GPT-6.1 Terra

What GPT-6.1 Terra is for

This model sits where a product needs to look at something and then write about it. Screenshots, interface states, diagrams and photographed pages can arrive in the same request as instructions, and the reply is text. Teams building internal tools or customer-facing assistants use it to replace a two-step pipeline with one call.

The model is a reasonable fit for products whose core loop is inspect-then-describe: an assistant that explains a chart, a support tool that reads what a customer pasted, a workflow that turns a design mockup into written specifications. Each of those needs the image and the instruction in one place, and each expects the answer as text that the surrounding application can render or forward.

Adoption is usually a question of surface area rather than novelty. One model that handles visual input, tool calls and writing means fewer integrations for a team to maintain, and the OpenAI-compatible API at RealRelay.ai lets existing client code carry over. Products end up with fewer seams to mind when a single feature touches a screenshot, a function and a paragraph.

Pricing

Input is priced from $1 per 1M tokens, covering both the text and the images a request carries, and output is priced from $5 per 1M tokens. Because the model is priced dynamically, the figure that applies to a particular request can sit above that floor, so planning should treat the published rate as a starting point.

Cached input reads are priced from $0.05 per 1M tokens, which matters for products that resend a large, stable prefix such as standing instructions, reference material or a long document that stays the same across requests. How a prompt is organised therefore shows up on the bill: the part that never changes can be billed at the cache read rate.

Splitting a forecast by workload is the practical approach. A request that submits a screenshot and asks for a short reply draws on the input rate for the image and text and on the output rate for the reply, with both applying to the same call. Teams tend to watch output volume, since each generated token is billed at the output rate.

Where teams use it

Support and operations teams use the model where a person would otherwise retype what is on a screen. A pasted screenshot of an error dialog, a customer's photograph of a receipt or a capture of an internal dashboard can be submitted with a question, and the written answer can be shown to the agent or sent onward.

Developer-facing products get a different benefit: the large context window lets a request carry a whole repository or a long set of documents alongside the instruction, and structured output keeps the reply parseable. Teams wire this into documentation helpers or specification drafts, where the input is big and the answer short.

For teams and commercial products, RealRelay issues API keys per member or project, each with its own quota and model allowlist.

Frequently asked questions

What is GPT-6.1 Terra?

GPT-6.1 Terra is an OpenAI model for visual reasoning, precise computer use and clear, natural writing. It accepts images and text in one request and returns text, so a product can send a screenshot, a diagram or a photographed page together with an instruction. API requests address it as openai/gpt-6.1-terra.

How is GPT-6.1 Terra priced?

Input is priced from $1 per 1M tokens and output from $5 per 1M tokens, with cached input reads starting at $0.05 per 1M tokens. The model is priced dynamically, so the rate applied to a given request may sit above the published floor. Teams plan capacity around those per-token rates.

How large is the context window of GPT-6.1 Terra?

The context window is 1,050,000 tokens, and a single response can extend to 128,000 output tokens. In practice a request can carry a whole repository, a long set of documents or a large batch of screenshots and still leave room for written output, so substantial reference material can stay in one call.

Which inputs and outputs does GPT-6.1 Terra support?

Requests can carry images and text, and responses are text only. That covers screenshots, interface captures, charts, diagrams and photographed pages alongside ordinary written prompts. Because output is plain text, results are simple to render in an interface, store in a database or hand to another service in the same pipeline.

Does GPT-6.1 Terra support tool calling and structured output?

Yes. A request can supply functions for the model to call, and responses can be constrained to a schema you define. Tool calling suits products that need the model to reach their own lookup or action endpoints, and a schema constraint keeps field names and types stable for downstream code.

How do I call GPT-6.1 Terra through the OpenAI-compatible API?

Send a chat request to RealRelay's OpenAI-compatible endpoint with the model ID openai/gpt-6.1-terra and content parts for any images you attach. Existing OpenAI client libraries work unchanged apart from the base URL and key. Reasoning, structured output and tool calls are all requested in that same call.

Can teams use GPT-6.1 Terra in production through RealRelay?

Yes. RealRelay serves models through model vendors' own APIs and the model services of cloud platforms, under RealRelay's own accounts. Teams can issue API keys to authorized members or projects, each with its own quota, model allowlist, expiry date and IP allowlist.