01
Benchmarks set the shortlist
Independent scores for intelligence, speed, and price narrow every lab’s catalog to a few candidates for the job.
Model agnostic
No lab leads on every job. We stay model agnostic, and choose per job on evidence.
The frontier moves every few weeks. One lab holds the top score this month and another the next, and the fastest or cheapest model is rarely the smartest. Staying model agnostic is how we keep each system on the best option for its job, and off any one vendor’s roadmap.
Below is the independent view we start from: benchmarks of intelligence, speed, and price from Artificial Analysis, refreshed twice a day and drawn in our own charts. Then we test the shortlist on your cases.
At a glance
The frontier, by the numbers.
57.6
The top Artificial Analysis Intelligence Index score today: Claude Opus 5.5 (Max, Default Fallback), from Anthropic.
4 labs
appear on the release-date frontier since Apr 2024, using current scores.
58
language-model families released in the last 90 days with a published Intelligence Index.
118
labs and providers across language, image, video, and speech. We can build with any of them.
Source: Artificial Analysis (artificialanalysis.ai), refreshed twice a day. A flagship is each lab’s highest-scoring model. * marks an AA estimated Intelligence Index.
Flagships
Every lab’s best model, on three measures.
The twelve labs with the highest-scoring models today, each by its flagship. The order changes with the measure: the smartest model is rarely the fastest or the cheapest.
Artificial Analysis Intelligence Index. Higher is better.
- 01AnthropicClaude Opus 5.5 (Max, Default Fallback)57.6
- 02MetaMuse Wonder Full (Max)55.8
- 03OpenAIGPT-6 Astra (Max)52.7
- 04GoogleGemini 4 Argon (High)52.6
- 05SpaceXAIGrok 4.7 (Xhigh)46.4
- 06XiaomiMiMo-V2.6-Pro46.3
- 07AlibabaQwen3.8 Max (0902)45.4
- 08Z AIGLM-5.3 (Max)44.8
- 09StepFunStep 5 Preview43.7
- 10KimiKimi K3 (Max)43.6
- 11InclusionAILing 3.1 Flash41.1
- 12DeepSeekDeepSeek V4.1 Flash (Max)39.5
Output tokens per second, the median published in AA’s website model measurements. Higher is better.
- 01InclusionAILing 3.1 Flash211
- 02DeepSeekDeepSeek V4.1 Flash (Max)185
- 03AnthropicClaude Opus 5.5 (Max, Default Fallback)95
- 04StepFunStep 5 Preview91
- 05SpaceXAIGrok 4.7 (Xhigh)82
- 06Z AIGLM-5.3 (Max)61
- 07OpenAIGPT-6 Astra (Max)57
- 08XiaomiMiMo-V2.6-Pro46
- 09KimiKimi K3 (Max)45
- 10AlibabaQwen3.8 Max (0902)42
- ·MetaMuse Wonder Full (Max)Not measured
- ·GoogleGemini 4 Argon (High)Not measured
US dollars per million tokens, a 3:1 blend of input and output. Lower is better.
- 01InclusionAILing 3.1 Flash$0.45
- 02DeepSeekDeepSeek V4.1 Flash (Max)$0.52
- 03XiaomiMiMo-V2.6-Pro$0.54
- 04StepFunStep 5 Preview$1.43
- 05Z AIGLM-5.3 (Max)$2.15
- 06SpaceXAIGrok 4.7 (Xhigh)$3.00
- 07AlibabaQwen3.8 Max (0902)$3.00
- 08GoogleGemini 4 Argon (High)$4.00
- 09KimiKimi K3 (Max)$6.00
- 10AnthropicClaude Opus 5.5 (Max, Default Fallback)$8.00
- 11OpenAIGPT-6 Astra (Max)$20
- ·MetaMuse Wonder Full (Max)Not measured
Over time
The frontier, by release date.
Using current benchmark scores and model release dates, 21 releases since Apr 2024 raise the frontier, from OpenAI, Anthropic, SpaceXAI, and Google. Retired models stay in this history, marked. Pick a lab above to see its own line.
