CoworkOS - τ3-Bench Banking: leaderboard
CoworkOS paper comparison of OpenHands, Codex, DeepSeek Harness and CoworkOS, all using GLM-5.3 High effort, 256K context and64K maximum output. Harness identities remain separate. Each benchmark uses its native scoring protocol; domain averages equally weight constituent benchmarks. GDPVal Elo uses all220 scheduled tasks, invalid/timeouts as losses, baseline1000; Productivity average substitutes exact-count tie-adjusted win rate for Elo. τ3-Bench Banking. 97 Banking Knowledge tasks, BM25/dense retrieval over698 documents, read-only shell,200 steps,1800-second simulation/2400-second outer limit; task pass rate.
Metric: Score (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.3 (high) (CoworkOS Run: harness=DeepSeek%20Harness; context=256K; output=64K) | 55.67 | |
| 2 | GLM-5.3 (high) (CoworkOS Run: harness=CoworkOS; context=256K; output=64K) | 49.48 | |
| 3 | GLM-5.3 (high) (CoworkOS Run: harness=Codex; context=256K; output=64K) | 46.39 | |
| 4 | GLM-5.3 (high) (CoworkOS Run: harness=OpenHands; context=256K; output=64K) | 40.2 |
Interactive version: theaggregate.ai/benchmark?slug=coworkos-3-bench-banking · How It Works · Data refreshed daily, snapshot 2026-11-05.