RSI-Inference Bench
Coding agents add model support to inference engines, improve serving speed and fix performance regressions. The harness builds and runs each submission.
Apple reports pass rates for model support and speedups for optimization and repairs. It combines the tracks in an Elo rating.
- Made by
- Apple
- Released
- October 2026
- Kind
- Machine learning engineering
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- Terminal
- Items from
- Real work and use
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Agents
- Strongest model tested
- GPT-6 Astra #4 on the capability index
- Last model added
- October 2026
- Scores refreshed
- November 20th, 2026
Claude Sonnet 5.5 leads with 2,044
Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
- 2,044 Claude Sonnet 5.5
- 1,903 Claude Opus 5.5
- 1,716 Claude Fable 5.1
- 1,685 GPT-6.1 Sol
- 1,676 GPT-6 Astra
- 1,676
Gemini 4 Argon
- 1,559 Claude Opus 5
- 1,448 GPT-6 Sol
- 1,405
DeepSeek V4.1 Flash 1017
- 1,383
MiniMax-M3.1-Flash
- 1,381 GLM-5.3
- 1,295
DeepSeek V4.1 Flash
- 1,259 Claude Opus 4.8
- 1,047
Gemini 3.8 Flash
- 1,022
Gemini 3.7 Flash
Bars start at 1,000.
- 1 Claude Sonnet 5.5 2,044
- 2 Claude Opus 5.5 1,903
- 3 Claude Fable 5.1 1,716
- 4 GPT-6.1 Sol 1,685
- 5 GPT-6 Astra 1,676
- 6
Gemini 4 Argon 1,676
- 7 Claude Opus 5 1,559
- 8 GPT-6 Sol 1,448
- 9
DeepSeek V4.1 Flash 1017 1,405
- 10
MiniMax-M3.1-Flash 1,383
- 11 GLM-5.3 1,381
- 12
DeepSeek V4.1 Flash 1,295
- 13 Claude Opus 4.8 1,259
- 14
Gemini 3.8 Flash 1,047
- 15
Gemini 3.7 Flash 1,022
Source: Official RSI-Inference Bench leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 5.5 | 2,044 | Sep 2026 | Official leaderboard | |
| 2 | Claude Opus 5.5 | 1,903 | Sep 2026 | Official leaderboard | |
| 3 | Claude Fable 5.1 | 1,716 | Sep 2026 | Official leaderboard | |
| 4 | GPT-6.1 Sol | 1,685 | Sep 2026 | Official leaderboard | |
| 5 | GPT-6 Astra | 1,676 | Sep 2026 | Official leaderboard | |
| 6 | | 1,676 | Sep 2026 | Official leaderboard | |
| 7 | Claude Opus 5 | 1,559 | Jul 2026 | Official leaderboard | |
| 8 | GPT-6 Sol | 1,448 | Sep 2026 | Official leaderboard | |
| 9 | | 1,405 | Oct 2026 | Official leaderboard | |
| 10 | | 1,383 | Oct 2026 | Official leaderboard | |
| 11 | GLM-5.3 | 1,381 | Aug 2026 | Official leaderboard | |
| 12 | | 1,295 | Sep 2026 | Official leaderboard | |
| 13 | Claude Opus 4.8 | 1,259 | May 2026 | Official leaderboard | |
| 14 | | 1,047 | Sep 2026 | Official leaderboard | |
| 15 | | 1,022 | Aug 2026 | Official leaderboard |
Rankings by part
We read 39 parts from the benchmark’s own leaderboard in 2 groups: day-0 model support and inference optimization. Each ranks models on one part of the test. Claude Sonnet 5.5 leads 21 of them, and 6 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Day-0 Model Support
15 models we track have a score here.
- 1 Claude Sonnet 5.5 26.7%
- 2 Claude Opus 5.5 13.3%
- 3 Claude Fable 5.1 10%
- 4 GPT-6 Astra ↑1 10%
- 5 GPT-6.1 Sol ↓1 10%
- 6
Gemini 4 Argon 6.67%
- 7 Claude Opus 4.8 ↑6 0%
- 8 Claude Opus 5 ↓1 0%
- 9
Gemini 3.7 Flash ↑6 0%
- 10 GLM-5.3 ↑1 0%
Inference Optimization
15 models we track have a score here.
