GPT‑6.1 Terra: what’s new in ChatGPT Work and Codex

Where GPT‑6.1 Terra is available

OpenAI announced the launch on October 26, 2026: ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu, and for developers an API named gpt-6.1-terra. This is the list in the announcement, not a promise of identical features in every interface. Free is not on the list. [1]

If you simply want to try the new model, first check the model choices available in your Work or Codex. You do not need to start by building an API integration. Subscriptions and API usage have separate billing; the amounts below apply to the API. [6]

What improved — and how to read the numbers

The update focuses on visual tasks, interfaces, writing and efficiency. Standard Terra API rates are 50% lower than GPT‑5.6 Terra and GPT‑6.1 Sol. [1]

These numbers apply to specified sets of tasks. They do not mean “16.7% errors in everyday ChatGPT” or “half of all software is finished.” Results and conditions are in the OpenAI announcement [1].

A separate system card clarifies that evaluation took place in a research environment or through the API; product behavior may differ because of tools and system instructions. Applying the results to your own project requires testing. [4]

Seeing a screen and working in an app are two steps

The model can analyze an image, but taking action requires an environment with appropriate tools. In a computer-use scenario, an application supplies the screen state, executes the model’s permitted actions and returns a new image for inspection. Uploading a screenshot alone does not give access to your computer. [5]

For example, you can ask it to find clipped text in a screenshot of a mobile page. To fix the page, the agent also needs the project files and a way to run it. To verify the result, inspect it again after the change. A screenshot, code changes and a user-flow check complement one another.

The same distinction applies to video: the API model card lists text output and image input, but does not support native audio or video modalities. Using tools in a video editor is not the same as Terra directly generating a video file. [2] For the creative process, see our separate guide to making a film or animation with AI.

Three small tasks to test for yourself

These are our starting exercises, not measured Terra demonstrations. Begin with a copy of your project and one outcome. If you are new to the approach, first read what vibe coding is.

1. Find a usability issue on a mobile page

Add your own screenshot and relevant files. Afterwards, check that the main button is accessible on a narrow screen and leads to the intended destination.

This screenshot shows my website on a phone. Find one problem that prevents the main action. First state what you observed and the criterion for a fix. Then make the smallest change to the supplied files. Check the page at widths of 320 and 390 px and test the button’s destination. Show the result and clearly state what you could not verify.

2. Check one game mechanic

For a small game, checking restart is more useful than immediately asking for ten new levels. Find more steps in the game creation library.

In this browser game, test the cycle: start → defeat → restart. After restarting, the score and state should return to their initial values and the controls should work. Reproduce any bug and fix only that bug. Do not add mechanics. Finally, distinguish code review from scenarios you actually ran.

3. Turn a document into a verifiable draft

Use your own de-identified document. Compare each row of the result with the original; a missing value should remain missing rather than become a plausible invention.

From the supplied de-identified document, prepare a table with field, value, source location and uncertainty. Do not guess unclear text. Mark conflicting and missing values separately. Return a draft for human review; do not submit external forms.

Download all three tasks as TXT ↓

How much GPT‑6.1 Terra costs via the API

Standard price per 1 million tokens according to the model card [2]:

Example calculation: 10,000 uncached input and 2,000 billed output tokens — $0.01 + $0.01 = $0,02. This is an arithmetic example for a standard short request, not the price of a finished project. Separate tools, cache writes, service tier and long context can change the bill. For input over 272,000 tokens the card lists higher rates for the entire request. [2]

Terra, Sol or Astra: where to start

The official guide positions Terra for fast visual and practical work, GPT‑6.1 Sol for complex work balancing capabilities and cost, and GPT‑6 Astra for the most demanding tasks. Terra’s novelty alone does not make it the best model for every complex fix or research task. [3]

For a first comparison, take three familiar tasks with known answers, provide the same materials and record the result, time, number of retries and actual cost. Where possible, evaluate the result without seeing the model name. This small table is more useful for your decision than someone else’s successful screenshot without the original task.

OpenAI reports improved resistance to malicious third-party instructions, but system-card evaluations do not guarantee protection for every agent. [4] For screen-based work, official documentation recommends an isolated environment, restricted access and verification of saved results. [5] For your own scenario, create an agent permissions map.

Sources and verification method

This article was prepared from primary sources. Test figures are OpenAI’s; we did not run an independent comparative benchmark or a paid API session for this article. The three tasks above are editorial examples for you to try. The Mia cover is an AI-assisted illustration, not an OpenAI screenshot.

1. OpenAI: GPT‑6.1 Terra announcement — launch, availability and comparative evaluations.
2. OpenAI API model card — pricing, modalities and long-context conditions.
3. Using GPT‑6 — model selection, reasoning effort and migration.
4. GPT‑6.1 Terra system card — methodology and limitations of safety evaluations.
5. Computer use documentation — environment, actions and result verification.
6. ChatGPT and API billing — separate billing systems.