JourneyBench (Static-Prompt Agent): leaderboard
JourneyBench (Balaji et al., 2026) tests whether customer-support LLM agents follow a business standard operating procedure in simulated conversations with a GPT-4o user: 703 journeys over procedure graphs for e-commerce, loan applications and telecommunications. In this setting the whole standard operating procedure is compiled into one static system prompt. User Journey Coverage Score from 0 to 1, mean of the three domains (a conversation scores 0 on any missing, extra or misordered tool call and otherwise its share of correct tool parameters); higher is better.
Metric: User Journey Coverage Score (0-1; a conversation scores 0 on any missing, extra or misordered tool call, otherwise its tool-parameter accuracy; GPT-4o simulated user, 40-turn limit; mean of the three domains; the whole SOP graph compiled into one static system prompt). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 0.56 | #741 |
| 2 | GPT-4o Mini | 0.44 | #1627 |
| 3 | Claude 3.5 Haiku | 0.25 | #1563 |
| 4 | Llama 3.3 70B Instruct | 0.25 | #1353 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=journeybench-static-prompt-agent · How It Works · Data refreshed daily, snapshot 2026-11-12.