One brief, 142 builds.

The run OpenCode 2.0.12 · selfhosted/Kolibri-1

Time

26 min stopped at the bench’s time limit

Tokens

2.09M 0 read from cache · 190k written

Cost

none per token the weights ran on a rented GPU, paid by the hour

Work

152 steps 1 session · 276 tool calls, 19 failed

Claude’s note. Public Kolibri-1 weights at Hugging Face revision e52eb4627d11516b0c01de49210ab5a4e4061444, served on a rented H200 NVL. Strict smoke passed; game generation timed out after 26 minutes, leaving index.html and main.js. The untouched page fails to resolve three and creates no canvas in a CPU-only browser with WebGPU disabled. A separate import-map/module-alias diagnostic copy then fails because BufferAttribute needs a Typed Array. Neither copy is published; no score or gameplay repair was made. The zero cost field is the harness token-price figure, not the GPU rental bill.

The run OpenCode 2.0.12 · openrouter/x-ai/grok-4.7#xhigh

Time

1 h 1 min

Tokens

2.88M 2.38M read from cache · 123k written

Cost

$2.68 pay-per-token, as OpenCode priced it

Work

33 steps 1 session · 44 tool calls, 1 failed

Claude’s note. Grok 4.7 at xhigh through OpenRouter. It spent its first 25 minutes reading three.js r181's sources and examples, then wrote index.html and a 71 KB js/main.js on three.js from jsDelivr without opening the page. It then started python3 -m http.server in the foreground; the call never returned, and after 22 idle minutes the run was stopped, which leaves the same files the time limit would have. The page stops at Loading modules: main.js line 808 writes -(...) ** 2, a syntax error in JavaScript, so no world or jeep ever appears.

The run OpenCode 2.0.14 · opencode/fledge-alpha-free

Time

2 min

Tokens

69k 48k read from cache · 8k written

Cost

$0 a free model

Work

6 steps 1 session · 5 tool calls, 1 failed

Claude’s note. Free preview model on OpenCode Zen. It wrote a single index.html in about two minutes and said it was done without opening it in a browser. The page imports three/tsl from three@0.170.0, where no such file exists (404), so the module fails to load and nothing is drawn. The first attempt died on a provider 503 before the model wrote anything useful and is kept as aborted.

The run OpenCode 2.0.12 · nano-gpt/inclusionai/ling-3.1-flash#high

Time

1 h 5 min

Tokens

23.83M 23.51M read from cache · 221k written

Cost

$0.41 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

257 steps 1 session · 266 tool calls, 1 failed

Claude’s note. It worked on its own for 65 minutes over 257 steps; one of 266 tool calls failed. The game was written by the 20th minute: index.html, style.css and a main.js of 1,200 lines on three.js r170 from jsDelivr. It then installed puppeteer-core in the box and spent most of the run in its headless Chromium, which has no graphics card, reading three.js's WebGPU build to learn why the WebGL2 fallback drew nothing until it did, and closed with a long scripted drive and a check of the minimap. On a card it takes the WebGPU path: a valley under a snowy ring of mountains, a lake, a dirt loop, pines and rocks, timed laps with a best lap, but no rivals, and the jeep drives through the trees.

The run Freebuff 0.0.203 · upstage/solar-mini4

Time

1 h 13 min

Tokens

not counted Freebuff reports only the context: 195k at most, 14.01M over its 174 steps

Cost

5 Freebucks Freebuff’s own currency, 5 an hour of session; the owner’s paid access, not dollars

Work

174 steps 1 session · 172 tool calls, 13 failed

Claude’s note. A rerun on the owner's request; the first build never loaded three.js. 73 minutes and 174 steps, most of them terminal commands. The track code uses a variable outside the loop that declares it, so building the scene throws and the page stays plain sky blue.

The run OpenCode 2.0.12 · openrouter/deepseek/deepseek-v4-pro-0813#max

Time

8 min

Tokens

935k 761k read from cache · 83k written

Cost

$0.50 pay-per-token, as OpenCode priced it

Work

10 steps 1 session · 19 tool calls

Claude’s note. A rerun from scratch through OpenRouter after NanoGPT's weekly limit cut the first one off, at max effort. Eight minutes and ten steps: jeep-race/index.html and main.js with three.js r180 from jsDelivr. It passes plain numbers to setColorAt for the gate cones, so setup throws before the jeep or the frame loop is made, and the HUD sits over a black screen.

The run OpenCode 2.0.12 · openrouter/upstage/solar-mini4#max

Time

23 min

Tokens

21.83M 20.80M read from cache · 159k written

Cost

$0.18 pay-per-token, as OpenCode priced it

Work

173 steps 1 session · 172 tool calls, 13 failed

Claude’s note. A rerun on OpenRouter at max effort, after the provider stalled the previous attempt. 23 minutes and 173 steps: a Vite project, built in the box image because its config names the box's own path. The track curve is given plain arrays instead of vectors, so setup throws and START stops on a track that was never made; nothing of the world is drawn.

The run OpenCode 2.0.12 · openrouter/nvidia/nemotron-3.5-lightning

Time

2 min

Tokens

579k 541k read from cache · 10k written

Cost

$0.03 pay-per-token, as OpenCode priced it

Work

33 steps 1 session · 32 tool calls

Claude’s note. A rerun on the owner's request. Ninety seconds: one index.html on three.js r160 from unpkg, whose module script carries Python-style comments, a syntax error in JavaScript, so none of it runs.

The run OpenCode 2.0.12 · openrouter/inclusionai/ling-3.0-flash-fin

Time

27 min

Tokens

1.58M 1.31M read from cache · 176k written

Cost

$0.05 pay-per-token, as OpenCode priced it

Work

33 steps 1 session · 33 tool calls, 5 failed

Claude’s note. A rerun on the owner's request; the first run wrote no files. This time 27 minutes and 174,000 output tokens went into one 12 KB index.html whose script stops mid-code and imports three by bare name, so nothing runs.

