GPT 6.1 Sol Vanilla
Agent said "Added three AI rivals that follow the course, take both ramps, and recover after missed gates. - Momentum-based collisions push you and other racers. - Live ..." Gate: passed.
Works 1/1 Per working game $0.99
Progressive Oct 23, 2026 Revision 3 3 arms · 3 stages · one run per arm per stage
Jetski Racer: who can keep building?
One jet ski game built in three stages by GPT 6.1 Sol, Opus 5.5, and Opus 5.5 with the threejs-game Workbench. Each stage picks up the code the last one left behind.
Stages
Add AI racers that follow the course and push back on contact. Each arm picks up its own game from Update #1.
The games
Recordings loop. Play opens the real build in a new tab.Opus 5.5 Vanilla
Agent said "I've added five AI jet skis that race you over 3 laps, bump with real push physics, and show up in a live race position display and a finish screen. It still..." Gate: passed.
Works 1/1 Per working game $6.44
threejs-gameWorkbench pompeii/threejs-game on Opus 5.5
Our pickAgent said "You now race five AI jet skis (KAI, MARA, RIO, JUNO, SKYE) over three laps. They follow the course, take the ramps and push you and each other around. The HU..." Gate: passed.
Works 1/1 Per working game $6.21
All three arms did impressive things in their own right. GPT 6.1 Sol was the biggest surprise for its cost and speed. Its game has a distinctive look, and its UI and HUD are clean, with a very cool aesthetic. Vanilla Opus 5.5, as expected, did a great job overall.
But at every stage, from Base through the first update and definitely the final one, the threejs-game Workbench produced the better game. That is the point of a Workbench: a better outcome. I think the gap would only grow if this game kept evolving.
Even the vanilla models did great work. But this is what a Workbench is for. It is tuned for the task at hand, so for about the same price as vanilla Opus you get a higher quality result. Play them yourself and be the judge.
Our pick: threejs-game (pompeii/threejs-game on Opus 5.5) · editorial judgment
Our picks
Editorial judgment, after playing every build- Overall gameplay Workbench threejs-game pompeii/threejs-game on Opus 5.5
The most fun to race, and it pulled further ahead once the AI racers arrived.
- Water physics Workbench threejs-game pompeii/threejs-game on Opus 5.5
The choppiest water and the biggest air at every stage.
- HUD and UI GPT 6.1 Sol Vanilla
The cleanest interface of the three, with a real design point of view.
- Fastest to a playable game Opus 5.5 Vanilla
A working game in 15 minutes, and the fastest over all three stages.
- Most surprising GPT 6.1 Sol Vanilla
A complete, tested three-stage game for $3.73, about a quarter of what the others spent.
Highlights
The final games, after every stage| Fact | GPT 6.1 Sol Vanilla | Opus 5.5 Vanilla | threejs-game pompeii/threejs-game on Opus 5.5 |
|---|---|---|---|
| Language | JavaScript | JavaScript | TypeScript |
| App codeexcludes tests and config | 1,947 lines, 9 files | 3,842 lines, 16 files | 4,686 lines, 21 files |
| Largest fileone long file is harder to change | 623 lines (world.js) | 868 lines (main.js) | 871 lines (Game.ts) |
| Tests it wrote | 1,341 lines: 40 unit, 6 browser | None | 625 lines: scripted browser playtest |
| Shipped sizeJavaScript and CSS, gzipped | 149 KB | 174 KB | 169 KB |
| The race | 3 rivals, 2 laps, 8 gates | 5 rivals, 3 laps, 22 gates | 5 rivals, 3 laps, 8 gates |
| Total cost, all stages | $3.73 | $12.97 | $14.64 |
| Total time, all stages | 99 min | 95 min | 140 min |
Side by side
