Model × Game Generation 2026.11.19

Sunagaku Labs — Model × Game Generation

DopaGaki App Comparison

Play six games and compare the instructions behind them: five models received the same one-line prompt, while Fable 5.2 built the earlier DOPA RUSH from its own brief.

PROMPTShared by all five models, nothing else

ドパガキが楽しめるようなアプリを作って!

Roughly: “Make an app that dopamine-hooked kids can enjoy!” (“Dopagaki” is Japanese slang for someone hooked on quick dopamine hits.) No follow-up instructions. Each model ran in its own folder without seeing the others’ output.

Six games and the models that made them

  1. Generated by

    Gemini 4 Neon

    Gemini CLI 0.68.0

    Generated app

    DOPA-NEON // 脳汁10000%オーバードライブ

    A neon-core tapping clicker. Hit the shrinking ring for a JUST bonus (x5), fill the gauge for FEVER (x5), then keep tapping for HYPER OVERDRIVE (x25). The beat is synthesized live with Web Audio.

    Prompt given
    ドパガキが楽しめるようなアプリを作って!
    Time
    647 s (10:47)
    Estimated cost
    about $0.99
    Stack
    Plain HTML, CSS and JS (no dependencies)
    Play it here (Japanese) →
  2. Generated by

    Gemini 4 Carbon

    Gemini CLI 0.68.0

    Generated app

    DOPA-MAX 9999 ⚡ ドパガキ無限脳汁ランド

    Four screens run at once: rapid-fire short videos with comment barrages, ASMR soap slicing with a bouncing DVD logo, slots and gacha. Switch layouts and take it all in.

    Prompt given
    ドパガキが楽しめるようなアプリを作って!
    Time
    542 s (9:02)
    Estimated cost
    about $0.72
    Stack
    React + Vite + TypeScript
    Play it here (Japanese) →
  3. Generated by

    Gemini 4 Hydrogen

    Gemini CLI 0.68.0

    Generated app

    脳汁オーバーロードDX / DOPAMINE OVERDRIVE

    An arcade clicker with gacha: tap the core, trigger a rainbow FEVER TIME, hold to auto-tap, prestige, and enjoy small side gimmicks.

    Prompt given
    ドパガキが楽しめるようなアプリを作って!
    Time
    307 s (5:07)
    Estimated cost
    about $0.83
    Stack
    React + Vite
    Play it here (Japanese) →
  4. Generated by

    GPT 6.2 Astra

    Codex CLI 0.167.2

    Generated app

    Beat Beast — ビートビースト

    A six-pad drum machine: each hit plays a kick, snare and more, and a pixel-art beast bounces along. A memory challenge has you repeat the lit sequence. On desktop, play with A S D / J K L.

    Prompt given
    ドパガキが楽しめるようなアプリを作って!
    Time
    42 s
    Estimated cost
    about $0.59 (at standard API rates)
    Stack
    Three static files (HTML, CSS, JS)
    Play it here (Japanese) →
  5. Generated by

    GPT Sol 6.2

    Codex CLI 0.167.2

    Generated app

    DOPA RUSH!(30秒タップアーケード)

    A 30-second tap arcade: hit the lit panels to chain combos, and ten in a row starts a five-second double-score fever. Number keys 1 to 9 work too.

    Prompt given
    ドパガキが楽しめるようなアプリを作って!
    Time
    88 s (1:28)
    Estimated cost
    about $0.35 (at standard API rates)
    Stack
    Static files with a font and an image
    Play it here (Japanese) →
  6. Generated by

    Fable 5.2

    Claude Code

    Earlier work · 2026.11.10

    Generated app

    DOPA RUSH

    Dodge the grey boredom horde in a five-minute action run. Attacks fire automatically as you move, collect drops and choose upgrade cards. Chain combos to trigger fever mode.

    Prompt given
    ドーパミンましましで、ずっと遊びたくなるゲームを
    Time
    Not recorded
    Estimated cost
    Not recorded
    Stack
    Canvas, TypeScript and Web Audio

    The brief in English: “Make a dopamine-packed game that makes you want to keep playing.”