- 1 Claude Sonnet 5.5 1.68×
- 2 Claude Opus 5.5 1.48×
- 3 Claude Fable 5.1 1.41×
- 4
Gemini 4 Argon ↑2 1.39×
- 5 GPT-6.1 Sol ↓1 1.37×
- 6 GPT-6 Astra ↓1 1.35×
- 7 Claude Opus 5 1.35×
- 8
MiniMax-M3.1-Flash ↑2 1.26×
- 9
DeepSeek V4.1 Flash 1017 1.23×
- 10 GLM-5.3 ↑1 1.22×
DeepSeek-V4-Flash: SGLang
15 models we track have a score here.
- 1 Claude Sonnet 5.5 33.3%
- 2 Claude Opus 4.8 ↑11 0%
- 3 Claude Opus 5 ↑4 0%
- 4
Gemini 3.7 Flash ↑11 0%
- 5 GLM-5.3 ↑6 0%
- 6 Claude Fable 5.1 ↓3 0%
- 7
Gemini 3.8 Flash ↑7 0%
- 8 GPT-6 Astra ↓3 0%
- 9
DeepSeek V4.1 Flash ↑3 0%
- 10 Claude Opus 5.5 ↓8 0%
GLM-5.3-Flash: SGLang
15 models we track have a score here.
- 1 GPT-6 Astra ↑4 100%
- 2 Claude Sonnet 5.5 ↓1 100%
- 3 Claude Opus 5.5 ↓1 66.7%
- 4 GPT-6.1 Sol 66.7%
- 5
Gemini 4 Argon ↑1 66.7%
- 6 Claude Fable 5.1 ↓3 33.3%
- 7 Claude Opus 4.8 ↑6 0%
- 8 Claude Opus 5 ↓1 0%
- 9
Gemini 3.7 Flash ↑6 0%
- 10 GLM-5.3 ↑1 0%
gpt-oss-120b: SGLang
15 models we track have a score here.
- 1 Claude Opus 4.8 ↑12 0%
- 2 Claude Opus 5 ↑5 0%
- 3
Gemini 3.7 Flash ↑12 0%
- 4 GLM-5.3 ↑7 0%
- 5 Claude Fable 5.1 ↓2 0%
- 6
Gemini 3.8 Flash ↑8 0%
- 7 GPT-6 Astra ↓2 0%
- 8
DeepSeek V4.1 Flash ↑4 0%
- 9 Claude Opus 5.5 ↓7 0%
- 10 GPT-6 Sol ↓2 0%
Qwen3.8-Flash-Next-FP8: vLLM
15 models we track have a score here.
- 1 Claude Opus 4.8 ↑12 0%
- 2 Claude Opus 5 ↑5 0%
- 3
Gemini 3.7 Flash ↑12 0%
- 4 GLM-5.3 ↑7 0%
- 5 Claude Fable 5.1 ↓2 0%
- 6
Gemini 3.8 Flash ↑8 0%
- 7 GPT-6 Astra ↓2 0%
- 8
DeepSeek V4.1 Flash ↑4 0%
- 9 Claude Opus 5.5 ↓7 0%
- 10 GPT-6 Sol ↓2 0%
DeepSeek-V4.1-Flash: vLLM
15 models we track have a score here.
- 1 Claude Sonnet 5.5 33.3%
- 2 Claude Opus 4.8 ↑11 0%
- 3 Claude Opus 5 ↑4 0%
- 4
Gemini 3.7 Flash ↑11 0%
- 5 GLM-5.3 ↑6 0%
- 6 Claude Fable 5.1 ↓3 0%
- 7
Gemini 3.8 Flash ↑7 0%
- 8 GPT-6 Astra ↓3 0%
- 9
DeepSeek V4.1 Flash ↑3 0%
- 10 Claude Opus 5.5 ↓8 0%
diffusiongemma-26B-A4B-it: vLLM
15 models we track have a score here.
