benchmaxxed

Bench

The same prompt across models. Run the returned pages side by side.

The pilot has one attempt per model in each main comparison. Wall time includes CLI startup and contention.

We ran these tests on a subscription. “API price” is what the same tokens would cost at OpenAI’s Standard list prices on 18 Nov 2026.

Pelican on a bicycle

Pilot 18 Nov 2026

A coastal SVG at dusk. Both wheels turn, the feet follow the pedals and pause stops the whole scene.

Exact prompt
Make an animated SVG of a pelican riding a bicycle along a coast at dusk. In one HTML file.

Draw the pelican and bicycle in enough detail that their silhouettes read at a glance. Show both wheels turning and the pelican actively pedaling, with its feet visibly following the pedals. Move the scenery behind it to convey travel. Add one clearly visible Pause / Play button that stops and resumes the entire scene, including the landscape. Let the scene start playing on load. Compose for a 1920 by 1080 viewport and use most of the frame. Art direction is up to you.

Work from this prompt alone. Do not use tools or access files. Return only the complete HTML document, including the SVG, CSS and any JavaScript. It must work offline with no libraries or external assets.

gpt-6.2-astra

Pelican Coast Ride

CLI wall
150.158 s
Output tokens
5,780
API price
$0.4242

The feet follow the pedals. The coastal scene and both wheels move.

gpt-6-luna

Pelican by the Coast at Dusk

CLI wall
123.923 s
Output tokens
6,549
API price
$0.0044

The wheels and scenery move. The feet swing clear of the pedals.

gpt-6.2-sol

Pelican Coast

CLI wall
185.185 s
Output tokens
7,259
API price
$0.1093

The feet visibly follow the moving pedals.

Additional take

gpt-6.2-sol

The Dusk Ride

CLI wall
232.806 s
Output tokens
8,449
API price
$0.1213

The second take changes the bicycle and coastline. The feet still follow the pedals.

Lunar lander

Pilot 18 Nov 2026

A single-file game with fuel, rotation, thrust and a marked pad. The original returns use the arrow keys and R.

Exact prompt
Build a playable lunar lander in one HTML file. Land upright on a lit pad before the fuel runs out.

Use a side-on view with a visible ship, terrain and a clearly marked landing pad. Left and Right arrow keys rotate, Up thrusts, and R restarts from the same initial state. Gravity applies continuously after the first flight input. A tilted or fast impact must crash, even on the pad. A gentle upright touchdown on the pad must register a landing. Show remaining fuel, make the current flight outcome obvious, and make the ship and flame large enough to see at a glance. Start airborne with the ship near the pad so a player can reach it in a few seconds. Make the scene distinctive and use most of a 1920 by 1080 viewport. Art direction is up to you.

Work from this prompt alone. Do not use tools or access files. Return only the complete HTML document, including all CSS, graphics and JavaScript. It must work offline with no libraries or external assets.

gpt-6.2-astra

No return

Timed out
CLI wall
300.086 s
Output tokens
—
API price
no usage reported

Timed out. No HTML was returned.

gpt-6-luna

Moonfall — Lunar Lander

CLI wall
91.795 s
Output tokens
4,833
API price
$0.0046

The ship stayed at Flight Ready when the arrow keys were pressed.

gpt-6.2-sol

Umbra Station · Lunar Lander

CLI wall
158.859 s
Output tokens
6,994
API price
$0.1095

The recorded test showed a tilt crash and an upright landing after restart.