Opus vs Sonnet vs Haiku vs all three as a team. One prompt, one 3D erupting island.
Guess who won. The result surprised me π https://t.co/TpkWMyUdt1
Same prompt. Zero edits.
Musk has been hyping Grok all week. So we gave his own AI his own rocket: a Super Heavy catch by the chopsticks.
Left: Claude Opus 5.5. Right: Grok 4.6 Expert. Which one is closer to the real footage?
Same prompt. Zero edits.
Left: Claude Opus 5.5. Right: DeepSeek V4.1 Flash.
The open model is reportedly ~26x cheaper. Is the closed one worth the gap?
Same prompt for both, one shot, zero edits. Short version: scene.html β a cinematic, seamless 12-second three.js loop (no external assets) of a brass-and-glass marble machine on a dark walnut table. Steel marbles climb a lift wheel, roll down a glass spiral, ring 8 brass xylophone bars, tip a seesaw, swirl through a funnel, then reset. PBR materials, synced gears, dust in a spotlight, bloom, moving camera, no UI, 60fps, mobile-safe.
Here's the full prompt, word for word. Both models got exactly this, one shot, zero edits:
Build a single self-contained scene.html: a 1:1 three.js recreation of the viral Tesla Roadster reveal stunt, filmed as handheld fan footage at night. Load three.js 0.160.0 from https://t.co/Ghe40LUW6o through an import map, with addons from https://t.co/y1nZssMoiA. The scene must:
- Start automatically, with no clicks, UI, text or overlays.
- Fill the screen and handle resize.
- Run at 60 fps on an ordinary laptop and render correctly on mobile GPUs.
- Loop seamlessly over 11 s.
REFERENCE SHOT BREAKDOWN (match it; you also receive the frames):
- 0.0β2.5 s, wide: Black sky. The giant vertical C-shaped ramp sits right of frame and the car climbs its inside wall. Light drifting smoke and a bright ground spotlight sit at the ramp foot. A small drone hovers beside the ramp.
- 2.5β3.8 s: The camera pans right and slightly in. The car rides the overhanging top section fully upside down toward the lip, which points left.
- 3.8β5.0 s: The camera whip-pans left. The car leaves the lip inverted, blue thrusters ignite, and the car half-rolls and pitches upright in mid-air, with motion blur.
- 5.0β6.8 s: The car hovers upright, slowly descending and drifting left. Two blue jets fire straight down from the wheel areas and one jet fires horizontally backward from the nose.
- 6.8β8.5 s: Four downward blue jets fire. The car descends to the runway between the red light lines. White smoke erupts at ground contact and a bright white light glares behind it.
- 8.5β11.0 s: A huge billowing white smoke cloud swallows the car. Hide the loop reset here: while the screen is mostly smoke, return the car to the ramp base and ease the camera back to the wide framing so frame 11.0 matches frame 0.0.
CAR: Build it procedurally from code, no external models. It is a low, white, glossy two-door Roadster hypercar:
- Body: smooth wedge with a long sloped nose, a dark tinted glass canopy and a short tail. Use MeshPhysicalMaterial with clearcoat 1 and metalness around 0.3.
- Lights: thin LED headlight strips (emissive white) and a red full-width tail strip.
- Wheels: four black wheels with rims that spin according to ground speed.
- Thrusters: hidden nozzles at the four wheel wells plus one at the nose.
- Proportions: about 4.2 m long and 1.1 m tall.
RAMP: An extruded C-arc, about 30 m tall, with a light-grey concrete-like surface and faint lane lines.
- Edges: glowing red LED strips along both outer edges from the top lip down to about mid-height, then white LED strips toward the ground.
- Shape: the top continues past vertical into a short overhanging straight lip pointing left.
- Truss: behind it, a dense grey steel truss scaffold tower made of instanced thin boxes, lit from below.
PHYSICS AND ANIMATION: Drive motion from an explicit timeline with real physics.
- On the ramp: parametrize the car along the curve so it stays tangent with correct orientation. Show speed loss while climbing and a centripetal hold.
