EdgeBench - Games: leaderboard
Interactive games tasks: mean best-so-far score (0-100) after a 12-hour continuous agent run with evaluator feedback, nominally three trials per task with only valid trajectories scored, averaged over the family's tasks; higher is better; harness/context: Claude Code 200k Source: EdgeBench paper Table 2 category score after 12 hours. Harness and compaction window are separate; unsuffixed category boards use Claude Code 200k. Published scores average valid trajectories under the source missing-case policy.
Metric: Interactive games tasks: mean best-so-far score (0-100) after a 12-hour continuous agent run with evaluator feedback, nominally three trials per task with only valid trajectories scored, averaged over the family's tasks; higher is better; harness/context: Claude Code 200k. Source: arxiv.org. Saturation forecast: Around November 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.1 | 29.3 | #179 |
| 2 | DeepSeek V4 Pro Preview | 16.9 | #227 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=edgebench-games · How It Works · Data refreshed daily, snapshot 2026-11-13.