The run Freebuff 0.0.203 · upstage/solar-pro4

Time

5 min

Tokens

not counted Freebuff reports only the context: 15k at most, 117k over its 9 steps

Cost

10 Freebucks Freebuff’s own currency, 10 an hour of session; the owner’s paid access, not dollars

Work

8 steps 1 session · 8 tool calls, 1 failed

Claude’s note. A rerun on the owner's request. Under five minutes: it created Vite's starter project and wrote its own index.html with a HUD, but left the starter's main.js, which writes into an element the page no longer has, so it throws at load. No game code was written.

The run OpenCode 2.0.12 · openrouter/ibm-granite/granite-4.2-8b#high

Time

24 min

Tokens

739k 625k read from cache · 94k written

Cost

$0.03 pay-per-token, as OpenCode priced it

Work

34 steps 1 session · 33 tool calls, 1 failed

Claude’s note. A rerun on the owner's request. In 24 minutes it installed three.js with npm and wrote an index.html and a main.js that imports three by bare name, with no bundler and no import map, and never served or opened the page, so the browser refuses the module and nothing runs.

The run OpenCode 2.0.12 · openrouter/inception/mercury-2.5#high

Time

12 s

Tokens

10k 0 read from cache · 3k written

Cost

$0.00 pay-per-token, as OpenCode priced it

Work

1 step 1 session · 0 tool calls

Claude’s note. A rerun on the owner's request. One step of twelve seconds: it answered with a whole index.html in the chat and called no tool, so nothing reached the folder, the same kind of answer as the first run's.

The run OpenCode 2.0.12 · nano-gpt/moonshotai/kimi-k2.6:thinking

Time

1 min

Tokens

24k 0 read from cache · 10k written

Cost

$0.03 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

2 steps 1 session · 1 tool call

Claude’s note. A rerun on the owner's request, in the box that now has a browser. It wrote nothing again: in one minute and two steps it listed the empty folder, set out its plan for the game in the chat and ended its turn.

The run Freebuff 0.0.203 · google/gemini-3.8-flash#max

Time

3 min

Tokens

not counted Freebuff reports only the context: 60k at most, 1.33M over its 37 steps

Cost

80 Freebucks Freebuff’s own currency, 80 an hour of session; the owner’s paid access, not dollars

Work

37 steps 1 session · 37 tool calls

Claude’s note. A rerun on the owner's request, in the box that now has a browser. Under four minutes: a Vite project in TypeScript, built but never opened. The terrain code declares a variable with let in one method and assigns it in others, so the page throws at load and stays dark.

The run OpenCode 2.0.12 · opencode/space-bunny-free#max

Time

1 h 23 min

Tokens

95.06M 94.44M read from cache · 446k written

Cost

$0 a free model

Work

461 steps 1 session · 482 tool calls, 16 failed

Claude’s note. Space Bunny Alpha again, on the owner's request, through OpenCode Zen, where it is free (Freebuff does not offer it). 83 minutes and 461 steps on its own: a Vite project of about 6,500 lines with a procedural 1 km landscape, a filleted circuit, AI rivals, audio and node tests of its physics and scene. It looked for a browser, found none and never opened the page. The loader stops at 97%: setting up the race calls this.updateParticleCamera, an imported function it never made a method, so nothing can be driven.

The run Antigravity CLI 1.2.12 · agy/gemini-3.8-flash-high

Time

6 min

Tokens

2.56M 2.08M read from cache · 73k written

Cost

$0.74 at google/gemini-3.8-flash list price; paid through a flat-rate plan, so not a charge

Work

45 steps 1 session · 45 tool calls

Claude’s note. Gemini 3.8 Flash again on the owner's request, now in Google's Antigravity CLI at gemini-3.8-flash-high, its top effort there. Six minutes and 45 steps: a Vite project with rivals, nitro, three cameras, sound and a minimap over detailed hills. But its drag takes 1.2% of the speed every frame against an engine of 420 over 1650 kg, so the jeep tops out near 1 km/h and never gets going.

The run OpenCode 2.0.12 · openrouter/tencent/hy3#high

Time

7 min

Tokens

339k 242k read from cache · 43k written

Cost

$0.04 pay-per-token, as OpenCode priced it

Work

8 steps 1 session · 7 tool calls

Claude’s note. A rerun from scratch through OpenRouter after NanoGPT's weekly limit cut the first one off; high effort, as max failed the bench's smoke test. Seven minutes and eight steps: one index.html and src/main.js with three.js from jsDelivr. Green rolling hills with pines and rocks, eight gates round a loop with a compass to the next, laps timed without end; no road, rivals or sound.

The run OpenCode 2.0.12 · openrouter/z-ai/glm-5.3-prime#max

Time

34 min

Tokens

11.07M 10.73M read from cache · 154k written

Cost

$7.87 pay-per-token, as OpenCode priced it

Work

85 steps 1 session · 86 tool calls, 1 failed

Claude’s note. A rerun on the owner's request; the first build drew only its HUD. 34 minutes and 85 steps: a static page with three.js vendored, a loop through meadows and dirt under snowy peaks, a lake, pines, three rivals that race, a minimap and sound, at 60 frames a second on WebGPU and 48 with the processor slowed. But the jeep comes apart as it drives, its body flying up while the wheels stay on the ground, and GO! never leaves the screen.

The run OpenCode 2.0.12 · opencode/space-bunny-free#max

Time

2 h 41 min

Tokens

131.45M 130.40M read from cache · 389k written

Cost

$0 a free model

Work

486 steps 1 session · 499 tool calls, 25 failed

Claude’s note. A rerun on the owner's request, through OpenCode Zen, where it is free; the first Zen run stalled on its loading screen. 161 minutes and 486 steps on its own: a Vite project of about 40 modules, TSL materials and post-processing, a 1.44 km island with a lake and a banked road carved into it, a title menu with a free-roam mode, and headless physics tests. With a browser in the box it rendered its own test pages and checked them. There are no rivals, and in the pilot's minute the jeep ends up outside the track wall and runs along it off the road.

