AstroAlertBench: leaderboard
Five-class classification of Zwicky Transient Facility alerts (bogus artifact, asteroid, supernova, active galactic nucleus, variable star) from the science, reference and difference image cutouts plus alert metadata, answered as a staged decision; AstroAlertBench (Chen et al., 2026). End-to-end accuracy in percent over 1,500 alerts with ALeRCE labels, 300 per class, where an unparsed answer counts as wrong; higher is better.
Metric: End-to-end five-class accuracy (%; 1,500 ZTF alerts, 300 each of bogus, asteroid, supernova, AGN and variable star; image triplet plus alert metadata; unparsed answers count as wrong). Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 (High) | 60.6 | #75 (Claude Opus 4.7) |
| 2 | GPT-5.4 (High) | 51.07 | #121 (GPT-5.4) |
| 3 | Kimi K2.5 (Thinking) | 49.34 | #252 (Kimi K2.5) |
| 4 | Claude Opus 4.7 (Non-reasoning) | 48.87 | #75 (Claude Opus 4.7) |
| 5 | Qwen 3.5 397B A17B (Thinking) | 44.27 | #242 (Qwen 3.5 397B A17B) |
| 6 | GPT-5.4 (Non-reasoning) | 43.67 | #121 (GPT-5.4) |
| 7 | Gemini 2.5 Pro | 41.93 | #262 |
| 8 | Gemini 2.5 Flash (Non-reasoning) | 36.27 | #452 (Gemini 2.5 Flash) |
| 9 | Qwen 3.5 397B A17B (Non-reasoning) | 34.98 | #242 (Qwen 3.5 397B A17B) |
| 10 | Qwen 3.5 35B A3B (Thinking) | 26.73 | #480 (Qwen 3.5 35B A3B) |
| 11 | Qwen 3.5 35B A3B (Non-reasoning) | 25.5 | #480 (Qwen 3.5 35B A3B) |
| 12 | Qwen 3.5 4B (Non-reasoning) | 22.32 | #1180 (Qwen 3.5 4B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=astroalertbench · How It Works · Data refreshed daily, snapshot 2026-11-12.