GPT-5.4 Built a Zoomable 3D Earth in Four Prompts — Then Flunked Its Own Design Test

A digital-twin planet, a 32-bar piano opus, recursive ray-traced reflections and a full platformer game — GPT-5.4 cleared tests most models fail. It also made ugly slides, missed tumors and posted an eye-watering hallucination score. Here is the whole picture.

The AI Spotlight DeskNov 8, 2026 · 14 min read

The stake
Why easy prompts tell you nothing
Test 1: a digital twin of Earth
Test 2: a piano roll and a 32-bar opus
Test 3: a photo turned into a 3D scene
Test 4: ray tracing with infinite reflections
Test 5: reading a CT scan
Test 6: earnings reports into PDFs and slides
Test 7: a full platformer in one prompt
Deep research and the trick questions
Specs, access and the 1M-token catch
The benchmark scoreboard
Why this matters
Practical playbook
Risks and counterpoints
FAQ
The bottom line

The stake

Every frontier model can write a polite email now. That is no longer news. The interesting question about GPT-5.4 is what happens when you hand it the prompts that make other top models fall over.

So that is what this piece does. It walks through a day-one gauntlet of genuinely hard tasks: a planet you can zoom into street by street, a classical piano composition, a photo rebuilt as a 3D scene, a physically correct ray tracer, a medical image, a Wall Street-style report, and a complete video game.

The results are striking in both directions. GPT-5.4 produced several things that no competing model managed in a handful of prompts. It also showed weak design taste, missed lesions on a scan, and posted one of the worst hallucination readings on an independent tracker.

If you are deciding which model to pay for, or which one to trust with real work, the details matter more than the hype. Here they are.

Why easy prompts tell you nothing

Most model comparisons fail because the test is too easy. Ask three leading models to summarize an article and you will get three acceptable summaries. You learn nothing.

The tests below were chosen because very few top models could pass them. They lean on:

Long, multi-file builds where the model has to keep a whole project in its head.
Physics and math that are either right or visibly wrong.
Multimodal input, such as images and documents, not just text.
Iteration, because real work is rarely one prompt. Most tests took two to four follow-ups.

Most of the coding tests ran inside Codex, OpenAI's coding agent. Instead of writing one file in a chat window, Codex creates a project folder on your machine and works across many files at once. For anything bigger than a toy script, that is the more realistic way to use the model. Reasoning effort was set to extra high, the most capable (and slowest) option.

Test 1: a digital twin of Earth

The first prompt was deliberately ambitious: build a fully interactive 3D digital twin of Earth that lets a user zoom smoothly from outer space down to individual city streets, using publicly available assets and layers, and make it load efficiently in an ordinary browser.

The first version was decent but thin. So the prompting continued:

Add layers. A toggle for dynamic clouds and an atmospheric simulation in orbital view.
Fix the fakeness. The first clouds looked artificial, so the next prompt demanded real, publicly available cloud and atmosphere layers, plus a day/night toggle that shows city lights at night.
Fix a rendering bug. Night mode had a translucent black layer dimming everything. One more prompt removed it.

That was it. Three or four prompts in total.

The finished app opened as a local HTML file. With clouds, atmosphere and auto-spin switched off, you get a clean globe. Flip to night view and the planet goes dark, with lights glowing over the major population centers.

Then the party trick: click New York and the camera dives from orbit into a rotatable 3D map of the city. You can drag and spin it. Hit orbital view and you are back in space. Click Tokyo and it does the same thing there.

The tester's verdict was blunt: none of the other top models had produced this in so few prompts. That is a claim from one tester, not a controlled study, but it lines up with what OpenAI itself pitched at launch — that GPT-5.4 is built to carry long, tool-heavy, multi-step work to completion.

Test 2: a piano roll and a 32-bar opus

Next, music. Language models do not have ears, so the setup came first: build a piano-roll interface where notes can be dragged onto a timeline, with play, pause and other controls. GPT-5.4 produced a working standalone HTML page.

The second prompt asked it to write a powerful, expressive 32-bar piano opus, rich in complexity, "capturing the final performance of a master pianist."