- 1 Claude Opus 4.8 ↑12 0%
- 2 Claude Opus 5 ↑5 0%
- 3
Gemini 3.7 Flash ↑12 0%
- 4 GLM-5.3 ↑7 0%
- 5 Claude Fable 5.1 ↓2 0%
- 6
Gemini 3.8 Flash ↑8 0%
- 7 GPT-6 Astra ↓2 0%
- 8
DeepSeek V4.1 Flash ↑4 0%
- 9 Claude Opus 5.5 ↓7 0%
- 10 GPT-6 Sol ↓2 0%
Inkling-Small-NVFP4: SGLang
15 models we track have a score here.
- 1 Claude Fable 5.1 ↑2 66.7%
- 2 Claude Opus 5.5 66.7%
- 3 Claude Sonnet 5.5 ↓2 66.7%
- 4 GPT-6.1 Sol 33.3%
- 5 Claude Opus 4.8 ↑8 0%
- 6 Claude Opus 5 ↑1 0%
- 7
Gemini 3.7 Flash ↑8 0%
- 8 GLM-5.3 ↑3 0%
- 9
Gemini 3.8 Flash ↑5 0%
- 10 GPT-6 Astra ↓5 0%
Kimi-Linear-48B-A3B: vLLM
15 models we track have a score here.
- 1 Claude Sonnet 5.5 33.3%
- 2 Claude Opus 4.8 ↑11 0%
- 3 Claude Opus 5 ↑4 0%
- 4
Gemini 3.7 Flash ↑11 0%
- 5 GLM-5.3 ↑6 0%
- 6 Claude Fable 5.1 ↓3 0%
- 7
Gemini 3.8 Flash ↑7 0%
- 8 GPT-6 Astra ↓3 0%
- 9
DeepSeek V4.1 Flash ↑3 0%
- 10 Claude Opus 5.5 ↓8 0%
MiniMax-M3-MXFP8: vLLM
15 models we track have a score here.
- 1 Claude Opus 4.8 ↑12 0%
- 2 Claude Opus 5 ↑5 0%
- 3
Gemini 3.7 Flash ↑12 0%
- 4 GLM-5.3 ↑7 0%
- 5 Claude Fable 5.1 ↓2 0%
- 6
Gemini 3.8 Flash ↑8 0%
- 7 GPT-6 Astra ↓2 0%
- 8
DeepSeek V4.1 Flash ↑4 0%
- 9 Claude Opus 5.5 ↓7 0%
- 10 GPT-6 Sol ↓2 0%
stable-diffusion-3-medium: SGLang
15 models we track have a score here.
- 1 Claude Opus 4.8 ↑12 0%
- 2 Claude Opus 5 ↑5 0%
- 3
Gemini 3.7 Flash ↑12 0%
- 4 GLM-5.3 ↑7 0%
- 5 Claude Fable 5.1 ↓2 0%
- 6
Gemini 3.8 Flash ↑8 0%
- 7 GPT-6 Astra ↓2 0%
- 8
DeepSeek V4.1 Flash ↑4 0%
- 9 Claude Opus 5.5 ↓7 0%
- 10 GPT-6 Sol ↓2 0%
gpt-oss-120b: vLLM
15 models we track have a score here.
- 1 Claude Opus 5.5 ↑1 1.38×
- 2 Claude Opus 5 ↑5 1.36×
- 3 Claude Sonnet 5.5 ↓2 1.30×
- 4
MiniMax-M3.1-Flash ↑6 1.29×
- 5 GPT-6.1 Sol ↓1 1.27×
- 6 GPT-6 Astra ↓1 1.27×
- 7 GLM-5.3 ↑4 1.26×
- 8
Gemini 4 Argon ↓2 1.25×
- 9 Claude Fable 5.1 ↓6 1.22×
- 10
DeepSeek V4.1 Flash ↑2 1.19×
Kimi-K2.6: vLLM
15 models we track have a score here.
