Agent Arena - Tool Hallucination: leaderboard
Agent Arena submetric measuring tool hallucination rate; lower values indicate fewer invented or invalid tool uses.
Metric: Tool Hallucination Effect (percentage points vs baseline; lower is better). Source: arena.ai. Saturation forecast: Around March 2028. 51 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K3 (Max) | -0.51 | #30 (Kimi K3) |
| 2 | GPT-6.1 Sol (Max) | -0.51 | #7 (GPT-6.1 Sol) |
| 3 | Claude Opus 5 (Max) | -0.5 | #12 (Claude Opus 5) |
| 4 | Step 5 Preview | -0.5 | #44 |
| 5 | GLM-5.2 (Max) | -0.48 | #73 (GLM-5.2) |
| 6 | Gemini 3.6 Flash (High) | -0.48 | #51 (Gemini 3.6 Flash) |
| 7 | Claude Fable 5 (High) | -0.47 | #16 (Claude Fable 5) |
| 8 | Claude Fable 5.1 (Max) | -0.46 | #9 (Claude Fable 5.1) |
| 9 | Grok 4.5 | -0.45 | #48 |
| 10 | GLM-5.3 (Max) | -0.45 | #39 (GLM-5.3) |
| 11 | GPT-6 Astra (Max) | -0.43 | #6 (GPT-6 Astra) |
| 12 | Claude Sonnet 5.5 (Max) | -0.43 | #11 (Claude Sonnet 5.5) |
| 13 | GLM-5.3 Flash | -0.42 | #65 |
| 14 | Muse Spark 1.3 (Max) | -0.4 | #24 (Muse Spark 1.3) |
| 15 | DeepSeek V4.1 Flash (Max) | -0.37 | #36 (DeepSeek V4.1 Flash) |
No result here: #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=agent-arena-tool-hallucination · How It Works · Data refreshed daily, snapshot 2026-10-20.