AlpsBench - Update: leaderboard
Metric: Semantic-matching F1 (0-1, times 100) of the memories the model outputs after reading a historical and a new dialogue against the human-annotated memories, matched one-to-one by a DeepSeek-V3.2 judge; AlpsBench's 2,500 human-verified instances per task built from long-term real user-LLM dialogues in WildChat; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 77.93 | #105 |
| 2 | Gemini 3 Flash (Preview) | 76.86 | #78 |
| 3 | Claude Sonnet 4.5 | 75.43 | #138 |
| 4 | DeepSeek V3.2 (Thinking) | 75.43 | #198 (DeepSeek V3.2) |
| 5 | Qwen 3 Max | 71.9 | #201 |
| 6 | Llama 4 Maverick | 68.24 | #451 |
| 7 | GPT-4.1 Mini | 62.14 | #346 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=alpsbench-update · How It Works · Data refreshed daily, snapshot 2026-10-11.