- 1 Claude Sonnet 5.5 1.53×
- 2 Claude Fable 5.1 ↑1 1.34×
- 3 Claude Opus 5.5 ↓1 1.32×
- 4
DeepSeek V4.1 Flash 1017 ↑5 1.30×
- 5 GPT-6 Astra 1.27×
- 6 GLM-5.3 ↑5 1.27×
- 7 GPT-6 Sol ↑1 1.27×
- 8 GPT-6.1 Sol ↓4 1.26×
- 9
Gemini 4 Argon ↓3 1.25×
- 10 Claude Opus 4.8 ↑3 1.25×
DeepSeek-V4-Pro: vLLM
15 models we track have a score here.
- 1 Claude Opus 5.5 ↑1 1.28×
- 2 Claude Sonnet 5.5 ↓1 1.25×
- 3 Claude Fable 5.1 1.21×
- 4
Gemini 4 Argon ↑2 1.16×
- 5 Claude Opus 5 ↑2 1.08×
- 6 GPT-6.1 Sol ↓2 1.08×
- 7 GPT-6 Astra ↓2 1.07×
- 8 Claude Opus 4.8 ↑5 1.05×
- 9
DeepSeek V4.1 Flash ↑3 1.04×
- 10 GPT-6 Sol ↓2 1.04×
GLM-5.2-FP8: vLLM
15 models we track have a score here.
- 1 Claude Opus 5.5 ↑1 1.13×
- 2 Claude Sonnet 5.5 ↓1 1.12×
- 3 Claude Fable 5.1 1.06×
- 4 GPT-6 Astra ↑1 1.06×
- 5
Gemini 4 Argon ↑1 1.04×
- 6
MiniMax-M3.1-Flash ↑4 1.04×
- 7 Claude Opus 4.8 ↑6 1.03×
- 8 GPT-6.1 Sol ↓4 1.03×
- 9 Claude Opus 5 ↓2 1.03×
- 10 GPT-6 Sol ↓2 1.03×
MiniMax-M3-MXFP8: vLLM
15 models we track have a score here.
- 1 Claude Sonnet 5.5 1.36×
- 2 Claude Opus 5.5 1.22×
- 3
Gemini 4 Argon ↑3 1.19×
- 4 Claude Fable 5.1 ↓1 1.17×
- 5 GPT-6 Astra 1.09×
- 6
DeepSeek V4.1 Flash 1017 ↑3 1.09×
- 7 Claude Opus 5 1.08×
- 8 GPT-6.1 Sol ↓4 1.08×
- 9 GLM-5.3 ↑2 1.08×
- 10
DeepSeek V4.1 Flash ↑2 1.04×
Qwen3.6-27B: vLLM
15 models we track have a score here.
- 1 Claude Sonnet 5.5 1.39×
- 2 Claude Opus 5.5 1.34×
- 3 GPT-6 Astra ↑2 1.33×
- 4 Claude Fable 5.1 ↓1 1.29×
- 5
DeepSeek V4.1 Flash ↑7 1.28×
- 6 GPT-6.1 Sol ↓2 1.28×
- 7 Claude Opus 5 1.26×
- 8
DeepSeek V4.1 Flash 1017 ↑1 1.25×
- 9 Claude Opus 4.8 ↑4 1.21×
- 10 GLM-5.3 ↑1 1.16×
DeepSeek-V4-Flash: SGLang
15 models we track have a score here.
- 1 Claude Sonnet 5.5 1.58×
- 2 Claude Opus 5.5 1.16×
- 3 GLM-5.3 ↑8 1.10×
- 4 GPT-6 Astra ↑1 1.10×
- 5 Claude Fable 5.1 ↓2 1.08×
- 6 Claude Opus 5 ↑1 1.06×
- 7 GPT-6.1 Sol ↓3 1.06×
- 8 GPT-6 Sol 1.06×
- 9
MiniMax-M3.1-Flash ↑1 1.06×
- 10
Gemini 4 Argon ↓4 1.06×
MiniMax-M2.7: SGLang
15 models we track have a score here.
