GlobeBench - Task Success: leaderboard
GlobeBench - Task Success. GlobeBench measures next-observation prediction on 1,632 execution-grounded text-world items across five domains and 19 skills. All 20 models retain the full item denominator; provider refusals and empty or errored responses count as failures. Scoring uses wm-hybrid-deterministic-llm/2, wm-grader-1.5, and response-judge/4: deterministic certification first, then GPT-5.6 Terra with GPT-5.6 Sol for missing or uncertain judgments. The primary repository binds every model key and serving route. Most subjects nominally use high effort; Qwen 3.8 Flash Next nominally uses xhigh. Nonusable responses were retried, sometimes with lower effort, larger budgets, or different routes. Completed answers and refusals were never resampled. These are mixed serving settings, not a claim that every item used one effort. Reported metric: Task Success (%).
Metric: Task Success (%). Source: raw.githubusercontent.com. Saturation forecast: Around 2029. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 49.57 | #6 |
| 2 | Grok 4.6 | 28.13 | #27 |
| 3 | Claude Opus 5.5 | 26.35 | #5 |
| 4 | Qwen 3.8 Max | 25.31 | #32 |
| 5 | GPT-5.5 | 25.18 | #30 |
| 6 | GPT-6 Sol | 25 | #16 |
| 7 | GPT-5.6 Sol | 23.65 | #19 |
| 8 | Gemini 3.7 Flash | 21.75 | #22 |
| 9 | Claude Fable 5.1 | 21.63 | #9 |
| 10 | Claude Opus 4.8 | 20.04 | #37 |
| 11 | GPT-6 Luna | 17.77 | #61 |
| 12 | Claude Opus 5 | 15.93 | #13 |
| 13 | GLM-5.3 | 14.77 | #38 |
| 14 | Kimi K3 | 13.97 | #33 |
| 15 | Inkling | 11.89 | #136 |
No result here: #7 GPT-6.1 Sol, #8 Gemini 4 Argon, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=globebench-task-success · How It Works · Data refreshed daily, snapshot 2026-10-25.