LLM Speed Chart

As of Oct 16, the fastest model by output speed is K2 Horizon 375B A23B.

  • Data updated
  • Arena published
  • Agents board released
  • 84 models
  • Sources & methodology
Top agentClaude Opus 5.517.18% Top in TextGemini 4 Argon1,523.9 Intelligence #1Claude Opus 5.557.6 Coding #1Claude Fable 5.181.6 Best valueMiMo-V2.6-Flash$0.28/1M out Open weights #1Kimi K3Arena 1,487.1

This week

Daily Agents records started on Oct 16, 2026; rank changes appear after 14 days.

New models

last 14 days
  1. Muse Spark OpenMetaOct 15
  2. Muse WonderMetaOct 12
  3. Muse Wonder FullMetaOct 12
  4. Claude Haiku 5.5AnthropicOct 7
  5. Mistral Large 4 PreviewMistralOct 6
Showing 36 modelsPrice unit: USD / million tokens

Scores and speed at a glance

0 20 40 60 0 200 400 600 800 1000 Intelligence Index Output speed · tokens/s (linear scale) M Muse Wonder · Speed 174/s A Claude Opus 5.5 · Speed 98/s · Effort tiers 42.3–57.6 A Claude Fable 5.1 · Speed 66/s · Effort tiers 46.8–53.4 G Gemini 3.8 Flash · Speed 137/s · Effort tiers 33.5–40.9 M Muse Spark 1.3 · Speed 135/s · Effort tiers 45.1–48.1 M Kimi K3 · Speed 45/s · Effort tiers 30.1–43.6 G Gemini 3.1 Pro Preview · Speed 139/s O GPT-6.1 Sol · Speed 59/s · Effort tiers 42.1–51.8 A Qwen3.8 Max · Speed 38/s X MiMo-V2.6-Pro · Speed 50/s Z GLM-5.3 · Speed 71/s · Effort tiers 34.3–44.8 O GPT-6 Astra · Speed 49/s · Effort tiers 45.8–52.7 A Claude Sonnet 5.5 · Speed 142/s · Effort tiers 35.9–56.0 D DeepSeek V4.1 Flash · Speed 180/s · Effort tiers 24.7–39.5 Z GLM-5.3 Flash · Speed 63/s O GPT-5.6 Terra · Speed 116/s · Effort tiers 20.8–42.1 D DeepSeek V4 Pro · Speed 93/s · Effort tiers 20.4–36.0 A Qwen3.7 Plus · Speed 55/s T Hy3 · Speed 87/s · Effort tiers 17.0–25.3 X MiMo-V2.6-Flash · Speed 51/s S Step 5 Preview · Speed 86/s x Grok 4.7 · Speed 84/s · Effort tiers 42.2–46.4 O GPT-6 Luna · Speed 128/s · Effort tiers 18.5–38.1 M MiniMax M3 · Speed 102/s A Qwen3.8 27B · Speed 40/s · Effort tiers 20.2–33.7 M Mistral Medium 3.5 · Speed 168/s A Qwen3.8-Flash-Next · Speed 48/s A Claude Haiku 5.5 · Speed 235/s · Effort tiers 29.5–43.4 D DeepSeek V4 Flash Vision · Speed 176/s I Ling 3.1 Flash · Speed 217/s I K2 Horizon 375B A23B · Speed 1021/s M Muse Spark Open · Speed 105/s M Kimi K2.7 Code · Speed 52/s O GPT-5.3 Codex · Speed 94/s A Qwen3.8 2.4T A95B · Speed 38/s S Agnes 2.5 Pro Beta · Speed 120/s
Speed-score frontierBubble size = output price (bigger = cheaper) $0.28–$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

Ranked by LMArena Arena score by default, switchable to the Artificial Analysis Intelligence Index; Arena is not an independent coding evaluation. Prices are official API list prices, with OpenRouter quotes shown separately as the lowest route price. Open weights do not mean unrestricted commercial use. Data sources

Models in the chart
  1. Claude Opus 5.557.698/s
  2. GPT-6.1 Sol51.859/s
  3. Claude Fable 5.153.466/s
  4. GPT-6 Astra52.749/s
  5. Claude Sonnet 5.556.0142/s
  6. Muse Wonder52.4174/s
  7. Gemini 3.8 Flash40.9137/s
  8. MiMo-V2.6-Pro46.350/s
  9. Muse Spark 1.348.1135/s
  10. Grok 4.746.484/s
  11. GPT-6 Luna38.1128/s
  12. DeepSeek V4.1 Flash39.5180/s
  13. MiMo-V2.6-Flash37.951/s
  14. GLM-5.344.871/s
  15. Kimi K343.645/s
  16. Qwen3.8 Max45.438/s
  17. Claude Haiku 5.543.4235/s
  18. GPT-5.6 Terra42.1116/s
  19. DeepSeek V4 Pro36.093/s
  20. GLM-5.3 Flash41.863/s
  21. Step 5 Preview43.786/s
  22. Qwen3.8-Flash-Next39.848/s
  23. Qwen3.8 27B33.740/s
  24. Gemini 3.1 Pro Preview29.7139/s
  25. Hy325.387/s
  26. Qwen3.7 Plus25.255/s
  27. MiniMax M329.2102/s
  28. Mistral Medium 3.514.2168/s
  29. Ling 3.1 Flash41.1217/s
  30. Qwen3.8 2.4T A95B39.938/s
  31. Muse Spark Open38.9105/s
  32. Agnes 2.5 Pro Beta35.2120/s
  33. DeepSeek V4 Flash Vision34.8176/s
  34. GPT-5.3 Codex32.594/s
  35. K2 Horizon 375B A23B30.51021/s
  36. Kimi K2.7 Code25.852/s

Not yet ranked by Arena 2

Tracked models that have no LMArena or Artificial Analysis score yet, newest first. They move into the table once scored.

