ComboShoppingBench - Claim Faithfulness: leaderboard
ComboShoppingBench (arXiv 2608.09282v1, Table 2): 22 tool-using agent configurations, 11 models each with thinking enabled and disabled, on 291 compositional shopping tasks over a frozen catalog with coupons and budgets. Share of responses whose settlement claims pass all four faithfulness criteria, judged by Gemini-3.1-Pro-Preview, also an evaluated agent; percent, higher is better.
Metric: Pass rate (%; LLM-judged). Source: arxiv.org. Saturation forecast: Estimated already saturated. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 (Non-reasoning) | 94.8 | #67 (Claude Opus 4.6) |
| 2 | Claude Opus 4.6 (Thinking) | 94.5 | #67 (Claude Opus 4.6) |
| 3 | Claude Opus 4.8 (Thinking) | 94.5 | #39 (Claude Opus 4.8) |
| 4 | GPT-5.5 | 90.4 | #33 |
| 5 | Gemini 3.1 Pro (Preview) (Thinking) | 88 | #64 (Gemini 3.1 Pro (Preview)) |
| 6 | Claude Sonnet 4.6 (Thinking) | 83.2 | #94 (Claude Sonnet 4.6) |
| 7 | Qwen 3.6 27B (Non-reasoning) | 81.1 | #182 (Qwen 3.6 27B) |
| 8 | Claude Opus 4.8 (Non-reasoning) | 78.7 | #39 (Claude Opus 4.8) |
| 9 | Kimi K2.6 (Thinking) | 78.4 | #113 (Kimi K2.6) |
| 10 | GLM-5.2 (Non-reasoning) | 78 | #80 (GLM-5.2) |
| 11 | Claude Sonnet 4.6 (Non-reasoning) | 77 | #94 (Claude Sonnet 4.6) |
| 12 | GPT-5.5 (Non-reasoning) | 70.8 | #33 (GPT-5.5) |
| 13 | DeepSeek V4 Pro (Non-reasoning) | 68.4 | #112 (DeepSeek V4 Pro) |
| 14 | MiniMax-M3 (Thinking) | 68 | #151 (MiniMax-M3) |
| 15 | Qwen 3.6 27B (Thinking) | 61.5 | #182 (Qwen 3.6 27B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=comboshoppingbench-claim-faithfulness · How It Works · Data refreshed daily, snapshot 2026-11-05.