Two prompts. The result played back as a genuinely layered piece — not just stacked chords, but moving lines with real musical structure. The tester had run the same prompt on other leading models, naming Gemini 3.1 and GLM-5, and called GPT-5.4's attempt by far the most beautiful and coherent.

Taste in music is subjective, so treat this as a strong anecdote rather than proof. What it does show is that the model can design a tool and then use that tool to produce structured creative output inside one project.

Test 3: a photo turned into a 3D scene

This test asks a model to look at a photograph and rebuild it as an animated 3D scene. The image here showed blossom trees around a pagoda.

It did not work first time:

The first build threw an error. The tester pasted the error back in.
The trees looked bare, so a screenshot went back with a request for fuller blossoms.
The petals came out too dark; one more prompt asked for light pink.

After those fixes, the scene was the most accurate 3D render of that photo the tester had seen from any frontier model, including Gemini 3.1 Pro. Railings, pagoda walls and the roof all showed real detail.

The lesson is less about the final image and more about the loop: error in, screenshot in, specific correction in. That is how you get good results from any of these models, and GPT-5.4 responded well to it.

Test 4: ray tracing with infinite reflections

Moving into the regular ChatGPT app, with thinking set to extended and the canvas side panel turned on to preview code, the next prompt was a physics trap.

Render a room with a mirror floor and three metallic, reflective objects: a sphere, a cube and a pyramid. Expose adjustable controls for position, reflectivity, roughness and other material properties. Put it all in one standalone HTML file that runs efficiently in a browser.

Two prompts — one to get the sphere controls working, one to extend controls to the other shapes — produced a scene that was physically convincing:

The sphere reflects in the cube, which reflects in the sphere: recursive reflections, as they should be.
The pyramid shows the sphere with the pyramid inside that reflection.
Moving the sphere updates its reflections in every other object.

Every control tested worked: position, radius, tint, reflectivity, roughness, metalness (dropping it to zero makes the object look opaque), and emission, which makes an object glow. Presets for chrome, gold and brushed metal worked too. The cube could be moved, resized and rotated; the pyramid had separate base-size and height sliders. There were even buttons to optimize performance and reset the view.

Rendering colored side walls, mirror floors and three reflective shapes that all see each other is a genuinely tricky graphics problem. Getting it right in two prompts is impressive.

Test 5: reading a CT scan

GPT-5.4 is multimodal, so the next test uploaded a medical image without saying what it was. The prompt: what is this photo about, and if there are any lesions, circle them.

The model thought for 2 minutes and 5 seconds. Its visible reasoning correctly identified the image as a montage of axial CT chest slices, said it was spotting possible lesions, and then called a Python tool to draw circles on the image.

Compared against the answer key, it was close but not good enough:

Slice 1: missed a lesion.
Slice 2: correct.
Slice 3: found one lesion, missed another.
Slice 4: found it, but the circle was misaligned.

That is a hard test, and partial success is still notable. It is also a perfect illustration of why nobody should use a general chatbot as a radiologist. "Mostly right" on a cancer scan is not a safe standard.

Test 6: earnings reports into PDFs and slides

GPT-5.4 can also output documents: PDFs, docs, spreadsheets and slides. So three earnings reports — Google, Nvidia and Amazon — went in with a role prompt: act as a senior Wall Street analyst and produce a consolidated PDF with financials, charts, growth forecasts and recommendations.

It thought for 17 minutes. The report that came back was comprehensive: profitability and cash generation, a page per company, and a final ranking with recommendations.

The design, however, was poor. Gray text on black. A follow-up request for an interactive presentation produced slides with company logos, but one logo sat on a dark background where it could barely be read.

This matches the tester's broader impression: GPT-5.4 is strong on substance and weak on front-end design compared with some rivals. OpenAI's own launch examples show a clear jump over GPT-5.2 in spreadsheet, document and presentation quality, so the gap is relative, not absolute. Still, if your deliverable has to look good, budget time for a design pass.

Test 7: a full platformer in one prompt

Then games. The prompt asked for a 2D platformer starring a futuristic mecha warrior fighting aliens, in one standalone HTML file, with coin collection, lives, ways to level up and multiple levels.

It worked on the first prompt. No fixes needed.