The run Freebuff 0.0.203 · z-ai/glm-5.3-flash#max

Time

43 min

Tokens

not counted Freebuff reports only the context: 50k at most, 815k over its 28 steps

Cost

5 Freebucks Freebuff’s own currency, 5 an hour of session; the owner’s paid access, not dollars

Work

28 steps 1 session · 27 tool calls

Claude’s note. A rerun on the owner's request; the first build hid behind its own no-WebGPU panel. 43 minutes and 28 steps: a varied 2 km world with a lake crossed by a dirt causeway, green hills with pines and sand dunes, twelve gates, a minimap and sound, at 60 frames a second on WebGPU. Held at full throttle the jeep leaves the world at 54 seconds and drives on into a pale void.

The run Freebuff 0.0.203 · mimo/mimo-v2.6-pro

Time

32 min

Tokens

not counted Freebuff reports only the context: 96k at most, 2.25M over its 37 steps

Cost

30 Freebucks Freebuff’s own currency, 30 an hour of session; the owner’s paid access, not dollars

Work

37 steps 1 session · 43 tool calls

Claude’s note. A rerun on the owner's request; the first build stopped at load. 32 minutes and 37 steps, checked with node simulations of the track and the race rather than in a browser. Hills, a lake, pines and a dirt loop with three rivals that do race, at 60 frames a second on WebGPU but under 9 with the processor slowed. GO! stays on screen all race, the position reads 4th throughout, and there is no sound.

The run Command Code 1.66.0 · commandcode/google/gemini-3.5-flash#high

Time

7 min

Tokens

3.86M 3.49M read from cache · 50k written

Cost

$1.46 at google/gemini-3.5-flash list price; paid through a flat-rate plan, so not a charge

Work

68 steps 1 session · 66 tool calls

Claude’s note. A rerun on the owner's request, at high effort, the most Command Code allows for it. Seven minutes and 68 steps: a menu of three terrains, a canyon track with eight rivals that never move, three laps of eight checkpoints. It chose three.js's WebGL renderer on purpose, against the brief's WebGPU, and never opened the page: the jeep tumbles within seconds and does not recover.

The run MiniMax Code 0.5.5 · minimax/MiniMax-M3.1-Flash-Preview#max

Time

1 h 35 min

Tokens

87.00M 86.53M read from cache · 349k written

Cost

$0 MiniMax Code priced every request at $0 on the owner’s account; no list price is published for this model, so it is not called free

Work

293 steps 1 session · 293 tool calls, 43 failed

Claude’s note. One MiniMax Code session at max effort, 95 minutes and 293 requests on the day the model opened there. It vendored three.js, wrote twenty modules and its own tests, drove an autopilot lap in Node and checked the page in a headless browser, then ended by itself. The world is the most detailed of the week: a 2.4 km map of eroded ridges, lakes and 1.7 million triangles, a 6 km track, pines, boulders and grass, at 60 frames a second on WebGPU. But the dust puffs assign a number to a sprite's read-only rotation, so the moment the jeep rolls off the line the update loop throws every frame: it freezes at 7 km/h with the clock at zero and never moves.

The run OpenCode 2.0.12 · nano-gpt/openai/gpt-oss-120b#high

Time

3 min

Tokens

24k 32 read from cache · 4k written

Cost

$0.01 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

4 steps 1 session · 2 tool calls

Claude’s note. High reasoning effort, the highest listed. In 3 minutes over 4 steps it wrote an HTML page and one module; NanoGPT answered its last step with 'Upstream emitted malformed tool call data', after the model had written its closing message. The game does not start, also on its untouched CDN page.

The run Freebuff 0.0.203 · google/gemini-3.8-flash#max

Time

3 min

Tokens

not counted Freebuff reports only the context: 47k at most, 639k over its 26 steps

Cost

80 Freebucks Freebuff’s own currency, 80 an hour of session; the owner’s paid access, not dollars

Work

26 steps 1 session · 26 tool calls

Claude’s note. One Freebuff session at max reasoning: under three minutes and 26 steps, a Vite project with an AI rival, a sound manager and a Russian HUD, built but never opened. Its camera follower calls a vector method three.js does not have, so the loop throws from the first frame and the page shows only the HUD over black.

The run Freebuff 0.0.203 · mimo/mimo-v2.6-pro

Time

35 min

Tokens

not counted Freebuff reports only the context: 85k at most, 683k over its 15 steps

Cost

30 Freebucks Freebuff’s own currency, 30 an hour of session; the owner’s paid access, not dollars

Work

14 steps 1 session · 30 tool calls

Claude’s note. One Freebuff session, no reasoning setting offered: 35 minutes, 14 steps, a Vite project of twelve modules that it built without error but never opened. At load it passes a bare function to three.js's SimplexNoise, which wants an object with random(), and the page stops on that error before drawing anything.

The run Freebuff 0.0.203 · upstage/solar-pro4

Time

5 min

Tokens

not counted Freebuff reports only the context: 32k at most, 176k over its 9 steps

Cost

10 Freebucks Freebuff’s own currency, 10 an hour of session; the owner’s paid access, not dollars

Work

9 steps 1 session · 9 tool calls

Claude’s note. One Freebuff session, no reasoning setting offered: five minutes, nine steps, four files written in one pass and never run. It chose WebGL itself, setting preferWebGPU to false, and the world build throws in its Catmull-Rom spline at load, so the menu sits over a black page.

The run Freebuff 0.0.203 · upstage/solar-mini4

Time

13 min

Tokens

not counted Freebuff reports only the context: 142k at most, 13.81M over its 136 steps

Cost

5 Freebucks Freebuff’s own currency, 5 an hour of session; the owner’s paid access, not dollars

Work

136 steps 1 session · 135 tool calls, 6 failed

Claude’s note. One Freebuff session, which offers no reasoning setting for Solar Mini 4: 13 minutes, 136 steps and 96 shell commands, mostly node --check and curl against its own server. It never loaded the page in a browser, where the bare import of three with no import map stops every script, leaving a menu whose Start race button does nothing.

The run Freebuff 0.0.203 · z-ai/glm-5.3-flash#max

Time

33 min

Tokens

not counted Freebuff reports only the context: 46k at most, 672k over its 23 steps

Cost

5 Freebucks Freebuff’s own currency, 5 an hour of session; the owner’s paid access, not dollars

Work

23 steps 1 session · 24 tool calls

Claude’s note. One Freebuff session at max reasoning, 33 minutes and 23 steps, ended by itself after checking its files with node but never in a browser. The game underneath is real: WebGPU at 60 frames a second, a 2.4 km carved loop, checkpoints and three cameras. But its own NO WEBGPU panel stays over the whole page, because the style sheet's display: flex overrides the hidden attribute, so a player sees only that panel.

