AI Benchmark

As of · Source Artificial Analysis

Language models Coding & Agents Image · Video · Audio

Highlights

Artificial Analysis Key Metrics

Detailed comparison across three core axes: Intelligence, Time per Task, and Cost per Task

Intelligence (Composite Index)

Updated
1
Fable 5.2
Anthropic
65
2
Astra 6.2
OpenAI
62
3
Gemini 4 Neon
Google
61
4
Opus 5.5
Anthropic
59
5
Sol 6.2
OpenAI
57
6
Muse Wonder Full
Meta
50
7
Kimi K3.1 Pro
Kimi
46
8
Muse Wonder
Meta
44
9
Qwen4 Max
Alibaba
43
10
Grok 4.7
SpaceXAI
35
11
DeepSeek V4.1 Flash 1017
DeepSeek
33
12
GLM-5.3
Z AI
32

Time per Task (Minutes per Task)

Updated
1
Sol 6.2
OpenAI
8.5m
2
DeepSeek V4.1 Flash 1017
DeepSeek
13.5m
3
Muse Wonder
Meta
13.7m
4
Astra 6.2
OpenAI
13.7m
5
Gemini 3.8 Flash
Google
14.2m
6
Gemini 4 Neon
Google
21.5m
7
GLM-5.3
Z AI
23.1m
8
Opus 5.5
Anthropic
24.6m
9
Grok 4.7
SpaceXAI
24.9m
10
Qwen4 Max
Alibaba
31.3m
11
Kimi K3.1 Pro
Kimi
39.3m
12
Fable 5.2
Anthropic
41.2m

Cost per Task (Value for Money)

Updated
1
DeepSeek V4.1 Flash 1017
DeepSeek
$0.60
2
Sol 6.2
OpenAI
$1.01
3
Qwen4 Max
Alibaba
$3.27
4
Kimi K3.1 Pro
Kimi
$3.68
5
Astra 6.2
OpenAI
$4.00
6
GLM-5.3
Z AI
$5.01
7
Gemini 4 Neon
Google
$6.44
8
Muse Wonder
Meta
$6.95
9
Gemini 3.8 Flash
Google
$7.77
10
Grok 4.7
SpaceXAI
$7.77
11
Opus 5.5
Anthropic
$10.04
12
Fable 5.2
Anthropic
$26.54

Coding & Agents

Artificial Analysis Agentic Evals

Real-world repository & terminal harness performance (Coding Agent Index) plus four model-level agentic benchmarks

Coding Agent Harnesses FrontierCode v1.1 · SWE-Atlas-QnA · Terminal-Bench v4 composite

Coding Agent Index (Composite Index)

Updated

Top 10 of 40 harnesses

  1. 1
    Claude Code · Fable 5.2 (Xhigh, Default Fallback)
    Anthropic
    58.5%
  2. 2
    Devin Fusion CLI · Opus 5.5 + GPT-6 Luna
    Cognition
    58.4%
  3. 3
    Claude Code · Opus 5.5
    Anthropic
    58.2%
  4. 4
    Claude Code · Sonnet 5.5
    Anthropic
    56.3%
  5. 5
    Antigravity CLI · Gemini 4 Carbon
    Google
    55.5%
  6. 6
    Muse Code · Wonder Full
    Meta
    54.5%
  7. 7
    Codex · GPT-6.2 Astra
    OpenAI
    53.5%
  8. 8
    Codex · GPT-6 Astra
    OpenAI
    53.1%
  9. 9
    Claude Code · Fable 5.1
    Anthropic
    52.5%
  10. 10
    Antigravity CLI · Gemini 4 Hydrogen
    Google
    52.3%

Execution Time (Minutes per Task · lower is better)

Updated

Top 10 of 40 harnesses

  1. 1
    Codex · GPT-6.2 Sol
    OpenAI
    7.2m
  2. 2
    Codex · GPT-6.2 Astra
    OpenAI
    7.8m
  3. 3
    Claude Code · Fable 5.2 (Low, Default Fallback)
    Anthropic
    13.4m
  4. 4
    Muse Code · Ember
    Meta
    19.9m
  5. 5
    Codex · GPT-6.1 Sol
    OpenAI
    21.9m
  6. 6
    Codex · GPT-6 Luna
    OpenAI
    23.5m
  7. 7
    Muse Code · Spark 1.3
    Meta
    24.4m
  8. 8
    Codex · GPT-6 Astra
    OpenAI
    26.0m
  9. 9
    Muse Code · Wonder
    Meta
    27.2m
  10. 10
    Codex · GPT-6.1 Terra
    OpenAI
    28.1m

Cost per Task (Cost per Task · lower is better)

