gain on routine work. Fix: Default to medium. Only escalate when medium fails a specific pattern of tasks.
Mistake 3: Skipping screen context.
The model can’t act on what it can’t see. Vague text descriptions produce vague actions. Fix: Always supply current screenshots or structured UI state when possible.
Mistake 4: No human review loop at the start.
Early errors compound. Fix: Force a review step for the first batch of every new workflow. Log failures and feed them back as negative examples.
Mistake 5: Ignoring the Responses API requirement for tools.
Chat Completions alone won’t give you full computer_use. Fix: Switch to the Responses endpoint when you need hosted tools.
Mistake 6: Overlooking prompt caching.
Repeated system instructions and tool schemas get charged at full price if you don’t structure them for caching. Fix: Put stable content first so the system can cache it.
When GPT-6.1 Terra computer use capabilities make the most sense
High-volume, structured desktop work is the sweet spot. Invoice processing, catalog matching from product photos, form population, technical drawing interpretation, simple video logging, proposal drafting that needs both data and clean language—these play to Terra’s strengths. The model is also noticeably better at technical communication than some higher-priced siblings according to OpenAI’s own commentary.
It is not the first choice for the absolute hardest coding agent benchmarks or for tasks that demand maximum reasoning depth on every turn. OpenAI still points teams toward Sol or Astra for those.
Availability is broad. Plus, Pro, Enterprise, and Edu users can select it inside ChatGPT Work and Codex today. Developers call it via the API. Safety evaluations treat it as Critical for cybersecurity under OpenAI’s Preparedness Framework, with the same safeguard stack used on the higher models.
Key Takeaways
- GPT-6.1 Terra delivers strong computer use and visual reasoning at roughly half the previous Terra’s price and a fraction of the cost of frontier models.
- On Agents’ Last Exam it matches or approaches much more expensive models while finishing tasks over three times faster.
- Best used for repetitive, screen-based office workflows rather than pure research or elite coding.
- Medium reasoning effort is the practical default for most teams.
- Start with one well-scoped task, add human review, measure your own success rate, then expand.
- Use the Responses API and the computer_use tool for full agent behavior.
- Cache stable prompts early to keep costs under control.
- Pair it with Sol or Astra when a job truly needs higher capability.
The practical upside is simple: more of the work that used to require a person staring at a screen can now run under supervision at a cost that actually scales. Pick one process this week, give Terra a clear brief and the necessary screen context, and watch how far it gets. Adjust from real results, not from the leaderboard.
FAQs
What exactly can GPT-6.1 Terra computer use capabilities do on a real desktop?
It can view the screen, identify UI elements, click, type, switch applications, and complete multi-step sequences such as reading an invoice and entering the data into another program. Performance depends on clear instructions and good screen context.
How does pricing for GPT-6.1 Terra computer use capabilities compare to other GPT-6 models?
Standard rates are $1 per million input tokens and $5 per million output tokens—about half of GPT-6.1 Sol and the prior Terra generation. Cached tokens cost less. Per-task cost on computer-use benchmarks is often dramatically lower than Astra or Claude Opus 5.5.
Is GPT-6.1 Terra computer use capabilities ready for production agents?
Yes for well-scoped, high-volume workflows with human oversight at the start. OpenAI reports solid gains on professional computer-use tests and supplies the computer_use tool via the Responses API. Always validate on your own data and keep review loops until the error rate is acceptable.
