LLM Speed Chart

This chart puts score and speed on one canvas: right = faster, up = stronger, so top-right is where you want to be. Bigger bubbles mean cheaper output.

Updated Sep 14 55 models Synced daily 3 pending scores ↓
Showing 30 modelsPrice unit: USD / million tokens

Scores and speed at a glance

0 20 40 60 100 200 300 Intelligence Index Output speed · tokens/s (linear scale) A Claude Fable 5.1 · Speed 70/s · Effort tiers 47.0–53.4 A Claude Opus 5 · Speed 58/s · Effort tiers 39.8–50.7 G Gemini 3.8 Flash · Speed 340/s · Effort tiers 33.8–41.2 A Claude Fable 5 · Speed 70/s G Gemini 3.7 Flash · Speed 337/s · Effort tiers 36.9–39.6 M Muse Spark 1.2 · Speed 211/s A Qwen3.8 Max · Speed 43/s G Gemini 3.1 Pro Preview · Speed 127/s G Gemini 3.6 Flash · Speed 222/s Z GLM-5.3 · Speed 68/s M Kimi K3 · Speed 37/s · Effort tiers 30.5–43.8 Z GLM-5.3 Flash · Speed 114/s Z GLM-5.2 · Speed 74/s · Effort tiers 22.4–34.0 X MiMo V2.5 Pro · Speed 42/s · Effort tiers 18.3–26.4 O GPT-5.6 Sol · Speed 68/s · Effort tiers 28.3–47.1 A Qwen3.7 Plus · Speed 74/s D DeepSeek V4 Pro · Speed 75/s x Grok 4.5 · Speed 61/s O GPT-5.6 Terra · Speed 125/s · Effort tiers 22.3–42.3 A Claude Sonnet 5 · Speed 89/s · Effort tiers 24.7–38.4 T Hy3 · Speed 97/s · Effort tiers 17.0–25.8 M MiniMax M3 · Speed 118/s D DeepSeek V4 Flash · Speed 226/s O GPT-5.6 Luna · Speed 118/s · Effort tiers 16.8–37.5 x Grok 4.6 · Speed 71/s · Effort tiers 35.4–44.4 X MiMo V2.5 · Speed 55/s D DeepSeek V4.1 Flash · Speed 227/s M Mistral Medium 3.5 · Speed 166/s M Kimi K2.7 Code · Speed 44/s O GPT-6 Astra · Speed 68/s · Effort tiers 46.0–52.8
Speed-score frontierBubble size = output price (bigger = cheaper) $0.1772–$50.00

How to read the chart

Higher means a better score; farther right means faster output. Bigger bubbles mean cheaper output. Select an icon for details.

The green line uses all models with a score and a measured speed; filters do not change this reference line.

Models without a measured speed or a score are not plotted; a model with no listed price gets the smallest bubble, which does not mean it is cheap.

Scoring method and data sources

Arena is LMArena’s overall rating, not an independent coding evaluation. Prices are from OpenRouter; open weights do not mean unrestricted commercial use. The score-to-price ratio is simply the Arena score divided by output price, not an independent benchmark. Data sources

Models in the chart
  1. Claude Fable 5.153.470/s
  2. GPT-6 Astra52.868/s
  3. Claude Opus 550.758/s
  4. Claude Fable 549.770/s
  5. GPT-5.6 Sol47.168/s
  6. GLM-5.344.968/s
  7. Grok 4.644.471/s
  8. Kimi K343.837/s
  9. GPT-5.6 Terra42.3125/s
  10. GLM-5.3 Flash41.9114/s
  11. Gemini 3.8 Flash41.2340/s
  12. Qwen3.8 Max40.343/s
  13. Muse Spark 1.239.8211/s
  14. DeepSeek V4.1 Flash39.5227/s
  15. Gemini 3.7 Flash39.4337/s
  16. Grok 4.539.161/s
  17. Claude Sonnet 538.489/s
  18. GPT-5.6 Luna37.5118/s
  19. DeepSeek V4 Pro36.375/s
  20. DeepSeek V4 Flash34.5226/s
  21. Gemini 3.6 Flash34.3222/s
  22. GLM-5.234.074/s
  23. Gemini 3.1 Pro Preview30.4127/s
  24. MiniMax M329.6118/s
  25. MiMo V2.5 Pro26.442/s
  26. Kimi K2.7 Code26.344/s
  27. Qwen3.7 Plus25.874/s
  28. Hy325.897/s
  29. MiMo V2.522.355/s
  30. Mistral Medium 3.514.9166/s
Top overall Claude Fable 5.1 Arena 1,507.6 Best score-to-price ratio DeepSeek V4 Flash 8,080.1 Highest-scoring open-weight model GLM-5.3 Arena 1,475.1 Newest DeepSeek V4.1 Flash Sep 10

Not yet ranked by Arena

The 3 models below are indexed here, but LMArena has not scored them yet, so they do not appear in the ranking above (we only show real ranks — no invented scores). Sorted newest first; they join the leaderboard automatically once Arena lists them.

New models in the last 14 days

Sorted by listing date; all models here already have Arena rankings. Unrated new models are listed in the “Not yet ranked by Arena” section.

Data sources

Unofficial mirror — no self-invented scores. Prices are OpenRouter quotes per million tokens; Arena scores use the Bradley-Terry scale, not Elo.

FAQ

Which LLM is the strongest right now?

Go by the overall ranking in this table; when scores are close, weigh price and open weights — there is no single strongest.

What is the difference between open-weight and closed LLMs?

Open weights can be self-hosted or deployed privately; closed models usually run behind an API. See the open-source LLM leaderboard.

Which LLM should I look at for coding?

Professional coding benchmarks are a different test. For a coding-oriented filter see the coding LLM leaderboard, currently sorted by overall score.

Which Chinese LLM is the strongest right now?

There is no single strongest. For open-weight Chinese models see the open-source LLM leaderboard; when scores are close, weigh price and self-hosting.

Why do prices differ from the official site?

Figures here are OpenRouter quotes in USD per million tokens and may differ from official APIs or local resellers.