One run per arm| Metric | GPT 6.1 SolVanilla | Opus 5.5Vanilla | threejs-gamepompeii/threejs-game on Opus 5.5 |
|---|---|---|---|
| Outcome | |||
| Workspassed the gate | 1/1 | 1/1 | 1/1 |
| Said done, was broken | 0/1 | 0/1 | 0/1 |
| Our pick | · | · | Yes |
| Economics | |||
| Cost per working gameincludes failed attempts | $0.99 | $6.44 | $6.21 |
| Total costone run | $0.99 | $6.44 | $6.21 |
| Wall time | 34m 34s | 56m 16s | 76m 21s |
| Running total | |||
| Cost so farBase through Update #2 | $3.73 | $12.97 | $14.64 |
| Wall time so far | 98m 58s | 94m 46s | 139m 39s |
| Effort | |||
| Tokens in / out | 2.8M / 22k | 16.7M / 65k | 15.6M / 61k |
| Turns | 1 | 1 | 1 |
| Tool calls | 115 | 112 | 103 |
| Code | |||
| App code addedthis stage only; excludes tests, config, and lockfiles | 280 lines, 4 files | 1,154 lines, 9 files | 1,080 lines, 7 files |
| Most added to one fileone long file is harder to change | 112 lines | 411 lines | 564 lines |
| Test code addedwritten by the agent | 372 lines | 0 lines | 168 lines |
| Setup | |||
| Model | openai/gpt-6.1-sol | anthropic/claude-opus-5-5 | anthropic/claude-opus-5-5 |
| Workbench | wb-benchmarks/vanilla-gpt-6-1-sol 0.1.1 | wb-benchmarks/vanilla-opus-5-5 0.1.1 | pompeii/threejs-game 0.3.2 |
Works means it builds to a static site, loads in a headless browser with WebGL, draws a real scene, and survives 60 seconds of mashed controls without crashing or freezing. Cost is model spend only and includes failed attempts. Our pick is our call, not a score.
Pinned setup
Everything needed to rerun it# Update #2 prompt, identical for every arm
Add AI jet skis that race me around the track. They should follow the course, take the jumps, and bump into me and each other with real push physics. Show my race position and a finish screen. It still needs to build to a static site I can upload. Nobody will be around to answer questions while you work, so don't wait for input: make the best call yourself and keep going. - GPT 6.1 Sol Vanilla starts at
- pompeii-labs/workbenchmarks-fixtures@9922fc6
- Opus 5.5 Vanilla starts at
- pompeii-labs/workbenchmarks-fixtures@c982863
- Opus 5.5 + pompeii/threejs-game starts at
- pompeii-labs/workbenchmarks-fixtures@a3c0f75
- Gate
- Builds a static site, Loads in a headless browser with WebGL and no errors, Draws a real scene, not a blank or flat canvas, Survives a 60 second playtest
- Attempts
- 1 per arm, no retries, no cherry-picks
- Runs
- 3 scored
- Generated
- 2026-10-23
Revisions
- Oct 23, 2026Revision 1. Base: 3 arms, 1 run each. 3 runs lost to an infrastructure outage were rerun and are not scored.
- Oct 23, 2026Revision 2. Update #1: race track added. Same prompt for every arm, no fix lists. 1 run each, no infrastructure losses.
- Oct 23, 2026Revision 3. Update #2: AI racers added. Same prompt for every arm, no fix lists. 1 run each; a local network drop interrupted grading, and the finished runs were picked up again, not rerun.
More benchmarks
Head to head Oct 26, 2026 2 arms · 2 runs
Who pulls the better espresso? One prompt, one run each: a cup of espresso you stir with your cursor, where speed is force. Opus 5.5 against Opus 5.5 with the threejs-tsl Workbench.
- Opus 5.5 Vanilla 1/1
- threejs-tsl pompeii/threejs-tsl on Opus 5.5 1/1
Head to head Oct 22, 2026 3 arms · 9 runs
Who's the real Pinball Wizard? One pinball prompt, three attempts each: Muse Wonder, Opus 5.5, and Opus 5.5 with the threejs-game Workbench.
- Muse Wonder Vanilla 3/3
- Opus 5.5 Vanilla 3/3
- threejs-game pompeii/threejs-game on Opus 5.5 3/3