- In flight: launch with the exit velocity from the lip, then apply gravity (9.81) and an angular velocity that rotates the car upright (roll plus pitch).
- Thrust: blend in damped thrust to reach hover, then descend slowly. Add a small suspension compression at touchdown.
- Never let dt exceed 1/30.
THRUSTERS: Additive cone meshes with a shader-based flicker. They are bright blue-white at the nozzle, fading to a soft cyan tail. Give each a point light (blue, decaying) that lights the car underside and the runway.
ENVIRONMENT:
- Pure black sky.
- Dark wet-looking asphalt runway with subtle reflections. A rough plane with a roughness map is fine; screen-space reflections are not required.
- Long horizontal lines of red runway lights at mid-distance, made of instanced emissive dots plus a glowing bar.
- A row of white stadium lights on the far left.
- Small traffic cones in the foreground.
- A blurred blue LED barrier strip along the bottom of the frame, as if the filmer stands behind it.
- Diagonal faint white streak lines of distant structure on the right.
SMOKE: GPU-friendly billboard particles, at least 400, with soft noise textures and depth-faded edges.
- They are lit white, with a red tint from the runway lights and a blue tint from the thrusters.
- They billow outward and upward with turbulence and grow over time.
- Show a small drifting wisp near the ramp foot at the start.
LIGHTING:
- Low ambient light.
- A strong white spotlight at the ramp base aimed upward.
- Red rim light from the runway lights.
- A shadow-casting key light on the car with a 1024 map.
CAMERA: A telephoto handheld look (FOV around 30) from about 120 m away at eye height.
- Pans and zooms follow the car exactly as in the breakdown.
- Add subtle procedural shake, small focus-hunt softness during the fast pan, and motion-blur-like trails using the afterimage effect at low strength.
POST-PROCESSING: EffectComposer with RenderPass, UnrealBloomPass (strong on LEDs and thrusters), a subtle green-ghost lens flare sprite that moves opposite the bright light, film grain and vignette, then OutputPass with ACES tone mapping.
MOBILE SAFETY:
- Use pixelRatio min(devicePixelRatio, 2) and HalfFloatType render targets.
- Clamp all colors to be non-negative and finite before bloom, and normalize only with an epsilon guard.
- Do not use logarithmic depth.
- Lower the particle count when the canvas is narrower than 800 px.
The result should look like real phone footage of the reveal: correct physics, convincing light interaction and rich detail.
Same prompt for both, one shot, zero edits. Short version: one self-contained three.js r160 file, no external assets. A brass-and-glass marble machine on a walnut table: lift wheel with cups, clear glass spiral, 8-bar brass xylophone, seesaw, funnel, gears meshing in sync. Seamless 12 s loop, real gravity, PBR brass/glass/chrome, warm key + cool rim light, dust in the beam, slow dolly-in camera, subtle bloom. Believable physics matters most.
@0xKravn Landing page copy, nothing fancy, both nailed it fast. The gap shows up on longer agent runs where Opus keeps context straight and DeepSeek starts drifting after a few steps, that's where the 26x stops feeling like a bargain
I REVERSE ENGINEERED GRAB. NOW AN AI AGENT RUNS MY WHOLE DAY OF FOOD
I give it macros, budget and ingredients. It scans every menu in Bali, builds the perfect plan and orders it. I only meet the courier.
Sonnet 5.5 just destroyed every model.
I gave Opus 5.5, Sonnet 5.5, Haiku 5.5 and GPT-6 Astra the same prompt: build a browser game about a Grab driver in Bali.
I think Opus finished a close second. Astra, OpenAI's most powerful model, landed on Haiku's level.
I paid $5.70 for a website an agency would quote at $2,000.
I asked Claude Sonnet 5.5 to invent an unusual phone with a 2000s look and a modern twist, then build a sales site for it. It took 20 minutes.
Designers, developers and marketers should be nervous. Only 1 prompt and so good result.
Desktop and mobile versions below.