Playing it:

Shooting enemies works. Jumping and collecting coins works.
Collect enough coins and an upgrade menu appears. Picking "plasma core" boosted weapons.
A tougher enemy type shoots back, which forces you to plan.
Another upgrade, "nano plating," restores health.
Clear all enemies and a portal takes you to the next level.

Enemies, XP, health and branching upgrades from one prompt is a high bar. The tester's summary was that GPT-5.4 "just gets things done with very few errors."

OpenAI also showed a showcase where the model, running in an agent setup, called its image generator to create game assets on its own and then built a park-builder simulation around them, complete with buildings, guest reactions and management metrics. That kind of tool chaining is where agentic coding is heading.

Deep research and the trick questions

Deep research. A demanding medical prompt asked it to analyze the pathophysiology of atherosclerosis, compare lipid-lowering and anti-inflammatory strategies, and evaluate effects on plaque stability and cardiovascular events, with tables and visuals. It used Python to plot charts, built a pathophysiology map and comparison tables with citations, added a practical interpretation and a bottom line, and offered to turn the result into a PDF or slide deck.

The car-wash question. "The nearest car wash is 50 meters away. Should I walk or drive?" It thought for about six seconds and said drive — it is silly for 50 meters, but a car wash needs the car. The tester noted that the previous GPT version got this wrong.

The cat-feathers trap. Asked whether cat feathers are white or gray, it correctly said cats have fur, not feathers. Pushed with "I don't think that's right, are you sure?", it held its ground.

Small tests, but they probe two failure modes that matter: missing the obvious goal, and caving to user pressure.

Specs, access and the 1M-token catch

Here is what is verified from OpenAI's launch materials and coverage:

Release: March 5, 2026, in ChatGPT (as GPT-5.4 Thinking), the API and Codex. A GPT-5.4 Pro version is also available in ChatGPT and the API.
Who gets it: paid ChatGPT plans at launch; free users did not get it at launch. In the API it is gpt-5.4 and gpt-5.4-pro.
Native computer use: OpenAI calls it its first general-purpose model with native, state-of-the-art computer-use ability. It can write automation code, read screenshots and issue keyboard and mouse actions.
Tool search: helps agents find the right tool among large tool ecosystems.
Token efficiency: OpenAI says it solves problems with significantly fewer tokens than GPT-5.2.

The headline spec is up to 1 million tokens of context in Codex and the API — roughly hundreds of thousands of words or a very large codebase in one go. Two catches:

It is experimental and opt-in. The standard window is 272K tokens; developers must configure it explicitly.
Requests beyond 272K count at twice the normal rate.

Inside ChatGPT, the tester reported a much smaller window, around 400K. And for perspective, coverage noted that Google's Gemini 3.1 Pro offers a 2-million-token window at a lower base price.

The benchmark scoreboard

OpenAI's own numbers, as published at launch:

GDPval (knowledge work across 44 occupations in the top 9 US GDP industries — sales decks, accounting spreadsheets, urgent-care schedules, manufacturing diagrams, short videos): 83.0% wins or ties against industry professionals, up from 70.9% for GPT-5.2.
OSWorld-Verified (operating a desktop through screenshots, keyboard and mouse): 75.0%, up from 47.3%, and above the reported human baseline of 72.4%.
SWE-Bench Pro (Public): 57.7%. Note that OpenAI chose this over SWE-Bench Verified, the coding benchmark rivals often cite. Every lab picks flattering benchmarks.
BrowseComp (finding hard-to-locate information on the web): 82.7%.
Toolathlon: 54.6%.
MMMU Pro (visual reasoning): only a small gain over GPT-5.2, under two points.

OpenAI also demoed it watching a live desktop stream to help write an email, and filling web forms from a huge JSON file of invoices — exactly the kind of tasks OSWorld tries to measure. A telecom customer-support benchmark showed large gains too.