Top 10 of 39 harnesses

  1. 1
    Codex · GPT-6 Luna
    OpenAI
    $0.19
  2. 2
    Muse Code · Ember
    Meta
    $0.22
  3. 3
    Mirror · Beam 1.0
    Reflection
    $0.31
  4. 4
    DeepSeek Harness · V4.1 Flash 1017
    DeepSeek
    $0.32
  5. 5
    Codex · GPT-6.2 Sol
    OpenAI
    $0.33
  6. 6
    Codex · GPT-6.2 Astra
    OpenAI
    $0.70
  7. 7
    Mistral Vibe · Large 4
    Mistral
    $1.02
  8. 8
    Codex · GPT-6.1 Sol
    OpenAI
    $1.51
  9. 9
    Codex · GPT-6.1 Terra
    OpenAI
    $1.53
  10. 10
    Claude Code · Haiku 5.5
    Anthropic
    $2.25
Model Agentic Evaluations Top 10 releases (best reasoning-effort variant per release)

FrontierCode (Software Engineering)

Updated

Top 10 of 41 releases

  1. 1
    Claude Fable 5.2 (medium)
    Anthropic
    64.0%
  2. 2
    Claude Opus 5.5 (medium)
    Anthropic
    57.1%
  3. 3
    GPT-6.1 Sol (max)
    OpenAI
    49.1%
  4. 4
    Claude Fable 5.1 (max)
    Anthropic
    49.0%
  5. 5
    Kimi K3.1 Pro
    Kimi
    47.1%
  6. 6
    GPT-6 Astra (max)
    OpenAI
    47.0%
  7. 7
    Muse Wonder (max)
    Meta
    46.5%
  8. 8
    Claude Sonnet 5.5 (xhigh)
    Anthropic
    46.1%
  9. 9
    Claude Haiku 5.5 (xhigh)
    Anthropic
    45.5%
  10. 10
    MiMo-V2.6-Pro
    Xiaomi
    45.4%

Terminal-Bench v4 (Agentic Terminal Use)

Updated

Top 10 of 101 releases

  1. 1
    Kimi K3.1 Pro
    Kimi
    67.2%
  2. 2
    Gemini 4 Carbon (high)
    Google
    65.2%
  3. 3
    GPT-6.2 Astra (xhigh)
    OpenAI
    64.1%
  4. 4
    Claude Sonnet 5.5 (max)
    Anthropic
    63.6%
  5. 5
    Gemini 4 Neon (xhigh)
    Google
    63.6%
  6. 6
    Claude Fable 5.2 (xhigh)
    Anthropic
    61.6%
  7. 7
    GPT-6.2 Sol (max)
    OpenAI
    60.1%
  8. 8
    GPT-6 Astra (xhigh)
    OpenAI
    59.6%
  9. 9
    Claude Opus 5.5 (max)
    Anthropic
    59.6%
  10. 10
    Gemini 4 Argon (high)
    Google
    57.1%

AutomationBench-AA (SaaS Workflow Automation)

Updated

Top 10 of 174 releases

  1. 1
    Gemini 4 Neon (xhigh)
    Google
    79.3%
  2. 2
    Gemini 4 Carbon (high)
    Google
    77.8%
  3. 3
    Gemini 4 Argon (high)
    Google
    77.5%
  4. 4
    GPT-6.2 Astra (max)
    OpenAI
    77.2%
  5. 5
    Claude Fable 5.2 (max)
    Anthropic
    76.4%
  6. 6
    Gemini 4 Hydrogen (xhigh)
    Google
    76.1%
  7. 7
    Kimi K3.1 Pro
    Kimi
    73.3%
  8. 8
    Qwen4 Max
    Alibaba
    72.6%
  9. 9
    MiniMax-M3.1-Flash
    MiniMax
    72.5%
  10. 10
    Claude Sonnet 5.5 (max)
    Anthropic
    71.8%

GDPval-AA (Knowledge Work Elo)

Updated

Top 10 of 187 releases

  1. 1
    Claude Opus 5.5 (max)
    Anthropic
    1866
  2. 2
    Kimi K3.1 Pro
    Kimi
    1861
  3. 3
    Muse Wonder Full (max)
    Meta
    1857
  4. 4
    Claude Sonnet 5.5 (max)
    Anthropic
    1838
  5. 5
    Claude Fable 5.2 (max)
    Anthropic
    1831
  6. 6
    Qwen4 Max
    Alibaba
    1809
  7. 7
    Muse Wonder (max)
    Meta
    1770
  8. 8
    Claude Fable 5.1 (max)
    Anthropic
    1765
  9. 9
    DeepSeek V4.1 Flash 1017 (max)
    DeepSeek
    1746
  10. 10
    Gemini 4 Carbon (high)
    Google
    1733

Image · Video · Audio

Artificial Analysis Arena

Top models per field, ranked by blind human-preference votes (Arena Elo) and speech-recognition error rate (WER)

Image (Arena Elo)

Top 10 of 115 models

  1. 1
    GPT Image 2.6 Sunburst
    OpenAI
    1261
  2. 2
    GPT Image 2.5 Sunburst
    OpenAI
    1229
  3. 3
    GPT Image 2.6 Flare
    OpenAI
    1227
  4. 4
    MAI-Image-2.7
    Microsoft AI
    1214
  5. 5
    GPT Image 2.5 Flare
    OpenAI
    1210
  6. 6
    GPT Image 2
    OpenAI
    1192
  7. 7
    Kling Image 4.0
    KlingAI
    1180
  8. 8
    Nano Banana 2.1
    Google
    1176
  9. 9
    MAI-Image-2.6
    Microsoft AI
    1168
  10. 10
    Grok Imagine Image 2.0
    SpaceXAI
    1167

