The News, Explained
OpenAI introduced GPT-6.1 Terra on October 26, 2026, targeting visual understanding, computer use and writing with an emphasis on speed and cost. Its proposed uses connect recognition with action, such as matching product photos to catalog entries or working with forms on screen. Source
The official Work guide lists Plus, Pro, Business, Enterprise and Edu access through ChatGPT Work and Codex. Terra can be selected from Advanced in the desktop app or on the web, and Work also supports it on mobile. It is unavailable in Chat and does not support UAE inference residency, the option to keep inference processing within the UAE. Enterprise access follows workspace availability and administrator controls; browser and computer use retain their own setup and access requirements. Source
Developers can request gpt-6.1-terra through the API, the connection used by software to send work to a model. Its model page specifies text and image input with text output. Calling tools—asking an external function to perform an action—requires the Responses API. Screen understanding and the permissions needed to act in an application therefore deserve attention together. Source
API billing distinguishes token categories. Tokens are the chunks used to process content and calculate charges. Input, generated output and reused context carry different rates. A cache stores context for reuse. The table gives US-dollar prices per million tokens for ordinary Standard text requests. Source
| Billing category | Standard base price | Condition |
|---|---|---|
| Ordinary input | $1 | Input sent to the model |
| Cached input | $0.05 | Reused stored context |
| Output | $5 | Tokens generated by the model |
| Cache writes | $1.25 / $2.50 | 30-minute / 12-hour retention |
Requests exceeding 272,000 input tokens apply twice the input and cache rates and 1.5 times the output rate to the full request. Fast mode costs twice Standard; regional processing adds 10% where available. Batch and Flex are 50% below Standard. Tool charges also need accounting. These are API rates; they cannot convert a Work or Codex subscription allowance into dollars. Source Source
For a hypothetical calculation, assume a Standard text request with no caching, images, tools, regional processing or Fast mode. Ordinary input of 100,000 tokens costs $0.10, while 10,000 output tokens cost $0.05: $0.15 combined. The input stays below the long-context threshold. A real workflow’s total must also include its retries and additional tools. Source
Performance figures need their task and settings attached. OpenAI reports that on an internal evaluation designed to elicit factual errors, the low-reasoning-effort error rate fell from 53.6% for GPT-5.6 Terra to 16.7% for GPT-6.1 Terra. Reasoning effort controls how much thinking the model uses to produce an answer. The system card says research or API evaluations can differ from production ChatGPT because of system prompts, tools and reasoning settings. Applying that figure to a particular workflow requires evaluating that workflow separately. Source Source
OYOPICK’s Take
Terra’s promise lies in making screen-based assistance economical enough to use more often. For work, the useful comparison is the cost of a reviewed result. Failed attempts and revision requests add to the illustrative bill above and change the benefit of a lower token rate.
Recording input size, required tools and correction attempts should therefore help model selection. If Terra can maintain accuracy while reducing retries on screen and document tasks, automation could become useful to more people. We would like services to make usage per completed task and its revision history easier to see, so readers can judge that possibility against their own work.