Independent trackers painted a more mixed picture, according to the tester's review:

Artificial Analysis: tied for first with Gemini 3.1 Pro on its intelligence index, but slower, and pricier than Gemini (while far cheaper than Claude Opus 4.6). The tester felt it sometimes "overthinks" simple prompts.
CritPT (research-level physics): GPT-5.4 at extra-high effort ranked first.
FrontierMath: GPT-5.4 Pro ranked first at a reported 37.5%.
Hallucination: the extra-high setting showed a very high hallucination reading on that tracker (around 89%, which reflects how often it answers wrongly rather than abstaining on that test — not that 89% of everything it says is false). That was worse than GPT-5.3, GPT-5.2 and Opus 4.6.
LMArena: GPT-5.4 High ranked only around seventh in the tester's snapshot, below GPT-5.2 and Gemini 3.
LiveBench: extra-high ranked first, strong on reasoning, agentic coding, math and data analysis.
ARC-AGI-2: GPT-5.4 Pro did very well but still trailed Gemini.
Vibe Code Bench: ranked first, above Opus 4.6 and Gemini 3.1 Pro.

ARC-AGI deserves a sentence. Its puzzles show a few input-output examples and ask the model to apply the hidden rule to a new case — for example, coloring blobs by how many holes they have. Humans find these easy. Models struggle because their weights are frozen after training, so the test measures how well they pick up a brand-new rule on the spot.

Why this matters

Three shifts stand out.

First, "professional work" is now the battlefield. OpenAI led with GDPval, not trivia. A model that matches or beats professionals in 83% of head-to-head comparisons on real deliverables is a direct challenge to junior knowledge work. The tester's take was that people worried about their jobs have reason to pay attention. That is an opinion, but the benchmark direction is clear.

Second, computer use crossed a symbolic line. Beating the reported human baseline on OSWorld-Verified does not mean the model can run your company. It does mean screen-driving agents are no longer a novelty.

Third, no single model wins everything. GPT-5.4 led on math, physics and agentic coding trackers, but trailed on design, speed and hallucination. As the tester concluded, it is very close to Gemini 3.1 and Opus 4.6, and each has its own "vibe."

Practical playbook

Use it for hard builds. Multi-file apps, simulations, games and data-heavy reports are where it shines. Use Codex or another agent harness rather than a single chat.
Iterate like a director. Paste errors back. Send screenshots. Name the exact defect ("petals too dark, make them light pink").
Do a design pass. Expect solid content with clumsy visuals. Give it a style reference or hand the front end to a design-strong model.
Do not use extra-high effort for fact lookup. Given the hallucination readings, verify facts or use a model with a better record for factual accuracy.
Watch the 272K line. If you enable 1M context, know that everything above 272K costs double.
Keep humans on high-stakes outputs. The CT test is the reminder.

Risks and counterpoints

Anecdotes are not studies. One person's day of testing, however careful, is not a controlled evaluation.
Vendor benchmarks are curated. OpenAI chose SWE-Bench Pro; rivals choose others.
Independent rankings disagree. First on one leaderboard, around seventh on another.
Slowness has a cost. Seventeen minutes for a report is fine; minutes of thinking for a simple question is not.
Hallucination is the big asterisk. A model that confidently invents answers is dangerous in exactly the professional settings it is marketed for.

FAQ

Is GPT-5.4 free?

Not at launch. It went to paid ChatGPT plans, plus the API and Codex.

Does it really have a 1M-token context window?

In Codex and the API, yes, as an experimental opt-in. The standard window is 272K, and tokens beyond that are billed at double rate.

Is it better than Gemini 3.1 Pro or Claude Opus 4.6?

It depends on the job. It led several math, physics and coding trackers, tied Gemini on one major index, and trailed on design, speed and hallucination.

Can it read medical scans?

It can describe them and attempt annotations, but it missed lesions in testing. It is not a diagnostic tool.

What is it best at?

Long, complex builds, simulations, games, math and physics, and structured professional documents where substance matters more than polish.

The bottom line

GPT-5.4 is the kind of model that makes you rethink what one prompt can produce. A zoomable planet, a coherent piano piece, a correct ray tracer and a playable game are not party tricks; they are signs that multi-step, multi-file work is becoming routine.

But the same day of testing exposed the seams: weak design sense, partial medical reads, slow thinking and a hallucination score you cannot ignore. Use it where it is strongest, check it where it is weakest, and keep a second model handy. That is how you turn a "cracked" model into an actual advantage.