Benchmarks / Reasoning
Hearing Benchcommunity
Models see a spectrogram image of a short sound clip and name the sound, with no tools. 100 clips across five tiers, from everyday sounds to spoken sentences.
Hearing Bench
Higher is better · best effort per model · Cost per clip at OpenRouter API list prices, 13 Oct 2026. Runs used each lab's own CLI harness.
full table below
| # | Model | Effort | Score |
|---|---|---|---|
| 1 | Gemini 4 | high | 44.0% |
| 2 | Opus 5.5 | xhigh | 43.0% |
| 3 | Gemini 3.8 Flash | high | 29.0% |
| 4 | GPT-6.1 Sol | xhigh | 26.0% |
| 5 | GPT-6 Astra | xhigh | 22.0% |
| 6 | Sonnet 5.5 | xhigh | 20.0% |
| 7 | Grok 4.7 | high | 17.0% |
| 8 | Haiku 5.5 | medium | 9.0% |
| 9 | GPT-6 Luna | xhigh | 9.0% |