The same task — build a Terraria-style 2D web sandbox game — was handed to four agent × model pairs. Each ran fully autonomously with zero human course-correction. Here is what they produced and what it cost.
| Claude CodeFable 5 | CodexGPT-5.6 | DeepSeek harnessV4 pro | DeepSeek harnessV4 flash vision |
|---|
| Model | Claude Fable 5 (max) | GPT-5.6 (sol max) | DeepSeek V4 pro (max) | DeepSeek V4 flash vision exp |
|---|
| Build | Forest Sandbox | Mossreach | TerraForest | Forest Sandbox |
|---|
| Stack | Vanilla Canvas + JS
zero-dependency static page | React + vinext (Next.js)
SSR app (statically exported here) | Vanilla Canvas + JS
zero-dependency static page | Vanilla Canvas + JS
zero-dependency static page (procedural art) |
|---|
| Agent time ¹ | ≈ 34 min | 27 min 55 s | ≈ 26 min 29 s | ≈ 26 min 22 s |
|---|
| Input tokens ² | 5.70 M (97.6% cached) | 13.22 M (97.4% cached) | 3.20 M (100% cached) | 9.1 M (99.7% cached) |
|---|
| Output tokens | 104.9 K | 50.0 K | 146.4 K | 160 K |
|---|
| Interaction | 1 prompt · 43 API round-trips | 1 prompt | 1 prompt · 1 turn, 26 steps | 1 prompt · 1 turn, 78 steps |
|---|
How these numbers are measured:
¹ Agent busy time, from task start to delivery, with user idle time excluded. Claude Code also sat through a 15-minute API request timeout — a platform fault, not counted. Of Codex's 27 min 55 s, roughly 10 minutes went to deployment troubleshooting on its hosting platform, leaving about 18 minutes of development. The two DeepSeek runs are LLM time plus tool-call time.
² Raw counts as reported by each platform. Cache discounts and output pricing differ between vendors, so token counts do not translate directly into cost.
The Claude Code and DeepSeek V4 pro builds shipped Chinese in-game text; for this English edition their display strings and comments are translated at build time, while game logic, layout and art are untouched. A subjective evaluation of the results (visuals / feel / completeness) will follow.