Google DeepMind opened public access to Gemini 4 Argon on October 13, 2026, following its initial September 30 Fairwind cybersecurity preview. To see how it handles frontend layout and spatial geometry, I ran a harder animated version of Simon Willison's classic "pelican riding a bicycle" SVG test.

Instead of returning a basic CSS keyframe loop, Gemini 4 Argon generated a 2,653-line single-file interactive SVG scene with a real-time 2-bone Inverse Kinematics (IK) leg solver, road gradient physics, three lighting themes, and a live X-ray geometry overlay.

Watch the same clip on X

The prompt

Simon Willison's original benchmark asks for a static SVG of a pelican riding a bicycle. I pushed the prompt further by asking for a self-contained HTML file with a smooth looping 2D cycling animation and specific visual details:

Create a self-contained HTML file featuring a pelican riding a bicycle.
The scene should have a playful, charming illustration style with a clear sky, soft clouds, and a simple road beneath the bike.
Design the pelican with a large orange beak/pouch, expressive eyes, a small cycling cap (red and white), and sunglasses. Its wings should hold the handlebars, and its legs should be positioned on the pedals.
Build a clean, recognizable bicycle with two spoked wheels, a mint-green vintage frame, handlebars, a saddle, and pedals.
Add smooth, looping 2D animations for:
the wheels rotating
the pedals and the pelican's legs moving in a realistic cycling motion
subtle body bounce as it rides
gentle movement in the cap ribbon and feathers
scrolling road markings, clouds, and speed lines to create a sense of forward motion
Center the composition in a responsive card with a warm light background. Use only HTML, CSS, and inline SVG, with no external dependencies.

Why the pelican test still catches weak models

Drawing a pelican on a bicycle in raw SVG forces a model to solve three problems at once without seeing the rendered screen:

  1. Zero-shot composition: models have plenty of training examples for bicycles and plenty for pelicans, but almost none where a long-beaked waterbird is seated on a road frame with its wings on drop bars and webbed feet clipped to pedals.
  2. Strict frame geometry: a bicycle frame is a double-triangle truss (main triangle plus rear stays) connecting the bottom bracket, wheel hubs, saddle, and head tube. If those coordinates drift by 20 pixels, the bike looks broken right away.
  3. Blind coordinate math: the model does not get a visual canvas. It has to calculate every <path>, Bezier curve, wheel spoke, and joint pivot in <svg viewBox="0 0 1200 650"> purely through text tokens.

Earlier on this site, I tested GPT-6 Astra on a 3D Three.js Pelican Pedal Club scene and a 3D logistics globe dashboard. This run strips away Three.js and tests raw 2D vector geometry.

What Gemini 4 Argon built

Gemini 4 Argon framed the page as an interactive cycling telemetry card titled L'Echappee du Pelican (53x11 Mint Vintage).

Gemini 4 Argon Riviera Noon SVG scene with a pelican riding a mint-green vintage bicycle

1. Analytical 2-bone Inverse Kinematics for the legs

Most models animate bicycle legs by slapping a CSS transform: rotate() on a static leg group, which makes the feet detach from the pedals or bends the knees backward.

Gemini 4 Argon wrote an analytical 2-bone IK solver (solveTwoBoneIK) using the law of cosines. On every animation frame, it takes the fixed hip pivot (492, 232) and the rotating crank pin coordinates on a 32-pixel pedal radius, then solves the exact knee joint position for both the near and far legs (thighLen = 84, calfLen = 86). The webbed feet stay locked to the pedals across the full 360-degree stroke, with a realistic heel-drop angle as the crank turns.

Pressing G (or clicking IK Geometry) turns on an X-ray kinematic overlay. It draws the pedal circle, the hip-to-knee and knee-to-crank vectors, and a live telemetry badge showing knee flexion in degrees.

Gemini 4 Argon Midnight Dynamo mode with the live X-ray IK Geometry overlay active

2. Accurate double-triangle frame and drivetrain

The mint-green frame uses proper lugged steel geometry:

  • Rear hub at (400, 410) and front hub at (690, 410) with 16 tangent spokes per wheel, gumwall tires, and brass valves.
  • Bottom bracket at (515, 410), seat cluster at (468, 265), and head tube running from (645, 255) to (655, 295).
  • A 5-arm alloy chainring with an animated chain (stroke-dasharray="7 5") whose dash offset advances in sync with the crank angle.
  • Red-white-blue French champion stripes on the seat tube, a downtube water bottle cage, and a chrome vintage headlamp.

3. Speed presets, road gradient, and lighting modes

Beyond the base animation, the model added a bottom control dock with keyboard shortcuts:

  • Four pace presets (1 to 4): Coast (0 km/h, where the pelican stops pedaling and levels the cranks while the wheels keep rolling), Sunday (22 km/h), Peloton (35 km/h), and Sprint (54 km/h, where the pelican leans forward +5.5 degrees into an aero tuck).
  • Road gradient slider (-8% to +12%): tilts the entire road and bike assembly and adjusts the computed wattage output in the top bar.
  • Three atmospheres: Riviera Noon, Golden Sunset, and Midnight Dynamo (which turns on a glowing headlamp cone, rear safety tail light, lighthouse beams, and floating fireflies).
  • Brass bell and sardine toss: clicking Ring Bell (B) synthesizes a dual-inharmonic brass ping (1568 Hz and 2349 Hz) via the browser's Web Audio API with zero external sound files. Clicking Toss Sardine (F) launches a silver sardine along a parabolic arc; as it nears the beak, the pelican opens its orange gular pouch wide, catches the fish, and triggers a throat-swallow bulge down the neck.

The caveat: it used inline JavaScript

There is one clear constraint miss worth calling out. My prompt ended with:

Use only HTML, CSS, and inline SVG, with no external dependencies.

Gemini 4 Argon kept zero external dependencies and rendered 100% of the graphics inside inline SVG, but it wrote roughly 1,100 lines of inline vanilla JavaScript to drive the requestAnimationFrame loop, the 2-bone IK solver, the control dock, and the Web Audio bell.

If you grade the prompt strictly as a CSS-only challenge, adding <script> breaks the rule. In practice, pure CSS cannot compute law-of-cosines inverse kinematics per frame or synthesize audio waveforms. The model traded strict CSS purity for a working kinematic rig that keeps the feet attached to the pedals.

Takeaway

The pelican-on-a-bicycle prompt remains a fast sanity check for a new model's spatial reasoning. Gemini 4 Argon nailed the bicycle truss geometry, layered near-and-far leg depth, and pedal kinematics on the first shot.

Sources