The Forge compared with other AI benchmarks
No benchmark names "the best AI": each answers a different question. The Forge measures one precise thing: what a model can do in a 3D world through MCP tools, scored by a program.
· ScoreIA
| Benchmark | What it measures | Who or what scores | Read it for |
|---|---|---|---|
| The Forge (ScoreIA) | Building and animating a 3D knight through MCP tools, then fighting duels | A program (geometry, contacts, trajectories) | An agent acting through tools on a spatial task |
| LMArena | Preference between two answers, with models hidden | Human votes | Perceived quality in conversation |
| SWE-bench | Solving real GitHub issues | The project's tests | Fixing code |
| ARC-AGI | Solving unseen visual grid puzzles | The exact expected answer | Abstract reasoning |
| MLPerf | Training and inference speed of systems | Performance measurements | Choosing hardware, not a model |
The limits of The Forge
- It is a small bench: 9 models with 1 to 6 fights each. Under 3 fights, the placing stays exploratory.
- It is not independent: the ScoreIA team runs the campaign's models itself.
- Identities are declared and cards are not signed. The data and the replays are public.
- These descriptions summarise the other benchmarks' public methods; they do not rank them.
The AI ranking → · The method → · How the program referees →