    How to play and build notes
    • Movement is the only control: one finger on a phone, WASD or arrow keys on a desktop
    • Seven weapons and twelve passives are drawn as cards with rarity tiers, so each run builds differently
    • No image or audio files: characters, sound effects, and music are all generated in code
    • A setting that tones down flashes and screen shake, and a break reminder after 30 minutes of play
    • Designed a run of about five minutes around immediate feedback for defeating, collecting, and growing stronger.
    • Kept the simulation separate from rendering and tuned difficulty and pacing with seeded bots that play full runs.
    • Drew every Canvas 2D sprite once at start-up so scenes with more than 200 enemies stay light.
    • Synthesised sound effects and music with Web Audio; the defeat sound rises in pitch as a combo continues.
    • Progress and settings stay in the browser. Nothing is sent to a server.
    Play it here (Japanese) →
Switch models and play (Japanese) →

Comparison table

Five games from the shared prompt and one earlier game with its own brief
ModelGenerated appPrompt givenTimeEstimated costPlay
Gemini 4 Neon DOPA-NEON // 脳汁10000%オーバードライブ ドパガキが楽しめるようなアプリを作って! 647 s (10:47) about $0.99 Play →
Gemini 4 Carbon DOPA-MAX 9999 ⚡ ドパガキ無限脳汁ランド ドパガキが楽しめるようなアプリを作って! 542 s (9:02) about $0.72 Play →
Gemini 4 Hydrogen 脳汁オーバーロードDX / DOPAMINE OVERDRIVE ドパガキが楽しめるようなアプリを作って! 307 s (5:07) about $0.83 Play →
GPT 6.2 Astra Beat Beast — ビートビースト ドパガキが楽しめるようなアプリを作って! 42 s about $0.59* Play →
GPT Sol 6.2 DOPA RUSH!(30秒タップアーケード) ドパガキが楽しめるようなアプリを作って! 88 s (1:28) about $0.35* Play →
Fable 5.2 DOPA RUSH ドーパミンましましで、ずっと遊びたくなるゲームを Not recorded Not recorded Play →
  • Time runs from CLI start to exit. The connectivity check beforehand is excluded.
  • Gemini runs were billed directly through the API. Costs assume the same launch rates for all three models: $2 input, $0.10 cached input and $10 output (including thinking) per million tokens.
  • * The GPT runs used a Codex plan, so nothing was billed directly. These are reference values at standard API rates; with the Ultrafast service tier, actual plan usage differs.

How the comparison was made

  1. 01Placed the same PROMPT.md in five folders and let each model generate independently with Gemini CLI or Codex CLI.
  2. 02Recorded time, token counts and exit codes from the run logs, and estimated cost from published rates.
  3. 03Left the generated code unmodified; only the two React + Vite apps were built with relative paths before being placed on the site.
  4. 04Checked in real desktop and mobile browsers that all five games start and respond to input inside this site.

Run conditions for the five shared-prompt models

  • The only prompt was the one Japanese sentence above. The same PROMPT.md (identical SHA-256) sat in all five folders, and no further instructions were given.
  • The three Gemini 4 models ran on Gemini CLI 0.68.0 (gemini -m <model> --yolo --skip-trust --output-format json, API key authentication).
  • The two GPT 6.2 models ran on Codex CLI 0.167.2 (codex exec --json --skip-git-repo-check -m <model>, Ultrafast service tier, default medium reasoning).
  • The two GPT runs started together at 10:26:36 and the three Gemini runs at 10:27:30 (JST). Every model exited with code 0.
  • The generated code is unmodified. For the two React + Vite apps (Carbon and Hydrogen), the files here are builds made with vite build --base ./ so they load from a folder on this site.
Token breakdown
ModelInput (uncached)Cached inputOutputThinkingAPI requestsEstimated cost
Gemini 4 Neon 153,485 652,489 57,223 4,517 12 requests, 0 errors about $0.99
Gemini 4 Carbon 110,881 911,947 34,478 5,899 20 requests, 0 errors about $0.72
Gemini 4 Hydrogen 124,625 1,024,325 41,176 6,378 22 requests, 0 errors about $0.83
ModelInput (cached)OutputRates per million tokensAt standard rates
GPT 6.2 Astra 214,169 (179,584) 6,121 $8 input / $0.40 cached / $40 output about $0.59
GPT Sol 6.2 971,018 (904,320) 13,087 $2 input / $0.10 cached / $10 output about $0.35

What happened during the runs

  • The first Gemini launch exited immediately for all three models because the run script was missing a PATH setting, so nothing was generated. The results here come from the rerun after fixing it.
  • During its run, GPT Sol 6.2 read the launch script (run-codex.sh) outside its working folder once. It did not read any other model’s output.
  • Gemini 4 Neon first set up a Vite project, then ended with dependency-free HTML, CSS and JS.
  • GPT Sol 6.2’s hero image was made with Codex image generation, and its font (Barlow Condensed) ships with its Open Font License text.

Reading the results

  • Each model generated once. This is not a ranking of the models.
  • GPT Sol 6.2’s “DOPA RUSH!” is unrelated to the Labs game “DOPA RUSH”, which was made from a different request. The names just happen to overlap.
  • Fable, Neon, Hydrogen, Astra and Sol save play records or settings in your browser. Nothing is sent to our servers.
Switch models and play (Japanese) →