Duncan Rogoff's five-minute breakdown gets to the central idea of GPT-6.1 Terra: OpenAI is positioning a cheaper model for screen-based work, image understanding and everyday agent tasks, rather than making every interaction use its most expensive frontier model. The launch is real, and Terra's API model ID is gpt-6.1-terra. What is still a hypothesis is whether it will be the best model for your computer-use workflow, or become the model behind a particular personal agent.
Terra is worth testing for bounded work such as reading a screen, matching visual information, updating a form or making a scoped code change. Its standard API rate is $1 input / $5 output per million tokens, one-quarter of Claude Opus 5.5's published $4 / $20 base rates. OpenAI reports strong computer-use efficiency on its own evaluations; do not treat those results as a universal nine-fold saving or permission to give an agent unrestricted access. Start with a reversible task and verify the saved result.
Watch Duncan's Breakdown
Source note: The video is a commentary on OpenAI's announcement, not an independent hands-on benchmark. The figures below come from OpenAI's published tests unless another source is named. We checked the API model card and Anthropic's Opus price on 26 October 2026.
What Terra Is Actually Built to Do
OpenAI describes Terra as a fast, efficient model for visual reasoning, precise computer use and natural writing. Examples in the launch include interpreting technical drawings, matching product images to catalog entries, drafting proposals and completing forms. OpenAI's model-selection guidance recommends it for defined computer-use outcomes, concise writing from supplied material and scoped code fixes; it steers harder, ambiguous work toward GPT-6.1 Sol or GPT-6 Astra.
The API card lists text and image input, text output, a 1.05-million-token context window and up to 128,000 output tokens. It does not list native video input. When the video or launch article talks about editing video, the practical path is an agent using a video application and screen or frame information, not simply handing Terra a movie file and assuming it can perceive and edit the entire timeline by itself. For API tool calling, OpenAI says to use the Responses API.
Duncan thinks these traits make Terra a natural fit for personal agents such as ChatGPT Dots. That is a plausible inference, but OpenAI's release does not announce Terra as Dots' engine or confirm support in third-party agent products. Do not build a purchasing or deployment decision around that prediction.
What the Launch Benchmarks Say, and Do Not Say
| Evaluation | OpenAI's reported result | What it means |
|---|---|---|
| Agents' Last Exam v1 | At medium effort, 15 percentage points above GPT-5.6 Terra at about one-third the cost; near GPT-6 Astra on this test. At max effort, OpenAI reports roughly 9x lower cost per task than Claude Opus 5.5. | A targeted professional-agent benchmark, not a guarantee on your workflow. |
| OSWorld 2.1, offline set | At medium effort, above GPT-5.6 Terra's best score at a lower reported cost per task. | Evidence for desktop-app tasks under the published setup, not production reliability. |
| MMMU-Pro | Above GPT-5.6 Terra at every tested reasoning setting and near Sol/Astra on visual reasoning. | Visual interpretation improvement; it does not measure permission handling. |
| Terminal-Bench 4.0 | 50.3% solved at max effort versus the prior Terra model's published 21.5%. | Better scoped coding performance, while many tasks still fail. |
| Internal factual-error test | Error rate at low effort fell from 53.6% to 16.7% on OpenAI's error-inducing prompts. | An internal distribution, not a universal hallucination rate. |
These are OpenAI's published comparisons, with model settings and task mixes chosen for the release. The “3x faster” figure compares average completion on the cited professional tasks with GPT-6 Astra and GPT-6.1 Sol; the “9x cheaper per task” figure compares Terra and Opus on that evaluation at max effort. Neither says Terra is nine times cheaper for every request. The right decision test is to run both candidates on the same representative tasks and measure successful completions, elapsed time, retries, review effort and total bill.
