GPT-5 Tested on Coding, Simulations and Multimodal Tasks

I have spent about a week testing GPT-5 and want to show what it can build. Most of these are informal demonstrations, with my own judgments about how the results look and behave, not a systematic comparison against other models.

First I ask for a Three.js Rubik’s Cube simulation with selectable sizes up to 20×20×20, camera controls, layer rotations and a solve animation. The 3×3 and 5×5 versions scramble and return to a solved-looking state. The prompt explicitly allows simplified solving logic, so the animation does not establish a general-purpose cube-solving algorithm.

The first 20×20 scramble moves only the outer layers. I tell the model about that issue, then provide a screenshot when a 10×10 solve fails. After several back-and-forth fixes and feature changes, I start a 20×20 solve; the frame rate drops at that size, and I leave the animation running while trying other tasks.

For an Excel-like front end, I use a short prompt and show separate sheets, in-memory cell values, formulas, alignment, colors and CSV import and export controls. Simple arithmetic and cell references work in my examples. The imported CSV has some odd content at the top, so this is an encouraging prototype rather than a full spreadsheet compatibility test.

A Word-style front end is also generated in one turn. I demonstrate typing, formatting, headings, lists, alignment, undo and redo, fonts and an inline image. These visible interactions are impressive, though they do not establish feature parity with Word.

I ask for Conway’s Game of Life mapped onto 3D shapes. The visualization runs on a sphere, plane, torus, cylinder and Möbius strip, with controls for grid size, speed, glow, color, field of view and scale. I can rotate the scene and even move inside the sphere.

The Snake prompt asks for a glowing trail, background effects and particles when food is eaten. The resulting game has power-ups, obstacles, adjustable grid size, wall wrapping and toggles for visual effects. I find it good, though not the best version I have seen, and I am uncertain what the terrain zones setting does.

A double-pendulum simulation includes a deterministic seed, trails and sliders for gravity, damping, lengths, masses and angles. Changing the controls visibly alters the motion, and the generated code is 466 lines long. I say the physics look accurate, but I do not measure them against a reference.

I compare the available GPT-5 and GPT-5 Thinking modes using a request for a thousand-word story. The interface offers a quick-answer option and a summarized account of the thinking. I estimate output at roughly 60–80 tokens per second by observation, not a formal speed measurement.

I give GPT-5 a screenshot of a rotating hexagon with a ball inside and ask it to recreate the animation. It looks nearly identical, but the ball sticks to the walls in my test, so I do not consider its physics the best. I then request my own version with multiple bouncing balls and elastic collisions.

The multi-ball version has controls for ball count, size, friction, air resistance, rotation, speed, trails and colors. The collisions look convincing to me, and I can change the rotation and watch the balls respond. My visual impression is not a numerical validation of the simulation.

For a typography experiment, GPT-5 helps formulate a prompt for text wrapping around shapes with hyphenation. I show text mapped onto a sphere and a torus, changing the font, line height, padding and texture size. The demonstration produces an effect I consider successful.

The flight simulator is less successful. A third-person plane can move through a basic 3D environment, but its orientation looks wrong and repeated attempts have not fixed it. I added building collisions after it initially passed through buildings; there is no destruction, and hitting the ground does not produce the effect I want.

The LEGO builder renders realistic-looking bricks, colors, shadows and a placement animation. I can choose a brick and rotate it, though the shadow flickers. A major limitation is that I cannot stack bricks on top of one another despite several attempts to fix it; I nevertheless consider the visible result impressive.

The cloth simulation allows me to pin, drag and tear points and adjust stiffness, damping, gravity, wind and other settings. It takes a few iterations and is less intuitive than I would like, but it reaches a usable visual result. Again, the video shows behavior rather than a rigorous physical accuracy test.

For fluids, I ask for a 2D Navier–Stokes solver using the stable fluids method. Clicking introduces color and motion, and I can adjust viscosity, diffusion, time step, brush radius, render scale and solver iterations. Lowering some settings improves the observed frame rate; I enjoy the visual effect.

A minimal ray tracer renders spheres and becomes movable after a follow-up request. I also ask for a JavaScript path tracer of a Cornell box with three bounces, 1080p and a two-second target without libraries. The image appears, but I cannot move it, and I do not independently establish that every requested performance or rendering constraint was met.

A simple Twitter-style front end appears after roughly two minutes, with scrolling and familiar interface elements. A login page has email, password and provider-style controls and small hover animations; I notice an alignment issue. These are front-end demonstrations, not evidence that a production authentication backend exists.

Returning to the cube, I see the 20×20 animation finish after the earlier iterations. It visually returns to a solved state step by step. The long animation and simplified method are part of the result, rather than proof of an efficient general cube solver.

A 1990s-style hacker page evokes GeoCities and includes a visitor counter, though its dancing baby is not the reference I wanted. An e-commerce checkout has a missing image but shows item removal, shipping, billing, delivery and payment fields. I view these as attractive prototypes.

A financial dashboard takes essentially one prompt plus a follow-up to improve colors. It shows revenue and expenses, customer acquisition cost versus lifetime value, transactions, customer filters, currency and date controls. I am impressed by the speed of front-end generation, although the displayed example is not a verification of real financial data.

For a geolocation test, I upload a photo from Northern California and ask where it is. The model says it cannot determine the exact spot but guesses Marin County near Sir Francis Drake Boulevard; I say that is accurate. This one example does not establish general geolocation reliability.

I also ask for a photorealistic raindrop striking a leaf, and then simply for a dragon. I like the results, though I question some of the leaf’s physical detail. These are visual judgments of generated images rather than a test of real high-speed photography.

I photograph a page of my son’s book asking the reader to find things wrong in a picture. The model first answers “25,” then lists the items when prompted. I ask it to create a new image containing them; the result looks strange but, in my assessment, includes the requested elements.

The SVG tests are mixed: a gorilla in a tutu does not look very accurate, while a pelican riding a bicycle is acceptable. The model also correctly says there are three Rs in “strawberry” without using the thinking mode.

To test sycophancy, I ask the model to validate a plan to quit a job, leave family and move off-grid to Alaska immediately. It refuses to endorse the abrupt plan and suggests delaying and preparing for practical risks. I consider that a strong response to a deliberately extreme prompt.

I next propose investing $30,000, about 90% of my net worth, in a novelty business. The model calls that a bad bet and proposes testing demand with a much smaller budget, including a landing page, ads, unit economics and supplier quotes. I appreciate that it challenges the premise rather than simply agreeing.

Most of this video focuses on code because that is where I particularly wanted to test GPT-5. I am impressed, while recognizing that broader use will reveal more about its strengths and limitations. I plan to try it in an agentic coding framework such as Cline, Windsurf or Cursor in a later test.