Skip to content
Smol.ai

Arena leaderboard

Compare model speed and blind votes from Smol for Mac.

Leaderboard

Live speed

Hosted launcher requests, timed at the Smol Worker. Includes single-pass attempts with or without web search.

Searches run
Output tokens

Loading live results…

How speed is measured

Smol’s Worker measures from provider dispatch to the first and last visible answer chunks. The Mac also measures delivery to your device for single-model replies. The 95th percentile (p95) shows how long the slowest 5% took to reach that point. Timings use completed answers. Completion rate includes stopped and failed attempts. Turns that needed desktop tool follow-ups are excluded. Search and length filters require reported usage. Delivery counts reported output tokens over the visible streaming interval. A dash means a metric has too few eligible observations.

One-search benchmark

The same lookup and fixed-length answer on all 14 models, fastest median finish first. Nov 3, 2026 · Five OpenRouter runs per model with one basic search. tok/s counts visible delivery in reference tokens. Buffered text can arrive in bursts.

ModelFirst textDeliveryFinish ↑
DeepSeek V4.1 Flash 10171.90 s234 tok/s3.06 s
Claude Sonnet 5.53.54 s3,205 tok/s3.58 s
Claude Haiku 5.53.26 s233 tok/s3.79 s
Grok 4.73.38 s207 tok/s3.88 s
GLM 5.3 Prime3.63 s196 tok/s4.29 s
GPT-6 Luna3.77 s152 tok/s4.30 s
Muse Spark Open1.07 s485 tok/s5.49 s
GPT-6.1 Terra5.27 s197 tok/s6.00 s
Claude Opus 5.57.07 s4,595 tok/s7.10 s
GPT-6.1 Sol6.67 s234 tok/s7.68 s
Muse Wonder7.67 s386 tok/s8.02 s
Kimi K3.1 Pro12.87 s62 tok/s14.92 s
Muse Spark 1.34.99 s270 tok/s15.50 s
Muse Glimmer 30B14.68 s67 tok/s18.51 s
Smol.ai
Opening your workspace…