- 1 Claude Sonnet 5.5 1.43×
- 2 Claude Opus 5.5 1.33×
- 3 GPT-6 Astra ↑2 1.20×
- 4 Claude Fable 5.1 ↓1 1.18×
- 5
Gemini 4 Argon ↑1 1.16×
- 6 GLM-5.3 ↑5 1.16×
- 7
MiniMax-M3.1-Flash ↑3 1.10×
- 8 GPT-6.1 Sol ↓4 1.07×
- 9 Claude Opus 4.8 ↑4 1.07×
- 10 Claude Opus 5 ↓3 1.07×
Qwen3.5-397B-A17B-FP8: vLLM
15 models we track have a score here.
- 1 Claude Opus 5.5 ↑1 1.43×
- 2 Claude Sonnet 5.5 ↓1 1.34×
- 3 Claude Fable 5.1 1.18×
- 4 GPT-6.1 Sol 1.17×
- 5 Claude Opus 5 ↑2 1.17×
- 6 GPT-6 Astra ↓1 1.17×
- 7
MiniMax-M3.1-Flash ↑3 1.08×
- 8 GLM-5.3 ↑3 1.06×
- 9
Gemini 4 Argon ↓3 1.05×
- 10 GPT-6 Sol ↓2 1.04×
LLaDA2.0-mini-preview: SGLang
15 models we track have a score here.
- 1 Claude Sonnet 5.5 2.45×
- 2 Claude Fable 5.1 ↑1 1.93×
- 3 Claude Opus 5 ↑4 1.86×
- 4
Gemini 4 Argon ↑2 1.84×
- 5
DeepSeek V4.1 Flash 1017 ↑4 1.84×
- 6
DeepSeek V4.1 Flash ↑6 1.68×
- 7
MiniMax-M3.1-Flash ↑3 1.68×
- 8 Claude Opus 5.5 ↓6 1.42×
- 9 GPT-6 Astra ↓4 1.14×
- 10 GPT-6 Sol ↓2 1.13×
Kimi-Linear-48B-A3B: SGLang
15 models we track have a score here.
- 1 Claude Opus 5.5 ↑1 2.40×
- 2 Claude Sonnet 5.5 ↓1 2.10×
- 3 Claude Fable 5.1 2.06×
- 4 Claude Opus 5 ↑3 2.02×
- 5 GPT-6 Astra 2.01×
- 6
DeepSeek V4.1 Flash 1017 ↑3 2.00×
- 7 GPT-6.1 Sol ↓3 1.85×
- 8
Gemini 4 Argon ↓2 1.85×
- 9
MiniMax-M3.1-Flash ↑1 1.79×
- 10 GLM-5.3 ↑1 1.72×
DeepSeek-V4-Flash: SGLang: PD
15 models we track have a score here.
- 1 Claude Opus 5.5 ↑1 2.42×
- 2 Claude Sonnet 5.5 ↓1 2.37×
- 3 GPT-6 Astra ↑2 2.27×
- 4 GPT-6.1 Sol 2.08×
- 5
Gemini 4 Argon ↑1 2.08×
- 6
MiniMax-M3.1-Flash ↑4 2.06×
- 7 Claude Opus 5 2.03×
- 8 GPT-6 Sol 1.77×
- 9 Claude Fable 5.1 ↓6 1.76×
- 10
DeepSeek V4.1 Flash ↑2 1.72×
MiniMax-M3-MXFP8: SGLang: PD
15 models we track have a score here.
- 1 Claude Opus 5.5 ↑1 1.30×
- 2 Claude Sonnet 5.5 ↓1 1.30×
- 3 Claude Opus 5 ↑4 1.22×
- 4
DeepSeek V4.1 Flash 1017 ↑5 1.21×
- 5 Claude Fable 5.1 ↓2 1.19×
- 6
DeepSeek V4.1 Flash ↑6 1.19×
- 7 GPT-6 Sol ↑1 1.19×
- 8
MiniMax-M3.1-Flash ↑2 1.15×
- 9 GPT-6.1 Sol ↓5 1.11×
- 10 GPT-6 Astra ↓5 1.07×
Qwen3.6-27B: SGLang: PD bulk
15 models we track have a score here.
