AI Benchmark
As of · Source Artificial Analysis
Language models Coding & Agents Image · Video · Audio
Highlights
Artificial Analysis Key MetricsDetailed comparison across three core axes: Intelligence, Time per Task, and Cost per Task
Intelligence (Composite Index)
Time per Task (Minutes per Task)
Cost per Task (Value for Money)
Coding & Agents
Artificial Analysis Agentic EvalsReal-world repository & terminal harness performance (Coding Agent Index) plus four model-level agentic benchmarks
Coding Agent Index (Composite Index)
Top 10 of 40 harnesses
- 1Claude Code · Fable 5.2 (Xhigh, Default Fallback)Anthropic58.5%
- 2Devin Fusion CLI · Opus 5.5 + GPT-6 LunaCognition58.4%
- 3Claude Code · Opus 5.5Anthropic58.2%
- 4Claude Code · Sonnet 5.5Anthropic56.3%
- 5Antigravity CLI · Gemini 4 CarbonGoogle55.5%
- 6Muse Code · Wonder FullMeta54.5%
- 7Codex · GPT-6.2 AstraOpenAI53.5%
- 8Codex · GPT-6 AstraOpenAI53.1%
- 9Claude Code · Fable 5.1Anthropic52.5%
- 10Antigravity CLI · Gemini 4 HydrogenGoogle52.3%
Execution Time (Minutes per Task · lower is better)
Top 10 of 40 harnesses
- 1Codex · GPT-6.2 SolOpenAI7.2m
- 2Codex · GPT-6.2 AstraOpenAI7.8m
- 3Claude Code · Fable 5.2 (Low, Default Fallback)Anthropic13.4m
- 4Muse Code · EmberMeta19.9m
- 5Codex · GPT-6.1 SolOpenAI21.9m
- 6Codex · GPT-6 LunaOpenAI23.5m
- 7Muse Code · Spark 1.3Meta24.4m
- 8Codex · GPT-6 AstraOpenAI26.0m
- 9Muse Code · WonderMeta27.2m
- 10Codex · GPT-6.1 TerraOpenAI28.1m
Cost per Task (Cost per Task · lower is better)
Top 10 of 39 harnesses
- 1Codex · GPT-6 LunaOpenAI$0.19
- 2Muse Code · EmberMeta$0.22
- 3Mirror · Beam 1.0Reflection$0.31
- 4DeepSeek Harness · V4.1 Flash 1017DeepSeek$0.32
- 5Codex · GPT-6.2 SolOpenAI$0.33
- 6Codex · GPT-6.2 AstraOpenAI$0.70
- 7Mistral Vibe · Large 4Mistral$1.02
- 8Codex · GPT-6.1 SolOpenAI$1.51
- 9Codex · GPT-6.1 TerraOpenAI$1.53
- 10Claude Code · Haiku 5.5Anthropic$2.25
FrontierCode (Software Engineering)
Top 10 of 41 releases
- 1Claude Fable 5.2 (medium)Anthropic64.0%
- 2Claude Opus 5.5 (medium)Anthropic57.1%
- 3GPT-6.1 Sol (max)OpenAI49.1%
- 4Claude Fable 5.1 (max)Anthropic49.0%
- 5Kimi K3.1 ProKimi47.1%
- 6GPT-6 Astra (max)OpenAI47.0%
- 7Muse Wonder (max)Meta46.5%
- 8Claude Sonnet 5.5 (xhigh)Anthropic46.1%
- 9Claude Haiku 5.5 (xhigh)Anthropic45.5%
- 10MiMo-V2.6-ProXiaomi45.4%
Terminal-Bench v4 (Agentic Terminal Use)
Top 10 of 101 releases
- 1Kimi K3.1 ProKimi67.2%
- 2Gemini 4 Carbon (high)Google65.2%
- 3GPT-6.2 Astra (xhigh)OpenAI64.1%
- 4Claude Sonnet 5.5 (max)Anthropic63.6%
- 5Gemini 4 Neon (xhigh)Google63.6%
- 6Claude Fable 5.2 (xhigh)Anthropic61.6%
- 7GPT-6.2 Sol (max)OpenAI60.1%
- 8GPT-6 Astra (xhigh)OpenAI59.6%
- 9Claude Opus 5.5 (max)Anthropic59.6%
- 10Gemini 4 Argon (high)Google57.1%
AutomationBench-AA (SaaS Workflow Automation)
Top 10 of 174 releases
- 1Gemini 4 Neon (xhigh)Google79.3%
- 2Gemini 4 Carbon (high)Google77.8%