Sources & methodology

As of Oct 16, 2026, the fastest model by measured output speed is K2 Horizon 375B A23B (1021 tokens/s); 5 models sit on the speed frontier (none is both faster and stronger). Speed and latency are Artificial Analysis medians.

  • arena.ai Agents

    Agents leaderboard: task success rate, 95% interval, sessions and rank (arena.ai Agents, via the public Hugging Face dataset lmarena-ai/leaderboard-dataset, CC BY 4.0). Hugging Face

  • LMArena

    Arena score, 95% confidence interval, votes and rank (overall Text leaderboard with style control, the same view arena.ai shows by default; dataset CC BY 4.0).

  • Artificial Analysis

    Intelligence Index, Coding Index, output speed, TTFT / first-answer latency, release dates and official prices (shown on this page only, not in the public API).

  • OpenRouter

    Per-provider prices (the cheapest becomes “lowest route”), context length and listing dates.

  • Vendor pricing pages

    “Input / Output” are official API list prices (USD per 1M tokens); when no official price exists we use the OpenRouter price and mark it with *.

How we compute it

  1. Ranking: by default, the arena.ai Agents task success rate (real users’ coding/tool agent tasks). When a model has several reasoning tiers (Max/High/xHigh), the best-scoring tier is used and labelled; fewer than 10,000 sessions or a 95% half-width above 3 points is marked Preliminary. Models without an agent result follow, ordered by LMArena Text (style control); the Text column is still sortable.
  2. Rank: by default, LMArena Text Arena score with style control (the same view arena.ai shows by default; Bradley-Terry, not Elo). The ± is the half-width of the 95% confidence interval, and neighbouring ranks often overlap. Models with fewer than 5,000 votes or ± above 10 are marked Preliminary. We do not invent a composite score.
  3. Intelligence / Coding Index: Artificial Analysis aggregates of public benchmarks. When a model has several reasoning-effort tiers, the table shows the best tier for the current sort column; switch to “by tier” to see each one.
  4. Prices are USD per 1M tokens. “Lowest route” is the cheapest OpenRouter provider’s input / output price; ⚠ marks routes more than 60% below list (usually promos or a single provider). The best-value board only includes models with Arena ≥ 1400 or Intelligence Index ≥ 40, sorted by official output price.
  5. Latency and speed: TTFT is the time to the first token (including thinking tokens); “Answer” is the time to the first answer token (includes the thinking phase for reasoning models); speed is output tokens per second. All are Artificial Analysis medians.
  6. Updates: Text and other data sync daily at 14:20 (UTC+8), the Agents board at 14:25. Agent rank changes appear only after 14 days of daily records; until then no arrows are shown.
  7. Updates: synced automatically every day at 14:20 (UTC+8). LMArena publishes on its own schedule, so the “Arena published” date can be older than the update date. “This week” compares against the daily archive from 7 days earlier.
  8. Open data: /api/public/llm-leaderboard.php (JSON, LMArena and OpenRouter fields only).

FAQ

What is the best AI model right now?

As of Oct 16, 2026, the arena.ai Agents leaderboard (success rate on real coding/tool tasks) is led by Claude Opus 5.5 (17.18%, 95% interval 13.80%–20.56%), ahead of GPT-6.1 Sol (14.78%). The top intervals overlap heavily, so the exact order is not reliable. On the LMArena Text leaderboard, Gemini 4 Argon leads (1,523.9).

Which good AI model is the cheapest?

Among models above the quality floor (Arena ≥ 1400 or Intelligence Index ≥ 40), MiMo-V2.6-Flash has the lowest official price: $0.14 input / $0.28 output per 1M tokens, followed by MiMo V2.5 ($0.28 output). See the best-value leaderboard.

What is the best open-source (open-weight) AI model?

By Arena score, the top open-weight model is Kimi K3 (Moonshot AI, 1,487.1), followed by MiMo-V2.6-Pro (1,479.2). We track 21 open-weight models with an Arena score — see the open-weight leaderboard. Open weights do not mean unrestricted commercial use; check each licence.

Which AI model is best for coding?

On the Artificial Analysis Coding Index, Claude Fable 5.1 is highest (81.6), followed by Claude Opus 5 (78.0). The index aggregates coding benchmarks and differs from the Arena score; see the full coding leaderboard.

Which AI model is the fastest?

The fastest measured output speed belongs to K2 Horizon 375B A23B (1021 tokens/s); the lowest time to first answer token belongs to K2 Horizon 375B A23B (3.65 s). These are Artificial Analysis medians; for reasoning models the first-answer latency includes thinking time.

Which AI models were released recently?

5 tracked models were released in the last 14 days, including Muse Spark Open (Oct 15, 2026), Muse Wonder (Oct 12, 2026), Muse Wonder Full (Oct 12, 2026), Claude Haiku 5.5 (Oct 7, 2026). New models usually take a few days to appear on LMArena; until then they are listed under “Not yet ranked on Arena”.

How often is the data updated, and where does it come from?

Synced automatically every day: the Agents board comes from arena.ai (public Hugging Face dataset, current release Oct 14, 2026), Text scores from LMArena (released Oct 15, 2026); intelligence and coding indices, speed and latency from Artificial Analysis; route prices and context from OpenRouter. This page was updated Oct 16, 2026.

Why do prices differ from the official site?

Input/output prices are official list prices in USD per million tokens. The lowest route price is the cheapest OpenRouter route, which may be a promotion or a single provider and can differ from local resellers.