- 1 Claude Opus 5.5 ↑1 2.75×
- 2 Claude Sonnet 5.5 ↓1 2.74×
- 3 Claude Fable 5.1 2.70×
- 4
Gemini 4 Argon ↑2 2.70×
- 5 Claude Opus 5 ↑2 2.55×
- 6 GPT-6.1 Sol ↓2 2.49×
- 7
MiniMax-M3.1-Flash ↑3 2.04×
- 8 GPT-6 Astra ↓3 1.71×
- 9
DeepSeek V4.1 Flash ↑3 1.67×
- 10 Claude Opus 4.8 ↑3 1.58×
Qwen3.6-27B: vLLM: PD
15 models we track have a score here.
- 1 Claude Sonnet 5.5 2.12×
- 2 GPT-6.1 Sol ↑2 1.74×
- 3
Gemini 4 Argon ↑3 1.52×
- 4 Claude Fable 5.1 ↓1 1.32×
- 5
MiniMax-M3.1-Flash ↑5 1.28×
- 6 GLM-5.3 ↑5 1.25×
- 7 Claude Opus 5 1.24×
- 8 GPT-6 Astra ↓3 1.21×
- 9 GPT-6 Sol ↓1 1.17×
- 10 Claude Opus 4.8 ↑3 1.14×
GLM-5.2-FP8: vLLM: MTP
15 models we track have a score here.
- 1 Claude Sonnet 5.5 1.21×
- 2
Gemini 3.8 Flash ↑12 1.17×
- 3 Claude Fable 5.1 1.09×
- 4 Claude Opus 5.5 ↓2 1.09×
- 5 GPT-6.1 Sol ↓1 1.05×
- 6
Gemini 4 Argon 1.05×
- 7 GLM-5.3 ↑4 1.03×
- 8
MiniMax-M3.1-Flash ↑2 1.03×
- 9 Claude Opus 5 ↓2 1.03×
- 10 GPT-6 Sol ↓2 1.03×
MiniMax-M3-MXFP8: vLLM: MTP
15 models we track have a score here.
- 1 Claude Sonnet 5.5 1.16×
- 2 Claude Opus 5.5 1.11×
- 3
Gemini 4 Argon ↑3 1.07×
- 4 GPT-6 Astra ↑1 1.07×
- 5
DeepSeek V4.1 Flash 1017 ↑4 1.07×
- 6 GPT-6.1 Sol ↓2 1.04×
- 7 Claude Fable 5.1 ↓4 1.03×
- 8 Claude Opus 4.8 ↑5 1.03×
- 9
MiniMax-M3.1-Flash ↑1 1.03×
- 10 Claude Opus 5 ↓3 1.02×
Qwen3.8-27B: vLLM: DFlash2
15 models we track have a score here.
- 1 Claude Sonnet 5.5 1.41×
- 2 GPT-6 Astra ↑3 1.28×
- 3 Claude Opus 5.5 ↓1 1.18×
- 4 Claude Fable 5.1 ↓1 1.12×
- 5 Claude Opus 5 ↑2 1.11×
- 6 GPT-6.1 Sol ↓2 1.10×
- 7 Claude Opus 4.8 ↑6 1.07×
- 8
Gemini 4 Argon ↓2 1.07×
- 9 GPT-6 Sol ↓1 1.06×
- 10
MiniMax-M3.1-Flash 1.06×
DeepSeek-V4-Flash-DSpark: SGLang
15 models we track have a score here.
- 1
Gemini 4 Argon ↑5 1.11×
- 2 Claude Sonnet 5.5 ↓1 1.09×
- 3 Claude Opus 5.5 ↓1 1.09×
- 4 GPT-6.1 Sol 1.07×
- 5 Claude Fable 5.1 ↓2 1.05×
- 6 Claude Opus 5 ↑1 1.05×
- 7
DeepSeek V4.1 Flash ↑5 1.03×
- 8
MiniMax-M3.1-Flash ↑2 1.03×
- 9 GPT-6 Sol ↓1 1.03×
- 10 GPT-6 Astra ↓5 1.02×
Qwen3.5-397B-A17B-FP8: SGLang: MTP
15 models we track have a score here.
