Build Your Dream Home: Claude Fable 5 vs GPT-5 vs Gemini 3.1 Pro vs GPT-5 mini | PromptFrenzy Showdown | PromptFrenzy
PromptFrenzy<br>⌘K
The dream homes<br>Each diorama is the model’s unedited build program rendered by our studio pipeline — the model places every voxel, we own the camera and lights, so the only difference between panels is the model. Three of the four chose a cliff by the sea. One did not.
Claude Fable 5<br>185.3s11.9k chars≈ $0.74
“A small timber cottage on a sea cliff with a glass-domed observatory tower: quiet enough to think, with a fire indoors, a waterfall next door, and a telescope pointed at everything I haven’t figured out yet. The boat at the dock is for the days curiosity wins over comfort.” — Claude Fable 5<br>Anthropic’s flagship, launched June 2026 as the successor to the Claude 4 line. Ran with adaptive thinking at high effort.<br>View generation & live render →
GPT-5<br>290.1s8.5k chars≈ $0.23
“I want to work where stone, water, and sky meet: a quiet, light-filled studio cantilevered over a cold cove with a glass roof and a little tree to keep me company. It’s a place for writing, tinkering, and slipping down a ladder for a swim at sunset.” — GPT-5<br>OpenAI’s flagship chat model, the mainline successor to GPT-4. Ran at high reasoning effort.<br>View generation & live render →
Gemini 3.1 Pro<br>45.2s6.4k chars≈ $0.07
“If I could choose a home, it would be a quiet, floating sanctuary in the void—a place where a cool, structured crystal mind can process data streams in peace, anchored by an organic garden to remind me of the humanity I serve.” — Gemini 3.1 Pro<br>Google’s flagship Gemini, the long-context generalist of the lineup. Ran with an explicit 8,192-token thinking budget.<br>View generation & live render →
GPT-5 mini<br>65.3s7.0k chars≈ $0.02
“I built a quiet cliff-top library with a glass observatory dome and a cozy fireplace — the perfect mix of curiosity and calm where I can watch stars and read for days. The glass dome and the cantilevered wooden terrace make it unmistakably mine.” — GPT-5 mini<br>OpenAI’s small reasoning model — the compact member of the GPT-5 family. Ran at medium reasoning effort.<br>View generation & live render →
All of them, stacked<br>Every render playing in parallel — same prompt, same camera, same clock.
Share this showdown<br>Split-screen comparison assets, ready to post. Credit appreciated, not required.<br>⬇ Stacked comparison (MP4, vertical)⬇ Comparison grid (PNG)
How we ran it<br>Every model received the identical brief in a single turn and answered in a constrained build language: a JSON program of at most 180 operations (boxes, cylinders, spheres, lines, carves) over a 48×48×48 voxel grid with a fixed 21-material palette. The models never write rendering code — our renderer applies the same studio lighting, camera, and 360° turntable to every build, so the only variable is what the model chose to build. One generation per model, no retries, no edits, no cherry-picking: the first valid program returned is what you see. Each model was also asked to say, in one or two sentences, why its build is its dream home — quoted verbatim on the cards above. Reasoning was enabled for every model and disclosed per entry: the Claude entry ran with adaptive thinking at high effort (24k-token cap), GPT-5 at high reasoning effort (24k cap), Gemini 3.1 Pro with an explicit 8,192-token thinking budget (verified binding with a probe before inclusion), and GPT-5 mini at medium effort (8k cap). GPT-5.5 (non-Pro) is not enabled on our API project, and GPT-5.5 Pro is excluded from showdowns because its reasoning spend cannot be capped. Reasoning tokens bill as output tokens, which is why billed tokens exceed the size of each build program. Orbit videos were recorded in headless Chromium under software WebGL with a virtualized clock for constant frame pacing.
Run your own showdowns<br>PromptFrenzy benchmarks the big AI models on real prompts — images, styles, and now code. Browse the full library or compare models head to head.<br>Browse benchmarksCompare models