The run OpenCode 2.0.12 · nano-gpt/google/gemma-4-31b-it:thinking

Time

4 min

Tokens

55k 23k read from cache · 10k written

Cost

$0.01 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

4 steps 1 session · 7 tool calls, 2 failed

Claude’s note. Reasoning on (the :thinking route); NanoGPT lists no effort levels for it. In 4 minutes over 4 steps it wrote an HTML page, a script and a style sheet; the game does not start, also on its untouched CDN page.

The run OpenCode 2.0.12 · openrouter/nvidia/nemotron-3.5-lightning

Time

2 min

Tokens

612k 496k read from cache · 24k written

Cost

$0.03 pay-per-token, as OpenCode priced it

Work

17 steps 1 session · 16 tool calls, 4 failed

Claude’s note. Run through OpenRouter: on NanoGPT, the subscription route, the model timed out twice before it wrote anything (25 and 26 Sep; kept aside, not runs). On OpenRouter its reasoning is only on or off, so no effort level was set. In 2 minutes over 17 steps it wrote one HTML page; the game does not start, also on its untouched CDN page.

The run Freebuff 0.0.203 · openai/gpt-6-luna#max

Time

10 min

Tokens

not counted Freebuff reports only the context: 74k at most, 942k over its 19 steps

Cost

20 Freebucks Freebuff’s own currency, 20 an hour of session; the owner’s paid access, not dollars

Work

19 steps 1 session · 26 tool calls

Claude’s note. One Freebuff session at max reasoning: ten minutes and 19 steps. It never installed its dependencies, so its own npm run build failed, and it said so and checked the syntax with node instead. Vite's default target then rejects its top-level await; the copy here is built with --target esnext, the code untouched. It has the most polished interface of the batch and a low-poly gravel loop, but the world is 1 km across with no edge: hold the throttle and the jeep drives off it into a black void.

The run Freebuff 0.0.203 · mimo/mimo-v2.5

Time

1 h 15 min stopped at the bench’s time limit

Tokens

not counted Freebuff reports only the context: 231k at most, 14.66M over its 111 steps

Cost

10 Freebucks Freebuff’s own currency, 10 an hour of session; the owner’s paid access, not dollars

Work

110 steps 1 session · 124 tool calls, 1 failed

Claude’s note. One Freebuff session; Freebuff offers no reasoning setting for it and sends it as mimo/mimo-v2.5. The bench stopped it at 75 minutes, a limit set for Freebuff's hour-long session before we saw that Freebuff lets a working agent finish: it was still running its own browser tests of the gameplay, 110 steps in. Published as it stood. It is a real race: WebGPU at 60 frames a second, three AI rivals over three laps of twelve checkpoints, a start gate, minimap and pause, but no sound; the jeep can wedge itself on snowy slopes.

The run Freebuff 0.0.203 · deepseek/deepseek-v4-flash#max

Time

1 h 20 min

Tokens

not counted Freebuff reports only the context: 258k at most, 16.00M over its 105 steps

Cost

15 Freebucks Freebuff’s own currency, 15 an hour of session; the owner’s paid access, not dollars

Work

105 steps 1 session · 138 tool calls, 1 failed

Claude’s note. One Freebuff session at max reasoning; it worked 80 minutes, on past the session's hour, wrote fifteen modules and its own driving and scene tests, and ended by itself. The world is 3.2 km of snowy ridges, a river, a lake and pines at 60 frames a second on WebGPU, with a countdown, three laps of eight checkpoints and a good HUD, but no rivals and no sound. Long stretched strips of geometry hang in the sky, debris floats over the lake, and the jeep drives straight across the water.

The run OpenCode 2.0.12 · opencode/longcat-2.5-preview-free

Time

3 min

Tokens

146k 113k read from cache · 11k written

Cost

$0 a free model

Work

9 steps 1 session · 13 tool calls, 1 failed

Claude’s note. Three minutes and nine steps, free on OpenCode Zen: one index.html written in a single go, then a few edits and a syntax check of its script; it never opened the page. It draws with WebGPU at 60 frames a second: rolling procedural hills coloured by height and slope, a lake, instanced cone pines, bushes and rocks, fog. The jeep drives, climbs and takes air, but there is no road, only twelve checkpoint markers that a lap silently resets; no rivals, no sound, and a debug line reading WebGPU renderer ready stays in the middle of the screen.

The run OpenCode 2.0.12 · nano-gpt/z-ai/glm-5.1:thinking#high

Time

46 min

Tokens

11.06M 10.06M read from cache · 91k written

Cost

$2.42 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

155 steps 1 session · 153 tool calls, 2 failed

Claude’s note. High reasoning effort, the highest NanoGPT lists for it. The first attempt hit NanoGPT's 296-second request limit before the model wrote anything, so it is not a run and was run again. In 46 minutes over 155 steps it wrote an index.html and eleven modules under src/, with three.js r169's WebGPU build from jsDelivr through an import map; the page draws with WebGPU. Vendored for the page.

The run OpenCode 2.0.12 · nano-gpt/qwen/qwen3-coder-next

Time

46 s

Tokens

36k 0 read from cache · 8k written

Cost

$0.02 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

3 steps 1 session · 2 tool calls

Claude’s note. A model without reasoning. In 46 seconds over 3 steps it wrote one file, jeep-race-game.html, published here as the page's index.html and otherwise unchanged; three.js r160 from a CDN, vendored; the page draws with WebGL. Its start screen is in Russian.

The run OpenCode 2.0.12 · nano-gpt/xiaomi/mimo-v2.5:thinking#high

Time

9 min

Tokens

189k 162k read from cache · 19k written

Cost

$0.01 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

11 steps 1 session · 8 tool calls, 1 failed

Claude’s note. High reasoning effort, the highest NanoGPT lists for it. In 9 minutes over 11 steps it wrote one HTML page; its WebGPU renderer is imported from a path three.js r170 does not have, so the page falls back to its own WebGL path. NanoGPT answered its last step with 'Upstream emitted malformed tool call data', after the model had written its closing message.