Releases that raise the frontier on current scores
| Released | Model | Lab | Index |
|---|---|---|---|
| Sep 2026 | Claude Opus 5.5 (Max, Default Fallback) | Anthropic | 57.6 |
| Sep 2026 | Claude Fable 5.1 (Max, Default Fallback) | Anthropic | 53.4 |
| Jul 2026 | Claude Opus 5 (Max) retired | Anthropic | 50.8 |
| Jun 2026 | Claude Fable 5 (Max, Opus 4.8 Fallback) retired | Anthropic | 49.6 |
| May 2026 | Claude Opus 4.8 (Max) retired | Anthropic | 41.8 |
| Apr 2026 | Claude Opus 4.7 (Max) retired | Anthropic | 40.7* |
| Mar 2026 | GPT-5.4 (Xhigh) retired | OpenAI | 39.0* |
| Feb 2026 | GPT-5.3 Codex (Xhigh) | OpenAI | 32.5* |
| Dec 2025 | GPT-5.2 (Xhigh) retired | OpenAI | 30.4* |
| Nov 2025 | Claude Opus 4.5 (Reasoning) retired | Anthropic | 29.1* |
| Nov 2025 | Gemini 3 Pro Preview (High) retired | 28.0* | |
| Sep 2025 | GPT-5 Codex (High) retired | OpenAI | 24.9* |
| Aug 2025 | GPT-5 (High) retired | OpenAI | 23.0* |
| Aug 2025 | Claude 4.1 Opus (Reasoning) retired | Anthropic | 22.8* |
| Jul 2025 | Grok 4 retired | SpaceXAI | 22.5* |
| Jun 2025 | o3-pro retired | OpenAI | 21.9* |
| May 2025 | Claude 4 Opus (Reasoning) retired | Anthropic | 20.6* |
| Apr 2025 | o3 | OpenAI | 20.2* |
| Feb 2025 | Claude 3.7 Sonnet (Reasoning) retired | Anthropic | 17.7* |
| Dec 2024 | o1 retired | OpenAI | 15.2* |
| Sep 2024 | o1-preview retired | OpenAI | 11.4* |
Value and speed
The smartest model is rarely the best value, or the fastest.
86 current language models released in the last 12 months. The orange line is the frontier: the most intelligence available at each price, or at each speed. It runs through 14 models from 10 labs. Ringed dots are each top lab’s flagship. Hover, or tab to a chart and use the arrow keys, for any model.
Price is a 3:1 blend of input and output price per million tokens, in US dollars, on a log scale. Reasoning models spend more tokens on each task, so their cost per task runs higher than their token price alone suggests.
Speed is output tokens per second, the median published in AA’s website model measurements. Reasoning models also think before they answer, so a fast model is not always the quickest to finish a task.
The models on either frontier
| Model | Lab | Frontier | Index | Price | Tokens/s |
|---|---|---|---|---|---|
| Claude Opus 5.5 (Max, Default Fallback) | Anthropic | Price and Speed | 57.6 | $8.00 | 95 |
| Claude Sonnet 5.5 (Max, Default Fallback) | Anthropic | Price and Speed | 56.0 | $4.00 | 135 |
| Muse Spark 1.3 (Max) | Meta | Price | 48.1 | $2.00 | 128 |
| MiMo-V2.6-Pro | Xiaomi | Price | 46.3 | $0.54 | 46 |
| Claude Haiku 5.5 (Max) | Anthropic | Price and Speed | 43.4 | $0.20 | 237 |
| MiMo-V2.6-Flash | Xiaomi | Price | 37.9 | $0.18 | 63 |
| Agnes 2.5 Pro Beta | Sapiens AI | Price | 35.2* | $0.15 | 102 |
| Ling-3.0-flash-Fin | InclusionAI | Price and Speed | 22.6 | $0.11 | 286 |
| Gemini 3.5 Flash-Lite | Speed | 22.2 | $0.85 | 425 | |
| Qwen3.5 4B (Reasoning) | Alibaba | Price | 13.1* | $0.06 | 24 |
| Mercury 2.5 | Inception | Speed | 12.3 | $0.38 | 766 |
| Granite 4.2 3B | IBM | Price | 9.1 | $0.05 | 215 |
| Gemma 4 E4B (Reasoning) | Price | 8.9* | $0.04 | 43 | |
| Celeris-1 | Celeris | Speed | 6.3 | $0.33 | 1532 |
Beyond language
Image, video, and speech, too.
The same discipline for generative media. Rankings are arena preference scores (Elo) from blind, side-by-side votes, so they reflect what people prefer, not what a lab claims.
Text to image · 102 ranked Image
01 GPT Image 2.5 Sunburst (max) OpenAI
1217 Elo
02 GPT Image 2.5 Flare (max) OpenAI
1209 Elo
03 GPT Image 2 (high) OpenAI
1188 Elo
04 Nano Banana 2.1 Google
1173 Elo
05 Grok Imagine Image 2.0 SpaceXAI
1167 Elo
Updated .
