OpenAI launches GPT-6.1 Terra for visual work and computer tasks

OpenAI released GPT-6.1 Terra on October 26, adding a model focused on visual work, computer interaction and everyday writing to its GPT-6 lineup. The release appears in the company's official API changelog for that date. Its positioning addresses a practical question for AI agents: how much capability an application needs when the assignment involves reading a screen and completing a defined job.

OpenAI's product documentation describes matching supplier photographs with catalog entries, updating listings and reviewing campaign images for cropping or text placement. Those examples connect image understanding with subsequent actions in software. For a business workflow, recognizing a product is only one stage; getting the corresponding record updated is the useful result. Browser and computer access still require their own setup and permissions.

Standard API text pricing is $1 per million uncached input tokens, $0.05 for cached input and $5 for output. Tokens are the units used to measure the material a model processes or produces. These are Standard rates, rather than a universal price for every request. Cache writes have separate charges, and service choices such as Fast, Batch or Flex alter the bill.

The model page lists a 1.05 million token context window, the space available for a request's working material, and up to 128,000 output tokens. It accepts text and images and returns text. Tool calling requires the Responses API. Large requests need particular cost attention: when input exceeds 272,000 tokens, input and cache rates double and output rates rise to 1.5 times the base rate for the entire request.

OpenAI reports a 15 percentage point improvement over GPT-5.6 Terra on Agents' Last Exam at medium reasoning effort, at roughly one-third the evaluated cost. The benchmark covers professional computer tasks. That is a company-reported test comparison, rather than an established saving for every customer. Its relevance is the attempt to improve completed work per unit of spending, where retries and task difficulty can change the economics.

The company also reports that errors on its internal set of difficult factual prompts fell from 53.6% with GPT-5.6 Terra to 16.7% with the new model at low reasoning effort. Terra's maximum-effort error rate was 4.3%, so the result does not mean fabricated answers have disappeared. For command-line software tasks in Terminal-Bench 4.0, OpenAI reports 50.3% completion at maximum effort, against GPT-5.6 Terra's published 21.5%.

Current ChatGPT Learn guidance lists access through ChatGPT Work and Codex on Plus, Pro, Business, Enterprise and Edu plans. Terra can be selected in Advanced on desktop or Work on the web, with mobile support in Work. It is unavailable in Chat. Enterprise access depends on workspace availability and administrator controls, and the model does not support UAE inference residency. Access through an eligible subscription therefore remains distinct from API billing and tool permissions.

The accompanying safety addendum treats Terra as meeting the Critical capability level for cybersecurity under OpenAI's Preparedness Framework and applies the safeguards used for GPT-6.1 Sol and GPT-6 Astra. This is a capability assessment, not a report of a particular attack. The document also cautions that research or API evaluations can differ from production ChatGPT because system instructions, available tools and reasoning settings vary, placing limits on how directly the published results describe everyday use.