---
title: "How Good Is GPT-6.1 Terra? Coding and Computer Use Tests"
url: "https://goremy.ai/blog/articles/gpt-6-1-terra-hands-on-test-results"
author: Luis Chavez-Mattos
editor: Luis Chavez-Mattos
editor_title: Director, Product
published: "2026-10-28T00:00:10.103Z"
updated: "2026-10-28T00:00:10.103Z"
topic: AI Model Reviews & Comparisons
description: "Hands-on GPT-6.1 Terra results: build times for a browser OS, skate game, subway FPS and wrestling game, plus a robot arm run, pricing and quirks."
---

# How Good Is GPT-6.1 Terra? Coding and Computer Use Tests

> Hands-on GPT-6.1 Terra results: build times for a browser OS, skate game, subway FPS and wrestling game, plus a robot arm run, pricing and quirks.

Edited by Luis Chavez-Mattos, Director, Product · Published 2026-10-28

## How good is GPT-6.1 Terra in practice for coding and computer use?

GPT-6.1 Terra is very fast and competent on simple-to-moderate builds, but it is not the most thorough model. In hands-on tests, it produced working games and a browser OS in roughly 2.5 to 15 minutes, with the quickest runs using fast and ultra modes. Weak spots showed up too: inverted controls in two separate 3D projects, shallow detail, and a robot arm task it abandoned early.

## TL;DR

- Terra's biggest strength is **raw speed**: a C++ skate game took 2 minutes 35 seconds in fast and ultra mode.
- First-pass results tend to be **competent but simple**, and one or two follow-up polish prompts usually fix the obvious problems.
- **Inverted controls** appeared in both the skate game and the wrestling game, so check direction and orientation first.
- On a robot arm task, Terra **gave up early** and needed coaxing, while GPT-6 Astra reportedly completed the same test impressively.
- Pricing is **$1 per million input tokens and $5 per million output tokens**, with a context window a bit over a million tokens.
- On the coding chart shown in the announcement, **GPT-6.1 Soul still looks stronger**, and Terra is the cheaper option.

## What are the GPT-6.1 Terra specs and price?

OpenAI's model card lists text and image input, so Terra can see images. The details a builder cares about:

| Spec | Value |
|---|---|
| Input | Text and image |
| Context length | A bit over 1 million tokens |
| Max output | 128K tokens |
| Knowledge cutoff | June 1, 2026 |
| Default reasoning effort | Medium |
| Price | $1 per million input, $5 per million output |

The reviewer positioned it relative to two siblings. Terra is not as cheap as Luna, but it appears noticeably more capable. Luna did not even appear on the comparison chart. On the coding chart, Terra scored 47 at low effort and 50.3% at max effort, and Soul still looked like the better coding model. The benchmark itself wasn't named in the coverage, so treat those numbers as a rough relative signal, not a verdict.

The announcement post is short. Most of its examples focus on computer use and agentic actions, with some visual and coding examples. It also includes a chart of factual error rate on difficult prompts, where lower is better.

## How fast is GPT-6.1 Terra?

Fast enough to change how you iterate. Speed depends on the mode:

- **Effort level.** Most tests ran at max effort, even though medium is the default.
- **Fast mode.** Trades nothing visible in the tests for much shorter waits.
- **Ultra mode.** Spawns sub agents that split a job into parts.

Here is how the timed runs compared.

| Test | Settings | Time |
|---|---|---|
| C++ skate game, first pass | Fast, ultra | 2 min 35 s |
| C++ skate game, first pass | Max effort, no fast mode | 7 min 41 s |
| C++ skate game, after polish pass | Max effort | About 12.5 min total |
| Browser OS | Max effort, fast | 14 min 30 s |
| Subway FPS, first pass | Max effort | 9 min 30 s |
| Subway FPS, polish pass | Ultra, fast | 6 min |
| 1990s wrestling game, first pass | Ultra, fast | 4 min 19 s |
| Wrestling game, inversion fix and upgrade | Not stated | 6 min 20 s |

The wrestling run stood out because the model used Blender and Godot to build the models and the game. The reviewer noticed an open Godot project where the model was playing its own build. That is an agentic loop: build, run, check.

## What did Terra actually build?

### Browser OS

The prompt asked for a browser-based operating system with a second game the model would enjoy playing. The reviewer removed the email client from the usual prompt and added a messages app with five recent messages the model had supposedly sent to other AI models.