Image editing · 69 ranked Editing
01 GPT Image 2.5 Sunburst (max) OpenAI
1198 Elo
02 GPT Image 2.5 Flare (max) OpenAI
1178 Elo
03 MAI-Image-2.6 Microsoft AI
1151 Elo
04 Nano Banana 2.1 Google
1150 Elo
05 MAI-Image-2.6-Flash Microsoft AI
1140 Elo
Updated .
Text to video · 32 ranked Video
01 Wan 3.0 Alibaba
1134 Elo
02 MiniMax H3 (768p) MiniMax
1121 Elo
03 MiniMax H3 Max Fal
1117 Elo
04 Dreamina Seedance 2.5 ByteDance Seed
1105 Elo
05 FLUX 3 Black Forest Labs
1095 Elo
Updated .
Image to video · 77 ranked Motion
01 Gemini Omni Flash Google
1368 Elo
02 Bach 1.0 Pro Video Rebirth
1362 Elo
03 Wan 3.0 Alibaba
1362 Elo
04 MiniMax H3 MiniMax
1356 Elo
05 Vimoo 1.0 Vimoo
1348 Elo
Updated .
Text to speech · 97 ranked Speech
01 Eleven v4 Turbo ElevenLabs
1330 Elo
02 Eleven v4 ElevenLabs
1328 Elo
03 Qwen-Audio-3.1-TTS-Plus Alibaba
1300 Elo
04 Sonic 3.6 Cartesia
1286 Elo
05 Gemini 3.8 Flash TTS Google
1274 Elo
Updated .
Task economics
What a task costs. How long it takes.
Each leading lab’s flagship, using AA’s weighted Intelligence Index task measurements. Time is decode time in minutes and excludes first-token latency and overhead.
| Flagship | Lab | USD / task | Minutes / task |
|---|---|---|---|
| Claude Opus 5.5 (Max, Default Fallback) | Anthropic | $5.982 | 13.15 |
| Muse Wonder Full (Max) | Meta | · | · |
| GPT-6 Astra (Max) | OpenAI | $3.258 | 7.84 |
| Gemini 4 Argon (High) | $1.990 | · | |
| Grok 4.7 (Xhigh) | SpaceXAI | $3.738 | 15.85 |
| MiMo-V2.6-Pro | Xiaomi | $0.133 | 22.61 |
| Qwen3.8 Max (0902) | Alibaba | $5.409 | 32.93 |
| GLM-5.3 (Max) | Z AI | $2.006 | 13.81 |
| Step 5 Preview | StepFun | $1.026 | 10.81 |
| Kimi K3 (Max) | Kimi | $2.000 | 16.01 |
| Ling 3.1 Flash | InclusionAI | $1.002 | 6.52 |
| DeepSeek V4.1 Flash (Max) | DeepSeek | $0.265 | 5.53 |
Source: Artificial Analysis (artificialanalysis.ai). Compare task cost or task time across all models.
AA website snapshot: . 699 models and settings; 186 with cost per task; 183 with time per task; 157 in AA's recommender cohort; 430 marked retired (deprecated) by AA. A dot means the source has not published that measurement. * marks an AA estimated Intelligence Index.
The directory
Every lab we can build with.
118 labs and providers, ranked by their best current language model, then by their best arena rank in image, video, and speech. Language models counts every model and reasoning setting listed. A lab whose language models are all retired shows its best retired one, marked. * marks an AA estimated Intelligence Index.