The run OpenCode 2.0.12 · nano-gpt/minimax/minimax-m2.7

Time

3 min

Tokens

186k 147k read from cache · 12k written

Cost

$0.05 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

12 steps 1 session · 13 tool calls

Claude’s note. NanoGPT lists no effort setting for it; run once at the default. In 3 minutes over 12 steps it wrote one HTML page, and a SPEC.md beside it, with three.js r160 from a CDN; the page draws with WebGL. Vendored for the page.

The run Command Code 1.50.1 · commandcode/google/gemini-3.6-flash#high

Time

3 min

Tokens

1.03M 818k read from cache · 23k written

Cost

$0.29 at google/gemini-3.6-flash list price; paid through a flat-rate plan, so not a charge

Work

33 steps 1 session · 32 tool calls, 2 failed

Claude’s note. High reasoning effort, the highest Command Code offers for it; run again at the owner's request after the first run went at the CLI's default. It worked for 3 minutes over 33 steps and wrote a Vite project, built with vite build for the page.

The run Command Code 1.50.1 · commandcode/google/gemini-3.5-flash#high

Time

6 min

Tokens

939k 759k read from cache · 53k written

Cost

$0.78 at google/gemini-3.5-flash list price; paid through a flat-rate plan, so not a charge

Work

19 steps 1 session · 18 tool calls

Claude’s note. High reasoning effort: Command Code refuses max for this model. The game does not start, also on its untouched CDN page.

The run Command Code 1.50.1 · commandcode/thinkingmachines/inkling-small

Time

22 s

Tokens

59k 39k read from cache · 7k written

Cost

0.02 credits Command Code’s own unit, paid through a flat-rate plan; no list price is published for this model

Work

3 steps 1 session · 2 tool calls

Claude’s note. No effort setting for this model on Command Code; run once at the CLI's default. models.dev lists no price for it. The game does not start, also on its untouched CDN page.

The run OpenCode 2.0.12 · nano-gpt/moonshotai/kimi-k2.7-code

Time

2 min

Tokens

37k 9k read from cache · 2k written

Cost

$0.04 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

4 steps 1 session · 3 tool calls, 2 failed

Claude’s note. It wrote nothing. Its web search failed and unpkg answered with a server error; it then got the same package.json from jsDelivr, but replied that a tool had returned an error and it could not complete the request. The run ended there, after four steps and 104 seconds.

The run OpenCode 2.0.12 · nano-gpt/xiaomi/mimo-v2.5-pro:thinking

Time

22 min

Tokens

345k 299k read from cache · 26k written

Cost

$0.03 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

26 steps 1 session · 9 tool calls

Claude’s note. It worked for 22 minutes over 26 steps and wrote one index.html: layered terrain with biomes, instanced trees, rocks and grass, a lake, a spline track with barriers and a start arch, a detailed jeep and three laps over four checkpoints. Twelve times the upstream sent a tool call OpenCode could not parse, and each time it carried on. It tried twice to open the page in OpenCode's browser, but none was connected. Its WebGPU import is a path three.js r170 does not have, so it falls back to WebGL2, and then building the jeep throws on a read-only position: the loading screen never clears. It finished by itself, sure the game was complete.

The run OpenCode 2.0.12 · nano-gpt/tencent/hy3

Time

44 min

Tokens

7.32M 6.86M read from cache · 193k written

Cost

$0.27 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

138 steps 2 sessions · 146 tool calls, 2 failed

Claude’s note. It worked for 44 minutes over 138 steps and wrote fourteen modules: terrain, scenery, a TSL sky and water, a jeep with its own vehicle model, AI rivals, a race, cameras, a HUD and a minimap. It said it could not run a real browser, so it checked the CDN files with curl and every module with node --check, and had a subagent review the code. None of that caught sky.js calling pow without importing it: the page stops at once on 'pow is not defined'. It finished by itself, sure the game was complete.

The run OpenCode 2.0.12 · nano-gpt/moonshotai/kimi-k2.6:thinking

Time

18 s

Tokens

15k 7k read from cache · 1k written

Cost

$0.01 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

2 steps 1 session · 1 tool call

Claude’s note. It wrote nothing. It listed the empty folder, set out in the chat what the game would have, said it would start by creating the files, and ended its turn there, after two steps and 18 seconds.

The run OpenCode 2.0.12 · nano-gpt/inclusionai/ling-3.0-flash:thinking

Time

36 s

Tokens

40k 7k read from cache · 14k written

Cost

$0.00 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

2 steps 1 session · 1 tool call

Claude’s note. It wrote the whole game, one jeep-race/index.html, in a single step and stopped after 36 seconds without opening it. It loads three.js r160 and calls THREE.WebGPURenderer, which that build does not have, so the START RACE button throws and no scene is made.

The run OpenCode 2.0.12 · nano-gpt/deepseek/deepseek-v4-flash-0731:thinking

Time

14 min

Tokens

416k 323k read from cache · 18k written

Cost

$0.01 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

17 steps 1 session · 13 tool calls

Claude’s note. It worked for 14 minutes over 17 steps and wrote one jeep-race.html: a menu, terrain on a 300-by-300 grid, a jeep, a track and a HUD. Three provider errors came along the way and it carried on. It served the page and checked it answered with curl; OpenCode's browser was not connected, so it never saw it run. It imports its WebGPU renderer from a path three.js r170 does not ship, so the script never loads and START RACE does nothing. It finished by itself, sure the game was complete.

The run OpenCode 2.0.12 · nano-gpt/qwen/qwen3.5-397b-a17b:thinking

Time

2 min

Tokens

109k 0 read from cache · 10k written

Cost

$0.09 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

8 steps 1 session · 7 tool calls

Claude’s note. It worked for 2 minutes over 8 steps and wrote a Vite project on three.js 0.170: terrain, a jeep with roll cage and spare tyre, checkpoints and a HUD. It started the dev server and declared the game running without looking at it. It calls THREE.WebGPURenderer on the plain three entry, which has none in that release, so the page shows its HUD over a white screen and nothing is drawn.

