Independent, real-world evaluation of AI models

Open-weight and hosted models take the same private test suite, graded by code instead of another model. Every score links to the run behind it.

What they build

1Claude Haiku 5.570Hosted225 tok/s$0.42 per run70$0.42ran 2026-10-07

2Muse Spark 1.3 Contributor66Hosted107 tok/s$0.41 per run66$0.41ran 2026-10-05

3DeepSeek V4.1 Flash64Hosted30 tok/s$0.39 per run64$0.39ran 2026-10-03

4GPT-6 Luna61Hosted90 tok/s$0.10 per run61$0.10ran 2026-10-04

5Beam 1.0 501B A23B57Hosted274 tok/s$0.83 per run57$0.83ran 2026-10-16

6Qwen3.8 27B48Hosted196 tok/s$1.28 per run48$1.28ran 2026-10-12

6MiMo-V2.6-Pro48Hosted55 tok/s$1.80 per run48$1.80ran 2026-10-09

8Ling 3.1 Flash45Hosted112 tok/s45Not measuredran 2026-10-04

8MiMo-V2.6-Flash45Hosted66 tok/s$0.33 per run45$0.33ran 2026-10-04

10GLM 5.3 Flash35Hosted60 tok/s$0.19 per run35$0.19ran 2026-10-03

11Underdog Saluki 27B33IQ2_XS10.2 GB at 32K23 tok/s on M5 Pro33Not measuredran 2026-10-15

12Qwen3.8 Flash32Hosted77 tok/s$0.46 per run32$0.46ran 2026-10-06

13Mistral Large 428Hosted56 tok/s$2.67 per run28$2.67ran 2026-10-06

14Apodex 1.1 Mini (free)25Hosted278 tok/s25Not measuredran 2026-10-05

15Mellum2.1 Thinking9Q8_012.9 GB at 32K90 tok/s on M5 Pro$2.46 per runest.9$2.46est.ran 2026-10-09

16LFM2.5 24B A2B7Q8_025.0 GB at 32K82 tok/s on M5 Pro7Not measuredran 2026-10-14

Each model ranks by its best build. Equal overall scores share a rank. How scores work

One email when a new model lands on the board or in the showcase. Unsubscribe anytime.