TRIP-Bench - Overall - Loose: leaderboard
Evaluates long-horizon interactive travel-planning agents on TRIP-Bench with 18 tools. Both scoring modes require zero feasibility failures. Loose success permits at most two planning-soundness failures and one user-constraint failure; strict success requires zero failures in all three groups. Explicit thinking and non-thinking model labels and dated checkpoints are retained. These are two metrics of the same evaluated task population. This series reports Loose Success (%).
Metric: Loose Success (%). Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (Thinking) | 45 | #115 (GPT-5.2) |
| 2 | DeepSeek V3.2 (Thinking) | 40 | #210 (DeepSeek V3.2) |
| 3 | Claude Sonnet 4.5 (Thinking) | 32 | #148 (Claude Sonnet 4.5) |
| 4 | GLM-4.7 (Thinking) | 20.3 | #202 (GLM-4.7) |
| 5 | DeepSeek V3.2 (Non-reasoning) | 18.5 | #210 (DeepSeek V3.2) |
| 6 | Claude Sonnet 4.5 (Non-reasoning) | 17.3 | #148 (Claude Sonnet 4.5) |
| 7 | GLM-4.7 (Non-reasoning) | 14.8 | #202 (GLM-4.7) |
| 8 | GPT-5.2 (Non-reasoning) | 13.3 | #115 (GPT-5.2) |
| 9 | Kimi K2 (Thinking) | 10.8 | #252 (Kimi K2) |
| 10 | Qwen 3 235B A22B 2507 Instruct | 5.8 | #315 |
| 11 | Kimi K2 0905 | 3.3 | #292 |
| 12 | Qwen 3 235B A22B 2507 (Thinking) | 0 | #263 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=trip-bench-overall-loose · How It Works · Data refreshed daily, snapshot 2026-11-05.