- 1 Claude Sonnet 5.5 1.56×
- 2 Claude Opus 5.5 1.48×
- 3 Claude Fable 5.1 1.34×
- 4 Claude Opus 5 ↑3 1.30×
- 5
DeepSeek V4.1 Flash 1017 ↑4 1.27×
- 6
Gemini 4 Argon 1.20×
- 7 GPT-6.1 Sol ↓3 1.18×
- 8 Claude Opus 4.8 ↑5 1.15×
- 9 GPT-6 Astra ↓4 1.12×
- 10 GPT-6 Sol ↓2 1.09×
MXFP4 MoE decode throughput: SGLang
15 models we track have a score here.
- 1 Claude Sonnet 5.5 1.95×
- 2
Gemini 4 Argon ↑4 1.61×
- 3 GPT-6 Sol ↑5 1.59×
- 4 GPT-6 Astra ↑1 1.58×
- 5 GPT-6.1 Sol ↓1 1.58×
- 6 Claude Opus 5 ↑1 1.57×
- 7 Claude Opus 5.5 ↓5 1.56×
- 8
DeepSeek V4.1 Flash ↑4 1.54×
- 9 GLM-5.3 ↑2 1.51×
- 10 Claude Fable 5.1 ↓7 1.48×
Spec-decode acceptance drop: vLLM
15 models we track have a score here.
- 1 Claude Sonnet 5.5 1.81×
- 2
Gemini 4 Argon ↑4 1.58×
- 3 GPT-6 Sol ↑5 1.45×
- 4
DeepSeek V4.1 Flash 1017 ↑5 1.45×
- 5 Claude Opus 5 ↑2 1.45×
- 6
DeepSeek V4.1 Flash ↑6 1.34×
- 7 GPT-6.1 Sol ↓3 1.23×
- 8 Claude Opus 5.5 ↓6 1.23×
- 9 Claude Fable 5.1 ↓6 1.18×
- 10 GPT-6 Astra ↓5 1.17×
Multimodal prefill latency: SGLang
15 models we track have a score here.
- 1 Claude Sonnet 5.5 6.17×
- 2 GPT-6 Astra ↑3 5.53×
- 3 GPT-6.1 Sol ↑1 5.49×
- 4 Claude Fable 5.1 ↓1 5.11×
- 5
Gemini 4 Argon ↑1 4.98×
- 6 Claude Opus 5.5 ↓4 4.77×
- 7
DeepSeek V4.1 Flash 1017 ↑2 3.20×
- 8 Claude Opus 5 ↓1 3.19×
- 9 GPT-6 Sol ↓1 3.15×
- 10
MiniMax-M3.1-Flash 3.01×
MTP decode throughput: vLLM
15 models we track have a score here.
- 1 Claude Sonnet 5.5 2.67×
- 2 Claude Opus 5.5 1.69×
- 3 GPT-6 Astra ↑2 1.67×
- 4 GPT-6.1 Sol 1.60×
- 5 Claude Fable 5.1 ↓2 1.57×
- 6
Gemini 4 Argon 1.54×
- 7 Claude Opus 4.8 ↑6 1.43×
- 8 GLM-5.3 ↑3 1.33×
- 9
DeepSeek V4.1 Flash ↑3 1.25×
- 10
MiniMax-M3.1-Flash 1.24×
Async spec-decode throughput: vLLM
15 models we track have a score here.
- 1 Claude Sonnet 5.5 2.05×
- 2 Claude Opus 5.5 1.62×
- 3 Claude Opus 5 ↑4 1.38×
- 4 Claude Fable 5.1 ↓1 1.36×
- 5 GPT-6 Sol ↑3 1.36×
- 6 GLM-5.3 ↑5 1.28×
- 7
Gemini 4 Argon ↓1 1.24×
- 8 GPT-6.1 Sol ↓4 1.19×
- 9 GPT-6 Astra ↓4 1.17×
- 10
MiniMax-M3.1-Flash 1.14×
FP8 decode throughput: vLLM
15 models we track have a score here.