- 3Gemini 4 Argon (high)Google77.5%
- 4GPT-6.2 Astra (max)OpenAI77.2%
- 5Claude Fable 5.2 (max)Anthropic76.4%
- 6Gemini 4 Hydrogen (xhigh)Google76.1%
- 7Kimi K3.1 ProKimi73.3%
- 8Qwen4 MaxAlibaba72.6%
- 9MiniMax-M3.1-FlashMiniMax72.5%
- 10Claude Sonnet 5.5 (max)Anthropic71.8%
GDPval-AA (Knowledge Work Elo)
Top 10 of 187 releases
- 1Claude Opus 5.5 (max)Anthropic1866
- 2Kimi K3.1 ProKimi1861
- 3Muse Wonder Full (max)Meta1857
- 4Claude Sonnet 5.5 (max)Anthropic1838
- 5Claude Fable 5.2 (max)Anthropic1831
- 6Qwen4 MaxAlibaba1809
- 7Muse Wonder (max)Meta1770
- 8Claude Fable 5.1 (max)Anthropic1765
- 9DeepSeek V4.1 Flash 1017 (max)DeepSeek1746
- 10Gemini 4 Carbon (high)Google1733
Image · Video · Audio
Artificial Analysis ArenaTop 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
- 1GPT Image 2.6 SunburstOpenAI1261
- 2GPT Image 2.5 SunburstOpenAI1229
- 3GPT Image 2.6 FlareOpenAI1227
- 4MAI-Image-2.7Microsoft AI1214
- 5GPT Image 2.5 FlareOpenAI1210
- 6GPT Image 2OpenAI1192
- 7Kling Image 4.0KlingAI1180
- 8Nano Banana 2.1Google1176
- 9MAI-Image-2.6Microsoft AI1168
- 10Grok Imagine Image 2.0SpaceXAI1167
Video (Arena Elo)
Top 10 of 34 models
- 1Wan 3.0Alibaba1174
- 2Utopai XUtopai Studios1168
- 3MiniMax H3MiniMax1159
- 4Dreamina Seedance 2.5ByteDance Seed1156
- 5Hyperflow H3Video Rebirth1154
- 6FLUX 3 MaxFal1152
- 7Gemini Omni Flash 1.2Google1151
- 8HeyGen Video 1HeyGen1149
- 9MiniMax H3 MaxFal1143
- 10FLUX 3Black Forest Labs1136
Audio (Arena Elo)
Top 10 of 101 models
- 1Eleven v4 TurboElevenLabs1341
- 2Eleven v4ElevenLabs1341
- 3Qwen-Audio-3.1-TTS-PlusAlibaba1299
- 4Sonic 3.6Cartesia1295
- 5Simba 3.5SpeechifyAI1293
- 6Gemini 3.8 Flash TTSGoogle1283
- 7Fish Audio Drama 3Fish Audio1279
- 8Lightning V3.1 ProSmallest.ai1278
- 9Realtime TTS-2Inworld1270
- 10Gemini 3.8 Flash-Lite TTSGoogle1262
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
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.
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.
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.
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 Scenario | Primary Metrics to Check | Practical Selection Guidance |
|---|---|---|
| System architecture, multi-file refactoring & deep research | Intelligence Index · FrontierCode · Terminal-Bench | Prioritize frontier reasoning tiers paired with proven coding harnesses; higher accuracy cuts costly human rework. |
| Interactive customer chat, search synthesis & live UI | Time per Task (min) · Low reasoning overhead | Choose fast non-reasoning or light-reasoning models to keep end-to-end latency snappy under concurrent traffic. |
| High-volume classification, translation & extraction pipelines | Cost per Task (USD) · Accuracy threshold | Filter 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 voiceovers | Task-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.