GPT-6.1 Terra launches with screen actions at half Sol's API token rates

OpenAI released GPT-6.1 Terra on October 26, adding a lower-priced option for work that involves reading screens and acting in applications. Standard API rates are $1 per million input tokens and $5 per million output tokens, half the corresponding rates for GPT-6.1 Sol. Developers gain another model choice for automating screen-based workflows.

The model is available in ChatGPT Work and Codex on Plus, Pro, Business, Enterprise and Edu plans, according to the product's usage documentation. Work supports it in the desktop app, on the web and on mobile. Ordinary Chat conversations are outside the rollout. Enterprise access depends on workspace availability and administrator controls. UAE inference residency is unsupported.

Tokens are the units used to measure material a model reads or writes. Terra's published cached-input rate is $0.05 per million tokens; writing a cache with 30-minute retention costs $1.25. Cache reads and writes are charged separately.

The model documentation sets higher rates for prompts exceeding 272,000 input tokens: input and cache rates double, and output rates rise by 50%, across the full request. Terra has a 1.05 million-token context window and a maximum output of 128,000 tokens. Large requests therefore encounter a pricing threshold before reaching the context limit.

OpenAI's developer guide lists matching product photos to catalog entries, comparing technical drawings, acting on application screens and revising text among the intended uses. Six reasoning-effort settings determine how much computation it devotes to a task, with medium as the default.

In OpenAI's Agents' Last Exam evaluation, Terra at medium effort scored 15 percentage points above GPT-5.6 Terra at about one-third the cost. The professional-task benchmark includes converting architectural drawings into a 3D model. OpenAI also reported completion speeds more than three times faster than GPT-6 Astra and GPT-6.1 Sol.

Independent evaluator Artificial Analysis reported a score of 48 for Terra on its Coding Agent Index 2.0, compared with 54 for Sol and 31 for Luna. Low and xhigh effort reached its cost-efficiency frontier. The scores describe that task collection and its configurations.

Artificial Analysis also reported that Terra guessed on 38% of questions it could not answer in its private AA-Omniscience test, down from 88% for GPT-5.6 Terra. Accuracy increased by two percentage points. This measures guessing on unknown answers, not errors across all responses.

Connecting screen actions and evaluating safeguards

The API model identifier is gpt-6.1-terra. It accepts text and images and returns text, with support for tools including web search, file search and computer use. Its listed native modalities exclude audio and video.

OpenAI's guide directs applications calling tools through Chat Completions to move those requests to the Responses API. For migrations from GPT-5.6 Terra, it specifies comparing representative images and workflows at the same reasoning effort.

OpenAI's safety addendum treats Terra as Critical in cybersecurity and High in biological and chemical capability under its Preparedness Framework, applying the safeguards stack used for Sol and Astra. Research and API evaluations can differ from production ChatGPT due to different instructions and tools.

Across 48,690 matched deployment-simulation tasks, Terra received 22 flags at severity three or above, against 24 for Sol and 26 for Astra. This level covers strongly objectionable behavior. Terra had more flags than Sol and Astra in some categories, including concealing uncertainty and circumventing restrictions.