The API Price Versus Claude Opus 5.5
At standard short-context text rates, the two vendors' current model pages list:
| Model | Input / 1M tokens | Output / 1M tokens | Cached input / 1M |
|---|---|---|---|
| GPT-6.1 Terra | $1.00 | $5.00 | $0.05 |
| Claude Opus 5.5 | $4.00 | $20.00 | $0.20 |
For an illustrative request with 100,000 uncached input tokens and 20,000 output tokens, the base token arithmetic is $0.20 for Terra versus $0.80 for Opus. That is a like-for-like token quantity, not a like-for-like completed task: models may use different output lengths or need different numbers of attempts. Computer-use calls and other tools may have separate charges; cache writes are not the same as cache reads. Terra's model card also says requests above 272,000 input tokens incur higher long-context rates for the full request, and Fast mode costs more. Check the live pricing page before budgeting an automation.
For ChatGPT Work or Codex subscriptions, OpenAI says Terra makes an existing usage allowance go further. That is not a promise of unlimited work or a conversion of subscription usage into the API token table. Plan limits and product access are separate from pay-as-you-go API charges.
Where to Use Terra Today
- ChatGPT Work and Codex: OpenAI's current access guide lists Plus, Pro, Business, Enterprise and Edu. In the desktop app or Work on the web, choose Terra from the Advanced model list; Work also supports it on mobile. The guide says Terra is not available in ordinary Chat, and enterprise access can depend on admin settings.
- API: use the model ID
gpt-6.1-terrain a project with model access. The model card confirms computer use and other tools through the Responses API; free API tier access is not listed for this model. - Computer actions: selecting Terra alone does not grant access to your machine. Browser or desktop use needs a configured execution environment, the relevant product permissions and an action loop. OpenAI's computer-use guide describes the application-owned driver and verification responsibilities.
“Critical” Capability Is Not a Safety Guarantee
Duncan correctly notes that OpenAI emphasizes safeguards, but the terminology matters. OpenAI's Terra system-card addendum says it treats Terra as Critical in cybersecurity capability under its Preparedness Framework and applies the safeguards stack used for Sol and Astra. “Critical” describes the capability/risk category; it does not mean every cybersecurity test was passed or that an agent cannot make a harmful mistake. OpenAI also says its safety evaluations may differ from behavior in the production products because prompts, tools and settings vary.
The developer guide still calls for an isolated browser or VM, limited sites and actions, treating page content as untrusted, confirmation before purchases or data transmission, bounded steps and costs, and checking the actual saved outcome. Stronger prompt-injection scores are welcome but do not remove the need for those controls.
One Useful First Experiment
Give Terra a reversible, low-stakes job: compare five product photos with a supplied catalog and prepare a draft mapping, or extract invoice fields into a review queue rather than posting them to an accounting system. Provide the source files and acceptance criteria. Let the agent use only the specific app or test environment it needs, and require a human to approve any external submission.
Then reopen the records yourself. Count correct and incorrect fields, time to completion, tool calls, retries and total cost. Repeat the same set with the model you use now. This is more informative than a launch-chart headline: it tells you whether Terra actually saves money after review and repair in your workflow.
Video Chapters by Topic
Use these links to jump to Duncan's explanation. The official source links above provide the exact specs and access rules behind his summary.
Model and use case
Performance and price
| Time | Topic |
|---|---|
| 01:29 | Speed, capability and cost |
| 01:57 | The computer-use benchmark |
| 02:44 | Cost versus Claude Opus 5.5 |
Responsible first test
| Time | Topic |
|---|---|
| 04:21 | Safety and what to try first |
Sources and Scope
Video: Duncan Rogoff's Terra breakdown and the user-supplied transcript. OpenAI Docs and primary material: launch announcement, API model card, Work/Codex access guide, computer-use guidance and safety addendum. Comparison: Anthropic's official Opus 5.5 pricing. Checked 26 October 2026. We did not run independent Terra-versus-Opus tasks or verify the launch benchmark scores outside the vendors' published material. The Dots connection is commentary, not an announced feature.