OpenAI GPT-6.1 Terra

Overview

GPT-6.1 Terra by OpenAI is a text model that can also read images. It reasons over screenshots, photos, and technical visuals while keeping responses clear and direct. It can request tool actions during a task, then use the tool results to finish the job. This is useful when you need one model to both understand what it sees and produce a final written answer.

What you can build

• Screenshot-based support agents that explain what went wrong and what to click next
• Visual QA for product photos, packaging, and catalog matching
• Technical drawing interpretation with written checks, measurements, and issue lists
• Form-filling helpers that extract fields from screenshots and draft the entries
• Coding assistants that read UI errors from screenshots and propose scoped fixes
• Agent workflows that combine vision, tool calls, and a final human-ready summary

Inputs

• A written instruction describing the task and the expected output format.
• Optional conversation turns as a message list, instead of a single instruction.
• Optional images for the model to analyze. Provide common formats like PNG or JPEG.
• Support for multiple images in one run. This listing supports up to 10 images.
• Optional system-level instructions that set tone, rules, and boundaries.
• An optional reasoning effort setting that ranges from no extra reasoning up to maximum.
• An optional verbosity setting that nudges outputs to be shorter or longer.
• An optional cap on how many tokens the model may generate in its reply.
• Optional tool definitions, provided as structured JSON, when you want tool calling.
• Optional tool behavior controls, such as letting the model choose tools automatically.
• Optional parallel tool calling, when you want multiple tool calls in one model turn.
• Optional user and session identifiers to keep chat history grouped correctly.
• Optional continuation inputs for resuming a prior run, including tool results.
• Optional web search enablement for runs where browsing is permitted.

Outputs

GPT-6.1 Terra returns a text response as its primary output.

When tools are enabled, it can also return structured tool call requests. Each request includes the tool name and a JSON arguments object. After you provide tool results back to the run, the model can continue and produce the final text answer.

If you use structured output modes, it can return machine-readable JSON content instead of free-form text.

Limitations

• It supports text input and text output.
• It supports image input, but it does not support audio or video inputs.
• Fine-tuning is not supported.
• It has a 1,050,000 token context window.
• It can generate up to 128,000 output tokens in a single response.
• Its knowledge cutoff is June 1, 2026.
• Tool calling only works when you provide tool definitions and accept tool calls.
• Low-quality images can degrade results. Expect errors with blurry screenshots, tiny UI text, glare, heavy compression, or scanned pages.
• If your inputs mix unrelated screenshots, the model may merge contexts unless you label each image clearly.

Safety & compliance

• Follow OpenAI’s usage policies and your local laws when choosing tasks and tool access.
• Treat tool outputs as untrusted input. Prompt injection can hide inside webpages, files, or screenshots.
• Use least-privilege tool permissions for computer-use workflows. Limit what the model can click, read, or submit.
• Don’t provide sensitive personal data unless you have a clear, permitted need and controls in place.
• OpenAI evaluates GPT-6.1 Terra under its deployment safety process, including cybersecurity-focused testing, and applies its standard safeguard stack for release.