The result was calm and pastel. It included a GTA-style game ("A city with no bedtime") that ran laggy and had flipped normals on the car. The second game, Grow a Little Galaxy, had you collect five glowing seeds and plant them on a center island. It beat the GTA clone and avoided the usual asteroid shooter. The messages app had gentle, on-theme prompts addressed to other models. A "save capsule" special feature saved and restored state correctly. There was no right-click menu, and the wallpaper motion was subtle at best.

### Skate game

The first pass was simple but had the foundations: a replay feature and a competent look. The reviewer said it felt like a distilled version of a larger GPT model. The controls were inverted (pressing A moved the wrong way), and the board's tail and nose were swapped. Pedestrians were not interactive.

A follow-up asked for a full polish pass and the inversion fixes. The second version had correct controls, lean effects, walking pedestrians, a pump effect, scaffolding, sound effects, a bigger block, and a face on the skater. Pedestrians still couldn't be bumped, and the pigeon didn't fly away.

### Subway FPS

This Three.js game has a train that arrives, the player boards it, and later waves follow. The first pass lacked right-click zoom and bullet holes, had muted sound, and locked movement during boarding. The polish pass in ultra and fast mode added zoom, ejected shell casings, better enemy models, a nicer boarding effect and a harder reload sequence. Bullet holes were still missing.

### Wrestling game

The first version had the characters facing backward, though the model had done simple rigging and included moves like a power slam. A follow-up fixed the inversion and made the style more aggressive and cartoonish, which the model delivered.

## How does Terra handle computer use?

Mixed. The announcement leads with computer use, but the robot arm test was the weakest result in the batch. Terra first said the camera angle wasn't good enough and asked for a fix. The tester adjusted the camera and even had the model show a photo of what it saw. Terra still said it couldn't see well enough. It then argued that the servo mapping looked wrong and declined to proceed. After encouragement it tried some more, then stopped, saying it lacked the right visuals.

The tester described it as hesitant, and noted another GPT-6 or 6.1 model showed similar behavior. GPT-6 Astra, by contrast, completed this same task in one of the most impressive ways the tester had seen. The takeaway is not that Terra can't do computer use. It's that it may quit early on ambiguous physical or visual tasks, and you may need to push it.

## What are Terra's weaknesses?

- **Inverted orientation.** Two projects shipped with backward controls or characters. Add an explicit check to your prompt or plan a follow-up.
- **Shallow detail.** The tester contrasted it with Claude models, which tend to add extra touches like pedestrian collisions. Terra does the core job and stops.
- **Early surrender.** The robot arm run showed it declining a task it could have attempted.
- **Missing small features.** Bullet holes, right-click menus and pedestrian interaction were all absent.

For context, the reviewer noted that results like these would have looked like a frontier jump a year ago.

## Is GPT-6.1 Terra worth using?

For rapid prototyping, yes. A first draft in a few minutes, followed by one or two polish passes, is a good workflow, and the $1/$5 pricing keeps those loops affordable. For tasks that demand depth, careful physical-world reasoning, or the best available code quality, the evidence here points to Soul or Astra instead. Terra looks like a strong middle option: faster and cheaper than the top tier, more capable than Luna.

## Frequently Asked Questions

### What do fast mode and ultra mode do in GPT-6.1 Terra?

Fast mode shortens run times. Ultra mode lets the model spawn sub agents to handle different parts of a job. In the tests, the combination produced a skate game in 2 minutes 35 seconds and a game polish pass in about 6 minutes.

### Which reasoning effort should I use?

Medium is the default. The tester ran most builds at max effort to see the model's ceiling. The coding chart shown scored 47 at low effort and 50.3% at max, so the gain from higher effort looked modest on that measure.

### Does GPT-6.1 Terra accept images?

Yes. The model card lists text and image input. Output is text, with a max of 128K tokens.

### Is Terra better than GPT-6.1 Soul at coding?

Not according to the chart discussed. Soul still looked like the stronger coding model, while Terra is cheaper and quicker.

### Why do Terra's games keep having inverted controls?

The sources don't explain it. The pattern appeared in both the skate and wrestling builds, and each was fixed with a single follow-up prompt, so ask for a direction check up front.