Give AI a photo of a tram and ask it to build a 3D model from it. What would it produce?

OpenAI's just-released GPT-6.1 Terra took on exactly that job.

Terra can understand images and screens, and can operate browsers, desktops, and applications, doing tasks for you on your computer.

In the official demo, it took a photo of a tram and then built the model in Blender.

Blender is software for making 3D models.

The body, windows, wheels, and roof were built in sequence, then the whole tram was rotated to show side and back views. The official video skipped part of the modeling process.

What the community has already done with it

The official demo used a tram; community user lindaniel tried a crab.

The modeling record he shared was 5 minutes 38 seconds.

The crab in the image had an orange shell, protruding eyes, two open claws, and articulated legs.

The same author also tried a parrot riding a bike, using Max, and it took about 2 minutes. But he specifically noted that the output was static.

Another community user kevechang used Fast to make a T-Rex riding a motorcycle HTML page. HTML is the language for writing web pages.

The screenshot shows the green dinosaur with short arms holding the handlebars, with hills, clouds, and sun in the background. The author says the original page moves.

Someone else flipped the riding task upside down.

Community user yunhe shared a webpage where the kangaroo, bicycle, road, and trees were all flipped over, making the page look like an Australian postcard viewed upside down.

A 9:13 island; the author still found gaps

Community user FGJ used Terra Max to make a 3D webpage, using Three.js to build a voxel-style island.

Three.js is a tool for web 3D effects. Voxel style means making buildings, trees, and terrain look like they're assembled from small cubes.

He shared a task record of 9 minutes 13 seconds, completing an island with a five-story pagoda, a torii gate, cherry blossom trees, and a stream.

Earlier, with Astra Max, he had done a similar scene in over 40 minutes. This time he praised Terra's speed.

However, he felt the overall quality and details still fell short of Astra. Below is the Astra work he had previously shared.

In this public set, Astra's tower eaves, railings, and island plants were more detailed.

Another community user wanted to build the black hole from Interstellar.

He asked for an accretion disk and gravitational lensing effects, mouse rotation, and side sliders for adjustable parameters.

The accretion disk is the glowing ring of matter around the black hole; gravitational lensing makes light bend.

The author selected Ultra mode in the app, recorded 8 minutes 43 seconds, and made both the work and code public.

But he was not satisfied with this version and felt that the GPT series always has something awkward when creating black hole effects.

Unit price dropped, how much is saved per task?

Terra's price cut applies to the API, the way developers connect the model to their own programs, billed by the amount of content processed.

Token is the unit of measurement for content processed by the model.

Standard prices are $1 per million input tokens, $5 per million output tokens, and $0.05 per million cached input tokens.

Input is the content given to the model; output is the content the model generates.

Compared with the previous GPT-5.6 Terra, standard input price dropped from $2 to $1 per million tokens, and output from $12 to $5.

DesignArena posted a side-by-side challenge: rebuild the Belem Tower at sunset.

It is a stone tower by the river in Lisbon. They asked to keep the rope-style carvings on the stone, open balconies, and the domed tower, plus gulls, sailboats, and a distant bridge.

The result: GPT-6.1 Terra took nearly 4 times longer than GPT-5.6 Terra, and used 48% more tokens.

But according to DesignArena's published bill, the new version cost $0.047 vs $0.077 for the old version, so the new one was about 38% cheaper.

By DesignArena's description, the new version produced layered arched windows, a domed corner tower, and sunset reflection on the water. The old version's tower was squarer and simpler, with no reflection on the water, and the distant bridge was reduced to two pillars.

Independent evaluation organization Artificial Analysis also tested a set of tasks.

In the comprehensive capability test, with max settings, the new Terra scored 6 points higher than the old version, with an average task cost of about $0.40 vs $1.40 for the old version.

The top of the chart shows the composite score, the bottom combines score and cost. Higher up means higher score, further left means lower cost.

Artificial Analysis also compared different reasoning strengths.

Reasoning strength is the thinking budget given to the model. Increasing the budget affects waiting time and cost.

Comparing combined capability and cost, Terra's low and medium tiers sit on the efficiency frontier. Among the tested options, none are both cheaper and higher-scoring.

Artificial Analysis placed models in engineering environments, had them complete tasks with tools, and aggregated a code agent index.

At max tier, Terra scored 48, Sol 54, Luna 31.

OpenAI also explicitly said that for the most complex programming tasks, GPT-6.1 Sol and GPT-6 Astra remain more suitable.

On the cost-efficiency frontier for coding tests, Terra's selected tiers are low and xhigh, differing from the low and medium in the comprehensive test.

As for whether the model fabricates, Artificial Analysis also ran the AA-Omniscience test.

This test checks whether the model keeps guessing when it doesn't know the answer.

The old Terra's hard-guess rate was 88%; the new Terra dropped to 38%. This is a metric from this test, not the error rate for all everyday answers.

Where to use, how to choose

GPT-6.1 Terra is now available in ChatGPT Work and Codex, not in regular ChatGPT for now, and mobile Work is supported.

Plus, Pro, Business, Enterprise, and Edu plans support it; enterprise workspaces depend on admin settings and actual availability.

In ChatGPT desktop or web Work, open Advanced in model settings and select GPT-6.1 Terra.

To have it operate the browser and computer, you need to complete the corresponding setup and grant it access permissions.

If you use the terminal, you can select it via Codex CLI.

Codex CLI is Codex running in the terminal.

Update to version 0.159.0 or later, then start with this command:

codex -m gpt-6.1-terra

Once in an interactive session, switch with /model.

For developers connecting via API, the model name is gpt-6.1-terra, with tool calls going through the Responses API.

It supports 1.05 million tokens of context, with max output of 128,000 tokens.

Reasoning tiers are none, low, medium, high, xhigh, and max, with medium as default.

API calls are billed separately from subscription quotas.

When input exceeds 272K tokens, input and cache prices per request double, and output price multiplies by 1.5.

In closing

In the Belem Tower test, the new version didn't use fewer tokens, but the bill was still lower than the old version.

With the same API budget, you can now call more tokens.

This might change how people give instructions to AI.

If a page was previously limited to one version due to cost, now there's room to try two other approaches.

If a small task wasn't worth spending money to call the model, after the price drop it might get integrated into a program.

Thank you for reading my article.

If you found it helpful, feel free to click like, watch, and share.🙂

If you want to receive updates first, you can give me a star⭐～Thank you for reading.

Open-source knowledge base (real-time updates, chat group): https://my.feishu.cn/wiki/OhQ8wqntFihcI1kWVDlcNdpznFf