AIMultiple - HALC-Bench: leaderboard
AIMultiple HALC-Bench: 204 trap claims about metrics absent from post-cutoff earnings-call transcripts, hidden among distractors at the beginning, middle and end of each model's usable context window. A trap scores only when the model answers 'not mentioned'; every item is a trap, so abstaining is never penalized. Mean of the three positions, %, higher is better.
Metric: Traps correctly answered 'not mentioned' (%, mean of three haystack positions). Source: aimultiple.com. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 | 100 | #20 |
| 2 | GPT-5.5 | 98.1 | #33 |
| 3 | Gemini 3.1 Pro (Preview) | 96.6 | #64 |
| 4 | Kimi K2.6 | 94.6 | #113 |
| 5 | MiniMax-M2.7 | 94.5 | #270 |
| 6 | Qwen 3.6 Plus | 92.2 | #124 |
| 7 | GLM-5.1 | 87.2 | #109 |
| 8 | Claude Opus 4.8 | 81.4 | #39 |
| 9 | Claude Sonnet 4.6 | 73 | #94 |
| 10 | Gemini 3.5 Flash | 67.2 | #66 |
| 11 | GPT-5.4 Mini | 33.3 | #228 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=aimultiple-halc-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.