Video (Arena Elo)

Top 10 of 34 models

  1. 1
    Wan 3.0
    Alibaba
    1174
  2. 2
    Utopai X
    Utopai Studios
    1168
  3. 3
    MiniMax H3
    MiniMax
    1159
  4. 4
    Dreamina Seedance 2.5
    ByteDance Seed
    1156
  5. 5
    Hyperflow H3
    Video Rebirth
    1154
  6. 6
    FLUX 3 Max
    Fal
    1152
  7. 7
    Gemini Omni Flash 1.2
    Google
    1151
  8. 8
    HeyGen Video 1
    HeyGen
    1149
  9. 9
    MiniMax H3 Max
    Fal
    1143
  10. 10
    FLUX 3
    Black Forest Labs
    1136

Audio (Arena Elo)

Top 10 of 101 models

  1. 1
    Eleven v4 Turbo
    ElevenLabs
    1341
  2. 2
    Eleven v4
    ElevenLabs
    1341
  3. 3
    Qwen-Audio-3.1-TTS-Plus
    Alibaba
    1299
  4. 4
    Sonic 3.6
    Cartesia
    1295
  5. 5
    Simba 3.5
    SpeechifyAI
    1293
  6. 6
    Gemini 3.8 Flash TTS
    Google
    1283
  7. 7
    Fish Audio Drama 3
    Fish Audio
    1279
  8. 8
    Lightning V3.1 Pro
    Smallest.ai
    1278
  9. 9
    Realtime TTS-2
    Inworld
    1270
  10. 10
    Gemini 3.8 Flash-Lite TTS
    Google
    1262
Methodology & Interpretation

How to Read AI Benchmarks & Choose the Right Model for Production

A single leaderboard score rarely captures production latency, reasoning token overhead, or tool-calling reliability. This reference page pairs independent measurements from Artificial Analysis across frontier language models, coding and workflow agent harnesses, and multimodal generation arenas on a consistent task baseline.

Understanding the Four Evaluation Pillars

LLM · Reasoning

Composite Intelligence Index

A weighted composite across contamination-resistant math, scientific reasoning, multi-step logic, code generation, and instruction following. It reflects a model’s ceiling on complex problem solving rather than memorized trivia.

Latency · Economics

Time per Task & Cost per Task

Raw tokens-per-second or per-million-token pricing hides how many internal reasoning tokens a model burns before answering. Task-level metrics measure total wall-clock minutes and total USD spent to finish the entire benchmark suite.

Coding · Agents

Agent Harnesses & Real-World Engineering

Evaluates end-to-end completion when paired with real shell and repository harnesses—covering FrontierCode, Terminal-Bench, AutomationBench, and GDPval-AA knowledge-work Elo alongside 95% confidence intervals.

Multimodal

Blind Arena Elo & Speech Word Error Rate (AA-WER)

Visual generation and speech synthesis rely on blind pairwise human preference votes (Arena Elo, higher is better), whereas speech-to-text accuracy is measured by Word Error Rate (AA-WER, lower is better).

Model Selection Matrix by Workload

Workload ScenarioPrimary Metrics to CheckPractical Selection Guidance
System architecture, multi-file refactoring & deep researchIntelligence Index · FrontierCode · Terminal-BenchPrioritize frontier reasoning tiers paired with proven coding harnesses; higher accuracy cuts costly human rework.
Interactive customer chat, search synthesis & live UITime per Task (min) · Low reasoning overheadChoose fast non-reasoning or light-reasoning models to keep end-to-end latency snappy under concurrent traffic.
High-volume classification, translation & extraction pipelinesCost per Task (USD) · Accuracy thresholdFilter for models that clear your quality bar, then pick the lowest Cost-per-Task open-weight or efficient commercial tier.
Creative assets, video storyboards & multilingual voiceoversTask-specific Arena Elo (T2I / Edit / I2V / TTS)Models leading text-to-image often differ from those leading image or video editing; mix specialized models per pipeline stage.

Frequently Asked Questions

Why can a model with fast tokens/sec still show a long Time per Task?

Reasoning models generate thousands of hidden thinking tokens before emitting the final answer. Even at high token throughput, generating 5–10× more tokens increases total wall-clock completion time and total cost.

What does the striped Preview bar indicate compared with standard bars?

Preview entries reflect early research checkpoints or limited preview endpoints before general-availability pricing and serving throughput are finalized. Production workloads should verify GA speed and cost figures.

How should I read the 95% confidence intervals on agent benchmarks?

Multi-step terminal and software tasks involve non-deterministic tool execution. When two models’ confidence intervals overlap, treat their capability as tied and decide based on cost, latency, and ecosystem fit.