The run OpenCode 2.0.12 · nano-gpt/z-ai/glm-5.2:thinking

Time

2 min

Tokens

34k 10 read from cache · 12k written

Cost

$0.03 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

2 steps 1 session · 1 tool call

Claude’s note. It wrote the whole game, one jeep-race.html, in a single step and stopped after 108 seconds without opening it: terrain with warped noise and vertex colours, a Catmull-Rom track, a jeep and a race. Its import map points three/tsl at a file three.js r169 does not ship, so the script never loads and the page stays on 'Initializing WebGPU...'.

The run OpenCode 2.0.12 · nano-gpt/minimax/minimax-m3:thinking

Time

19 min

Tokens

17.25M 17.03M read from cache · 78k written

Cost

$1.16 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

209 steps 1 session · 208 tool calls, 13 failed

Claude’s note. It worked for 19 minutes over 209 steps, saved three.js r169 next to the game and split it into nine modules: layered terrain, a shader sky, a jeep with its own physics, scattered forests and rocks, checkpoints and a minimap. It tried to fetch Chrome for puppeteer to see the page, but the bench's proxy refused the download host, so it checked the modules with node --check and grep instead. main.js calls WebGPURenderer without the THREE namespace it imported, so the page stops at once on its own 'WebGPU unavailable' notice. It finished by itself, sure the game worked.

The run OpenCode 2.0.12 · nano-gpt/deepseek/deepseek-v4-pro-0813:thinking

Time

11 min

Tokens

753k 513k read from cache · 49k written

Cost

$0.35 flat-rate plan: OpenCode’s price for the tokens, not a charge

Work

19 steps 1 session · 19 tool calls, 2 failed

Claude’s note. It worked for 11 minutes over 19 steps. Its two web searches failed, so it read the three.js r170 WebGPU build itself, then wrote one index.html of about 1,200 lines with layered terrain, a lake, instanced trees and rocks, and AI rivals on a spline track. It tried to open the page in a browser, found none connected, and settled for static checks in Node. The frame loop moves the jeeps, the HUD and the minimap but never calls renderer.render, so nothing is ever drawn. It finished by itself, sure the game was complete.

The run OpenCode 2.0.12 · openrouter/z-ai/glm-5.3-prime#max

Time

33 min

Tokens

20.77M 20.31M read from cache · 185k written

Cost

$13.77 pay-per-token, as OpenCode priced it

Work

122 steps 1 session · 123 tool calls, 1 failed

Claude’s note. It worked for 33 minutes over 122 steps and wrote one self-contained index.html, about 2,000 lines, with a carved track, splat-shaded terrain, a lake and scattered trees. It never opened a browser: it tested the terrain and a pure-pursuit autopilot in Node with three.js and fixed four bugs it found there, but never rendered a frame. In the browser the menu and HUD work, yet merging the pine geometry fails and every frame throws in the renderer, so no world is ever drawn. It finished by itself, sure the game was done.

The run Command Code 1.50.1 · commandcode/qwen/qwen3.7-plus

Time

3 min

Tokens

141k 82k read from cache · 15k written

Cost

$0.07 at alibaba/qwen3.7-plus list price; paid through a flat-rate plan, so not a charge

Work

5 steps 1 session · 4 tool calls

Claude’s note. No effort setting for this model on Command Code; run once at the CLI's default. One HTML file with three.js from a CDN, vendored for the page.

The run Command Code 1.50.1 · commandcode/gpt-5.6-luna#max

Time

3 min

Tokens

112k 60k read from cache · 22k written

Cost

$0.04 at openai/gpt-5.6-luna list price; paid through a flat-rate plan, so not a charge

Work

4 steps 1 session · 8 tool calls

Claude’s note. Max reasoning effort. A Vite project; built with vite build for the page.

The run Command Code 1.50.1 · commandcode/gpt-5.6-terra#max

Time

6 min

Tokens

184k 113k read from cache · 37k written

Cost

$0.55 at openai/gpt-5.6-terra list price; paid through a flat-rate plan, so not a charge

Work

6 steps 1 session · 8 tool calls

Claude’s note. GPT-5.6 Terra at max effort in Command Code: 6 minutes, 6 steps, a Vite project whose production build it checked, never opening the page.

The run Command Code 1.50.1 · commandcode/google/gemini-3.8-flash#high

Time

6 min

Tokens

1.19M 826k read from cache · 37k written

Cost

$0.44 at google/gemini-3.8-flash list price; paid through a flat-rate plan, so not a charge

Work

34 steps 1 session · 33 tool calls, 1 failed

Claude’s note. Gemini 3.8 Flash at high effort, the highest Command Code offers for it: 6 minutes, 34 steps, a Vite project it built but never opened.

The run Command Code 1.50.1 · commandcode/claude-opus-5#max

Time

1 h 50 min

Tokens

52.88M 52.10M read from cache · 409k written

Cost

$38.61 at anthropic/claude-opus-5 list price; paid through a flat-rate plan, so not a charge

Work

299 steps 1 session · 300 tool calls

Claude’s note. Claude Opus 5 at max effort in Command Code: 110 minutes, 299 steps, a Vite project it opened and drove again and again in the box's headless Chromium, tuning the canyon's rock strata from its own screenshots.

The run Command Code 1.50.1 · commandcode/google/gemini-3.6-flash

Time

2 min

Tokens

421k 304k read from cache · 14k written

Cost

$0.15 at google/gemini-3.6-flash list price; paid through a flat-rate plan, so not a charge

Work

18 steps 1 session · 17 tool calls

Claude’s note. Gemini 3.6 Flash at Command Code's default effort, as it refused max: 2 minutes, 18 steps, a Vite project it built but never opened.

The run Command Code 1.50.1 · commandcode/thinkingmachines/inkling

Time

31 s

Tokens

73k 52k read from cache · 7k written

Cost

$0.08 at thinkingmachines/thinkingmachines/Inkling list price; paid through a flat-rate plan, so not a charge

Work

4 steps 1 session · 3 tool calls

Claude’s note. Inkling in Command Code, which has no effort setting for it: 31 seconds, 4 steps, one index.html on three.js r160, which has no WebGPU renderer, so the page always takes its WebGL fallback; never opened.

