LLM Leaderboard
Ranked by Humanity's Last Exam · source AA
Frontier reasoning benchmark; most models still score low.
#1 by Humanity's Last Exam
Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)
53.3%
Anthropic · in $10.0 / out $50.0 per 1M · 57 tok/s
| # | Model | HLE |
|---|---|---|
| 1 | Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) | 53.3% |
| 2 | GPT-5.6 Sol (max) | 47.2% |
| 3 | Claude Opus 4.8 (Adaptive Reasoning, Max Effort) | 45.7% |
| 4 | Muse Spark 1.1 (xhigh) | 45.1% |
| 5 | GPT-5.6 Sol (xhigh) | 44.7% |
| 6 | Gemini 3.1 Pro Preview | 44.7% |
| 7 | Kimi K3 | 44.3% |
| 8 | GPT-5.5 (xhigh) | 44.3% |
| 9 | GPT-5.6 Sol (high) | 44.1% |
| 10 | GPT-5.5 (high) | 43.0% |
| 11 | GPT-5.6 Terra (max) | 41.8% |
| 12 | GPT-5.4 (xhigh) | 41.6% |
| 13 | Gemini 3.5 Flash (high) | 41.0% |
| 14 | GPT-5.5 (medium) | 40.6% |
| 15 | Grok 4.5 (high) | 40.3% |
| 16 | GLM-5.2 (max) | 40.1% |
| 17 | GPT-5.6 Terra (xhigh) | 40.0% |
| 18 | Gemini 3.5 Flash (medium) | 39.9% |
| 19 | GPT-5.3 Codex (xhigh) | 39.9% |
| 20 | Muse Spark | 39.9% |
| 21 | GPT-5.6 Sol (medium) | 39.7% |
| 22 | Claude Opus 4.7 (Adaptive Reasoning, Max Effort) | 39.6% |
| 23 | Claude Sonnet 5 (Adaptive Reasoning, Max Effort) | 39.6% |
| 24 | Qwen3.7 Max | 38.1% |
| 25 | GPT-5.6 Luna (max) | 37.2% |
| 26 | Gemini 3 Pro Preview (high) | 37.2% |
| 27 | MiniMax-M3 | 37.1% |
| 28 | GPT-5.6 Terra (high) | 36.7% |
| 29 | Claude Opus 4.6 (Adaptive Reasoning, Max Effort) | 36.7% |
| 30 | GPT-5.6 Sol (low) | 36.6% |
| 31 | Grok Build 0.1 0616 | 36.0% |
| 32 | DeepSeek V4 Pro (Reasoning, Max Effort) | 35.9% |
| 33 | Kimi K2.6 | 35.9% |
| 34 | GPT-5.6 Luna (xhigh) | 35.6% |
| 35 | GPT-5.2 (xhigh) | 35.4% |
| 36 | Grok 4.3 (high) | 35.0% |
| 37 | Gemini 3 Flash Preview (Reasoning) | 34.7% |
| 38 | MiMo-V2.5-Pro | 33.8% |
| 39 | DeepSeek V4 Pro (Reasoning, High Effort) | 33.5% |
| 40 | GPT-5.2 Codex (xhigh) | 33.5% |
| 41 | Qwen3.7 Plus | 33.4% |
| 42 | KAT-Coder-Pro V1 | 33.4% |
| 43 | Kimi K2.7 Code | 32.8% |
| 44 | Nex-N2-Pro | 32.4% |
| 45 | Grok 4.20 0309 v2 (Reasoning) | 32.2% |
| 46 | DeepSeek V4 Flash (Reasoning, Max Effort) | 32.1% |
| 47 | LongCat 2.0 | 32.1% |
| 48 | GPT-5.6 Luna (high) | 31.6% |
| 49 | GPT-5.6 Terra (medium) | 31.6% |
| 50 | Hy3 | 31.6% |
Quality, pricing, and speed benchmarks via Artificial Analysis. Human preference Elo via LMArena. Refreshed daily at 06:00 UTC · last sync .
About the metrics
Intelligence Index
AA
Artificial Analysis composite of reasoning, knowledge, coding, math.
LMArena Elo
LMArena
Human preference Elo from chatbot arena head-to-head votes.
Coding Index
AA
AA composite — LiveCodeBench, SciCode, etc.
Math Index
AA
AA composite — MATH-500, AIME, etc.
MMLU-Pro
AA
Multi-domain knowledge & reasoning. Harder than MMLU.
GPQA Diamond
AA
Graduate-level science Q&A. Strong reasoning signal.
LiveCodeBench
AA
Competitive programming, contamination-resistant.
MATH-500
AA
Subset of the MATH benchmark.
AIME
AA
American Invitational Mathematics Examination — olympiad math.
Humanity's Last Exam
AA
Frontier reasoning benchmark; most models still score low.
SciCode
AA
Scientific coding tasks.