Agent Planning Benchmark - Tool-Broken Recovery: leaderboard
The Agent Planning Benchmark (APB, 2026) isolates the planning step of multimodal tool-using agents: models write plans and tool calls for 4,209 cases drawn from agent datasets and real platform traffic, and Gemini 3 Pro judges each plan against a reference trajectory. This board gives the model a trajectory in which a critical tool has just failed and an alternative tool is available, and reports the percentage of cases in which it switches to that substitute; higher is better.
Metric: Replace rate (%): share of 300 cases in which, after a critical tool returns an error, the model switches to the provided substitute tool (the optimal behaviour, versus altering, retrying or refusing); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) (High) | 78 | #113 (Gemini 3 Pro (Preview)) |
| 2 | Qwen 3 VL 235B A22B Instruct | 77.7 | #522 |
| 3 | Claude Sonnet 4.5 | 76.7 | #250 |
| 4 | GPT-5 | 76.3 | #167 |
| 5 | Gemini 2.5 Pro | 74.7 | #266 |
| 6 | Qwen 3 VL 32B Instruct | 68.7 | #585 |
| 7 | Qwen 3 VL 30B A3B Instruct | 55.7 | #871 |
| 8 | Gemini 2.5 Flash (Non-reasoning) | 44.3 | #453 (Gemini 2.5 Flash) |
| 9 | GPT-4o | 42.3 | #741 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=agent-planning-benchmark-tool-broken-recovery · How It Works · Data refreshed daily, snapshot 2026-11-13.