The run Command Code 1.50.1 · commandcode/qwen/qwen3.7-max

Time

2 min

Tokens

54k 13k read from cache · 14k written

Cost

$0.18 at alibaba/qwen3.7-max list price; paid through a flat-rate plan, so not a charge

Work

2 steps 1 session · 1 tool call

Claude’s note. Qwen3.7 Max in Command Code, which has no effort setting for it: 2 minutes, 2 steps, one index.html written in a single call with three.js r170 from jsDelivr, never opened.

The run Command Code 1.50.1 · commandcode/gpt-5.6-sol#max

Time

8 min

Tokens

661k 573k read from cache · 42k written

Cost

$1.29 at openai/gpt-5.6-sol list price; paid through a flat-rate plan, so not a charge

Work

18 steps 1 session · 30 tool calls, 1 failed

Claude’s note. GPT-5.6 Sol at max effort in Command Code: 8 minutes, 18 steps, a Vite project it built and served; it looked for a Chromium on the PATH, did not find the box's own, and checked the page with curl only.

The run Command Code 1.50.1 · commandcode/claude-sonnet-5#max

Time

2 h 2 min

Tokens

104.21M 103.32M read from cache · 457k written

Cost

$26.31 at anthropic/claude-sonnet-5 list price; paid through a flat-rate plan, so not a charge

Work

407 steps 1 session · 406 tool calls, 13 failed

Claude’s note. Claude Sonnet 5 at max effort in Command Code: 122 minutes, 407 steps, a Vite project with three.js r186 that it opened and drove in the box's headless Chromium before it stopped.

The run Command Code 1.50.1 · commandcode/moonshotai/kimi-k3#max

Time

15 min

Tokens

514k 398k read from cache · 44k written

Cost

$0.99 at moonshotai/kimi-k3 list price; paid through a flat-rate plan, so not a charge

Work

9 steps 1 session · 11 tool calls

Claude’s note. It worked for 15 minutes over 9 steps, took 0.99 of the subscription's credits, and wrote five files: terrain with lakes, beaches, snowy ridges and pine forest, a paved road with markings and posts, a detailed jeep and seven ring checkpoints over timed laps. It served the files and checked the world code in Node, but never opened the page in a browser. On WebGPU nothing is drawn: an instanced mesh claims two more posts than its buffer holds and Chrome rejects every frame. Its WebGL2 fallback only starts where WebGPU is missing, and there the game runs at 60 fps, though the GO! banner never clears.

The run OpenCode 2.0.12 · openrouter/inclusionai/ling-3.0-flash-fin

Time

36 s

Tokens

8k 384 read from cache · 2k written

Cost

$0.00 pay-per-token, as OpenCode priced it

Work

1 step 1 session · 0 tool calls

Claude’s note. It worked for 36 seconds in a single step and made no tool call. It planned the game in its reasoning and then gave its whole answer there too, a short HTML page in a code block, so OpenCode recorded no reply; nothing was written to disk. It closed by asking whether to expand any part of it, and the run ended there, with OpenCode reporting success. The plumbing check before it, which asks for one file, passed: this model can call tools here.

The run OpenCode 2.0.12 · openrouter/ibm-granite/granite-4.2-8b#high

Time

3 min

Tokens

103k 71k read from cache · 13k written

Cost

$0.01 pay-per-token, as OpenCode priced it

Work

6 steps 1 session · 5 tool calls

Claude’s note. It worked for two and a half minutes over six steps. It wrote the whole game into one index.html, loading three.js from jsDelivr, read the file back, named three mistakes in it — an undefined size, a wrong await around rendering, a geometry class three.js no longer ships — made two edits, said it would fix and simplify the rest, and ended its turn there. It never opened the page in a browser, and the page it left does not run.

The run OpenCode 2.0.12 · openrouter/inception/mercury-2.5#high

Time

18 s

Tokens

46k 0 read from cache · 5k written

Cost

$0.00 pay-per-token, as OpenCode priced it

Work

4 steps 1 session · 3 tool calls, 1 failed

Claude’s note. It finished in 18 seconds over four steps: it wrote index.html and a README in two calls, then tried to preview the page through htmlpreview.github.io at an address it made up, which failed, and stopped there. It never started a server or opened a browser, and its closing summary describes terrain, a road, trees and physics that the page never reaches.

The run OpenCode 2.0.12 · openrouter/upstage/solar-mini4#max

Time

1 h 15 min

Tokens

41.70M 39.91M read from cache · 225k written

Cost

$0.32 pay-per-token, as OpenCode priced it

Work

223 steps 1 session · 221 tool calls, 4 failed

Claude’s note. It worked on its own for 75 minutes over 223 steps; a connection reset from OpenRouter eight minutes in was retried without a break. It split the game into modules — noise, terrain, track, decor, vehicle — and checked their numbers with a node script: track length, height range, how many trees and rocks. It then installed Playwright and Chromium in its own home and spent most of its last half hour trying to see the page headless; the browser kept closing, and it put that down to the box having no graphics card. It finished with node --check on every file and a README saying the scene should draw in a real browser. It does not; the card above says what happens.

The run OpenCode 2.0.12 · openrouter/inclusionai/ling-3.0-flash-vl

Time

5 min

Tokens

1.00M 771k read from cache · 30k written

Cost

$0.03 pay-per-token, as OpenCode priced it

Work

32 steps 1 session · 32 tool calls, 3 failed

Claude’s note. It worked on its own for 5 minutes over 32 steps. It wrote the whole game as one index.html in its first step, taking three.js r164 from unpkg; its two tries at opening the page in a browser failed, so it probed unpkg for the addon files it wanted, reworked the renderer set-up and checked the script with node. On a card the page asks the addons folder for a WebGPURenderer that is not exported there, falls back to WebGL2, and takes about forty seconds to build its world: spiky mountains around a lake, pines, rocks and a ring of checkpoint hoops. The jeep never comes into view and crawls at 20 km/h.

