GPT 6.1 Terra Brings Visual Reasoning to Everyday Work

Updating a product catalog from a supplier's lookbook sounds routine until several items look almost identical. Someone must inspect each photo, find the right record, change the listing, and explain what happened. GPT 6.1 Terra is designed for work like this, where understanding an image is one step toward completing a task.

OpenAI released the model, officially named GPT-6.1 Terra, on October 26, 2026. The company describes it as a fast, efficient model for visual reasoning, precise computer use, and clear writing. Those are useful priorities for product teams, but the real test is whether Terra can carry a task from the source material through to a correct result.

What Terra can take in and produce

Terra accepts text and images and produces text. It is not a new GPT Image model or a native video generator, and its model page lists audio and video as unsupported. It can call an image-generation tool, but images created by that tool are separate from Terra's own text output.

The official model page gives the core specifications:

The large context window leaves room for substantial reference material in a request. That capacity does not guarantee the model will use every relevant detail correctly. Identify the source images and matching rules explicitly, especially when an upload contains several versions of the same product or document.

From recognizing an image to changing a record

Consider a catalog team evaluating Terra on a small batch of supplier photos. A useful trial would require it to match each photo to an existing item, explain ambiguous matches, and leave uncertain records for review. Include a few near-identical products to see whether it asks for missing evidence or guesses.

Input quality matters here. OpenAI's general vision guidance notes that models can misread small or rotated text and struggle to locate objects precisely. Enlarging a tiny label can improve readability. For the catalog trial, supply a close-up of the item code alongside the full photo, then check the match against the source record.

Acting on a screen requires more than an image upload. In OpenAI's computer-use integration, the developer's application connects the model to a browser or desktop environment. The application executes requested actions and returns screenshots so the model can inspect the result. It must keep that environment running between calls; continuing the model conversation does not restore a lost browser session.

A useful catalog test would therefore reopen each saved listing and check its contents. Clicking Save is only an intermediate step. The final report should identify which records changed and which still need attention. OpenAI also recommends setting limits on a run and keeping users in control of consequential actions, including destructive changes.

Writing that finishes the task

The writing component is easy to overlook when evaluating screen interaction. A task can end with correct changes and an unhelpful explanation. A catalog manager needs to know which items require review without reading a transcript of every click.

OpenAI says Terra favors direct results and plain language. In this example, ask for the updated item IDs, unresolved matches, and the evidence needed to resolve them. Put unresolved items first so the manager can act on them immediately. A polished summary that leaves out an uncertain match would still fail the task.

OpenAI recommends evaluating Terra for writing based on supplied material and computer tasks with a defined outcome. If your team already uses GPT-6.1 Sol, compare the two on identical inputs and reasoning effort. Judge the writing by whether it preserves the relevant facts and suits its audience; brevity alone is a poor measure of a useful handover.

The cost of using a million-token window

As of October 27, 2026, Standard API text-token pricing for requests with up to 272,000 input tokens is $1 per million input tokens and $5 per million output tokens. Cached input costs $0.05 per million tokens. Cache-write rates are $1.25 per million tokens for 30-minute retention and $2.50 for 12-hour retention, according to OpenAI's pricing documentation.

Long requests change the calculation. Above 272,000 input tokens, Terra's model documentation specifies double the input and cache rates and 1.5 times the output rate for the entire request. At Standard rates, that makes uncached input $2 and output $7.50 per million tokens. Budget for the whole request at the higher rates once it crosses the threshold. Image inputs and any tool charges also need to be included in a workflow's total cost.

For API integrations, OpenAI recommends starting at medium reasoning effort and increasing it when better results justify the added time and tokens. Tool calling requires the Responses API. Applications that previously called tools through Chat Completions need to move those requests to Responses when adopting Terra.

Where Terra is available

OpenAI's launch-week product update lists Terra in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu plans. Users can select it under Advanced in the desktop app or ChatGPT Work on the web; Work also supports it on mobile. It is not available in Chat. Enterprise access depends on workspace availability and administrator controls, and UAE inference residency is unsupported.

The catalog example offers a manageable pilot: a small batch, source records, and a clear rule for uncertain matches. Compare the saved listings and the final explanation with a human-checked reference. Choose Terra based on whether it matches the right items, saves correct edits, and flags uncertainty in a report the team can use.