LOCA-bench: leaderboard
Metric: Mean task success rate (%) over the seven environment description lengths (8K to 256K tokens, 75 instances each, 525 in all); 15 seed agentic tasks adapted from Toolathlon with five seeded environment states each, on local mock servers (Canvas, email, calendar, BigQuery, Snowflake, WooCommerce, spreadsheets and more); ReAct scaffold, each model at its maximum context window with the oldest context truncated beyond it; success checked by rule-based scripts on the final environment state; higher is better. Source: arxiv.org. Saturation forecast: Around April 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (Medium) | 51.2 | #105 (GPT-5.2) |
| 2 | DeepSeek V3.2 (Thinking) | 42.7 | #198 (DeepSeek V3.2) |
| 3 | Gemini 3 Flash | 38.3 | #93 |
| 4 | MiniMax-M2.1 | 35.8 | #283 |
| 5 | GLM-4.7 | 35.2 | #185 |
| 6 | Kimi K2 (Thinking) | 31.2 | #236 (Kimi K2) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=loca-bench · How It Works · Data refreshed daily, snapshot 2026-10-11.