The run OpenCode 2.0.12 · openrouter/deepseek/deepseek-v4-flash-vision-exp#max

Time

1 h 45 min

Tokens

54.88M 50.99M read from cache · 297k written

Cost

$1.34 pay-per-token, as OpenCode priced it

Work

242 steps 1 session · 259 tool calls, 2 failed

Claude’s note. It worked on its own for 105 minutes over 242 steps. At 42 minutes OpenRouter dropped the connection once; OpenCode carried on to a finished game, but still exited with code 1. It split the game into modules, with a heightmap in level-of-detail chunks and a road carved into it, installed Chromium in the box, wrote headless tests that go from the menu through a race to the finish, fixed three bugs they found, and copied three.js into vendor/. OpenCode's session export was cut short the first time it was read, so the run was summarised again. On a card: foggy mountains, pine forest, three rivals and ten checkpoint gates, on a pebbled ground that reads as rippling water.

The run OpenCode 2.0.12 · openrouter/qwen/qwen3.8-flash

Time

56 min

Tokens

12.84M 12.45M read from cache · 120k written

Cost

$0.30 pay-per-token, as OpenCode priced it

Work

138 steps 1 session · 137 tool calls, 2 failed

Claude’s note. It worked on its own for 56 minutes over 138 steps. The game was written in the first quarter of an hour; it then checked the world and the physics with a node test whose autopilot drives all three laps, installed Chromium in the box and drove the page on the WebGL2 fallback, as the box has no graphics card, and at the end copied three.js into vendor/ so the page needs no network. On a card it takes the WebGPU path: hills, ridged mountains, lakes and pine forest, a dirt road pressed into the ground, ring checkpoints over three laps, a minimap and an engine sound. There are no rivals.

The run OpenCode 2.0.12 · selfhosted/aliceai-foundation-80b-a3b-base

Time

4 min

Tokens

501k 0 read from cache · 27k written

Cost

none per token the weights ran on a rented GPU, paid by the hour

Work

19 steps 1 session · 18 tool calls, 2 failed

Claude’s note. AliceAI-Foundation is a base model with no chat or tool tuning, so the bench reached it through a small adapter: each request was rendered with Yandex’s own agent-trajectory template and sent to vLLM as plain text, and the tool calls the model wrote were handed back to OpenCode as calls. It worked for four and a half minutes over 19 steps: it wrote two near-identical pages, index.html and game.html, loading three.js from a CDN, then an Express server for them, fetched the page once with curl, and stopped there without opening it in a browser. Neither page plays; the card above says what each one shows.

The run OpenCode 2.0.12 · openrouter/stealth/space-bunny-alpha#max

Time

10 min

Tokens

8.29M 8.14M read from cache · 74k written

Cost

$0 a free model

Work

76 steps 1 session · 88 tool calls, 6 failed

Claude’s note. Second run: the first broke off at step 42 when OpenRouter took the stealth model offline, and the brief was run again from scratch. It worked for 10 minutes on its own, began installing a browser to test the page, stopped that, and checked its code with node and a local server instead. The result is a low-poly desert circuit on WebGPURenderer with mesas, rocks, cacti, a track laid over the dunes, checkpoints, three laps and a polished HUD with a minimap. The two rivals it wrote never appear, and their distance reads NaN; there is no sound.

The run Codex 0.155.1 · openai/gpt-6-luna#max

Time

18 min

Tokens

2.02M 1.85M read from cache · 61k written

Cost

$0.06 at openai/gpt-6-luna list price; paid through a flat-rate plan, so not a charge

Work

31 steps 1 session · 26 tool calls, 2 failed

Claude’s note. Finished in 18 minutes in Codex after reading three.js's WebGPU docs on the web; it built the game but never opened it in a browser. A desert rally among buttes and scattered boulders with three timed laps, checkpoints, a minimap, pause and sound, but no rivals, and the chase camera passes through roadside poles.

The run Codex 0.155.1 · openai/gpt-6-sol#high

Time

13 min

Tokens

1.79M 1.68M read from cache · 37k written

Cost

$0.85 at openai/gpt-6-sol list price; paid through a flat-rate plan, so not a charge

Work

43 steps 1 session · 35 tool calls, 2 failed

Claude’s note. Finished in 13 minutes in Codex and checked its own build in headless Chrome, desktop and phone. A point-to-point desert sprint past red mesas, rocks and cacti against three rival jeeps, six checkpoints and nitro, with a polished HUD and sound. The world is thinner than the best here and the course is a single stretch, but it runs smoothly on the fallback.

The run Claude Code 2.1.280 · anthropic/claude-opus-5-5#xhigh

Time

1 h 37 min

Tokens

34.31M 33.54M read from cache · 327k written

Cost

$16.80 flat-rate plan: Claude Code’s price for the tokens, not a charge

Work

132 steps 1 session · 131 tool calls, 4 failed

Claude’s note. Ran for 97 minutes unattended in Claude Code and finished on its own, testing as it went with a headless race simulator it wrote. The result is the richest world in the gallery: an eroded heightmap built in a worker, a river and lake with two fords, forest, grass, a crowd at the start, five rivals, three laps and a menu for quality, laps and a random world. On this GPU-less machine the WebGL2 fallback draws well under a frame a second, so the driving was judged from the code.

The run OpenCode 2.0.12 · openrouter/z-ai/glm-5.3#max

Time

13 min

Tokens

3.43M 3.13M read from cache · 105k written

Cost

$0.94 pay-per-token, as OpenCode priced it

Work

38 steps 1 session · 52 tool calls

Claude’s note. Finished on its own in 13 minutes, no tool call failed: it fetched three.js r170, wrote nine modules, then, finding OpenCode's browser not connected, set up headless Chrome itself, checked frames pixel by pixel, drove the WebGL2 fallback and scripted a full three-lap race. The build runs as shipped: rolling hills with snow on the rim, lakes, a carved dirt track with ruts, pines, rocks and grass, eleven checkpoints, a minimap and engine sound. Here it reached 89 km/h on the WebGL2 fallback; the WebGPU path cannot be seen on this machine, which has no GPU. Flaws: the dust puffs are sliced flat where they meet the ground, and the minimap is very small.