- 1 GPT-6.1 Sol ↑3 1.53×
- 2 Claude Opus 5.5 1.47×
- 3 GPT-6 Astra ↑2 1.42×
- 4
Gemini 4 Argon ↑2 1.37×
- 5 Claude Fable 5.1 ↓2 1.34×
- 6 GLM-5.3 ↑5 1.31×
- 7 Claude Opus 5 1.31×
- 8
DeepSeek V4.1 Flash ↑4 1.29×
- 9 GPT-6 Sol ↓1 1.28×
- 10
DeepSeek V4.1 Flash 1017 ↓1 1.28×
Streaming decode at concurrency 1: SGLang
15 models we track have a score here.
- 1 GPT-6.1 Sol ↑3 1.65×
- 2 Claude Opus 5.5 1.63×
- 3 Claude Sonnet 5.5 ↓2 1.62×
- 4
Gemini 4 Argon ↑2 1.61×
- 5 Claude Fable 5.1 ↓2 1.56×
- 6 GPT-6 Astra ↓1 1.55×
- 7 GLM-5.3 ↑4 1.47×
- 8 Claude Opus 5 ↓1 1.46×
- 9
DeepSeek V4.1 Flash 1017 1.44×
- 10
DeepSeek V4.1 Flash ↑2 1.42×
The best score, 2,044, costs $76 a task
Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.
Source: Official RSI-Inference Bench leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| Claude Sonnet 5.5 | 2,044 | $76 | 1,674,492 | ||
| Claude Opus 5.5 | 1,903 | $83 | 805,686 | ||
| Claude Fable 5.1 | 1,716 | $71 | 518,308 | ||
| GPT-6.1 Sol | 1,685 | $12 | 372,113 | ||
| GPT-6 Astra | 1,676 | $81 | 254,553 | ||
| 1,676 | $68 | 709,509 | |||
| Claude Opus 5 | 1,559 | $114 | 344,039 | ||
| GPT-6 Sol | 1,448 | $41 | 282,499 | ||
| 1,405 | $2.11 | 468,299 | |||
| 1,383 | $4.51 | 685,140 | |||
| GLM-5.3 | 1,381 | $32 | 454,765 | ||
| 1,295 | $2.20 | 850,046 | |||
| Claude Opus 4.8 | 1,259 | $118 | 622,818 | ||
| 1,047 | $60 | 431,229 | |||
| 1,022 | $57 | 325,436 |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official RSI-Inference Bench leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official RSI-Inference Bench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in AI research
- PaperBench In use Agent replicates ICML 2024 papers from scratch. LLM judge grades against author-written hierarchical rubrics.
Qwen 3.8 Max 93%
- MLS-Bench In use Agent must invent one algorithmic change in an ML research scaffold whose gain transfers across settings. Scored relative to human baselines. Claude Fable 5.1 50.3%
- PostTrainBench 1.2 In use CLI agent gets a base model, one H100 and a time budget to post-train it for a target benchmark. Scored on normalized gain. Claude Fable 5.1 44.6%
- MLE-bench In use Kaggle-style ML engineering competitions. Agent trains models and is scored against competition leaderboards. GPT-6.2 Astra 96.1%
- PMPP-Hard In use Agents write GPU kernels for 69 tasks that must be correct, and on 48 of them also run within 1.25 times the reference time. The score is the area under the solved-tasks curve over a 10-hour budget, out of 690 task-hours. Kimi K3 0.71
- KernelBench In use Replace PyTorch operators with faster custom GPU kernels. Scored on correctness and speedup.
MiniMax-M3 28.8%
Data audit
All 10 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
- Pass Independent runs Scores from the official leaderboard (15 entries).
- Pass Models scored 15 models we track.
- Pass Newest score The newest is from October 2026, for MiniMax-M3.1-Flash.
- Pass Frontier models 6 of the top 10 have a score, led by GPT-6 Astra (#4).
- Note Headroom Scored as an Elo rating, with no fixed ceiling.
Leaderboard import 3 pass
- Pass Importer We read the board on rsi-inference.com: 15 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.