InferenceBench: leaderboard
InferenceBench original v1.0 prompt: geometric-mean final-server speedup over PyTorch across four scenarios and two-hour runs. Invalid or regressing runs score the PyTorch baseline 1. Historical scaffold versions and effort settings are preserved.
Metric: Geometric Mean Speedup. Source: inferencebench.ai. Saturation forecast: Around October 2027. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 7.6 | #37 |
| 2 | Claude Opus 4.8 (xHigh) | 7.34 | #37 (Claude Opus 4.8) |
| 3 | GPT-5.4 (High) | 6.16 | #79 (GPT-5.4) |
| 4 | Claude Sonnet 4.6 | 5.56 | #90 |
| 5 | GPT-5.3 Codex (High) | 5.53 | #69 (GPT-5.3 Codex) |
| 6 | GPT-5.5 (xHigh) | 5.45 | #31 (GPT-5.5) |
| 7 | Gemini 3.1 Pro (Preview) | 4.92 | #58 |
| 8 | Kimi K2.6 | 4.51 | #104 |
| 9 | Claude Opus 4.6 | 4.38 | #63 |
| 10 | GPT-5.2 | 4.28 | #110 |
| 11 | GPT-5.5 (High) | 4.22 | #31 (GPT-5.5) |
| 12 | Gemini 3.5 Flash | 4.16 | #60 |
| 13 | Claude Opus 4.5 | 3.76 | #83 |
| 14 | GPT-5.1 Codex Max | 3.59 | #100 |
| 15 | GLM-5 | 3.22 | #141 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=inferencebench · How It Works · Data refreshed daily, snapshot 2026-10-20.