| # | Lab | Flagship language model | Intelligence | Language models | Image | Video | Speech |
|---|---|---|---|---|---|---|---|
| 01 | Anthropic | Claude Opus 5.5 (Max, Default Fallback) | 57.6 | 63 | · | · | · |
| 02 | Meta | Muse Wonder Full (Max) | 55.8 | 24 | #7 | · | · |
| 03 | OpenAI | GPT-6 Astra (Max) | 52.7 | 108 | #1 | · | #37 |
| 04 | Gemini 4 Argon (High) | 52.6 | 66 | #4 | #1 | #5 | |
| 05 | SpaceXAI | Grok 4.7 (Xhigh) | 46.4 | 29 | #5 | #8 | #27 |
| 06 | Xiaomi | MiMo-V2.6-Pro | 46.3 | 11 | · | · | #61 |
| 07 | Alibaba | Qwen3.8 Max (0902) | 45.4 | 89 | #17 | #1 | #3 |
| 08 | Z AI | GLM-5.3 (Max) | 44.8 | 23 | #65 | · | · |
| 09 | StepFun | Step 5 Preview | 43.7 | 5 | #40 | · | #15 |
| 10 | Kimi | Kimi K3 (Max) | 43.6 | 11 | · | · | · |
| 11 | InclusionAI | Ling 3.1 Flash | 41.1 | 13 | #43 | · | · |
| 12 | DeepSeek | DeepSeek V4.1 Flash (Max) | 39.5 | 37 | · | · | · |
| 13 | Mistral | Mistral Large 4 Preview | 38.4 | 33 | · | · | #45 |
| 14 | Sapiens AI | Agnes 2.5 Pro Beta | 35.2* | 1 | · | #9 | · |
| 15 | China Mobile | JT-4.1 Flash 236B A21B (Reasoning) | 33.9 | 4 | · | · | · |
| 16 | Motif Technologies | Motif 3 | 33.6* | 3 | · | · | · |
| 17 | Institute of Foundation Models | K2 Horizon 375B A23B | 30.5 | 9 | · | · | · |
| 18 | MiniMax | MiniMax-M3 | 29.2 | 7 | · | #2 | #21 |
| 19 | Nex AGI | Nex-N2-Pro (Based on Qwen3.5-397B-A17B) | 28.2* | 1 | · | · | · |
| 20 | Upstage | Solar Pro 4 | 28.2* | 10 | · | · | · |
| 21 | Multiverse Computing | Quasar 438B (Max, Based on GLM-5.2) | 26.7 | 2 | · | · | · |
| 22 | Apodex | Apodex 1.1 | 26.4 | 1 | · | · | · |
| 23 | Thinking Machines | Inkling Small | 25.7 | 2 | · | · | · |
| 24 | Tencent | Hy3 | 25.3 | 3 | #17 | #61 | · |
| 25 | NVIDIA | Nemotron 3 Ultra 550B A55B (Reasoning) | 22.9 | 19 | #47 | #20 | #55 |
| 26 | AI9Stars | G9v3-39A5B | 21.8* | 2 | · | · | · |
| 27 | KwaiKAT | KAT Coder Pro V2 | 21.7* | 2 | · | · | · |
| 28 | SK Telecom | A.X-K2 | 21.5 | 1 | · | · | · |
| 29 | LG AI Research | K-EXAONE 2.0 0803 | 19.7* | 9 | · | · | · |
| 30 | LongCat | LongCat 2.0 | 19.1 | 2 | · | · | · |
| 31 | ByteDance Seed | Doubao Seed Code | 16.9* | 2 | #9 | #4 | · |
| 32 | Baidu | ERNIE 5.0 Thinking Preview | 14.3* | 2 | #76 | · | · |
| 33 | Amazon | Nova 2.0 Pro Preview (Medium) | 14.2* | 14 | · | · | #54 |
| 34 | ServiceNow | Apriel-v1.6-15B-Thinker | 13.4* | 2 | · | · | · |
| 35 | Cohere | Command A+ | 13.1 | 6 | · | · | · |
| 36 | IBM | Granite 4.2 30B | 12.7 | 13 | · | · | · |
| 37 | OpenBMB | MiniCPM5-2B | 12.5 | 4 | · | · | · |
| 38 | Inception | Mercury 2.5 | 12.3 | 2 | · | · | · |
| 39 | Naver | HyperCLOVA X SEED Think (32B) | 11.4* | 1 | · | · | · |
| 40 | Korea Telecom | Mi:dm K 2.5 Pro | 11.0* | 2 | · | · | · |
| 41 | Arcee AI | Trinity Large Thinking | 10.8 | 1 | · | · | · |
| 42 | Prime Intellect | INTELLECT-3 (Based on GLM-4.5-Air) | 10.6* | 1 | · | · | · |
| 43 | Trillion Labs | Tri-21B-Think | 9.0* | 2 | · | · | · |
| 44 | Sarvam | Sarvam 105B (High) | 8.8* | 3 | · | · | · |
| 45 | Nanbeige | Nanbeige4.1-3B | 8.4* | 1 | · | · | · |
| 46 | Liquid AI | LFM2.5-2.6B | 8.4* | 10 | · | · | · |
| 47 | Nous Research | Hermes 4 - Llama-3.1 70B (Reasoning) | 7.9* | 7 | · | · | · |
| 48 | TII UAE | Falcon-H1R-7B | 7.8* | 1 | · | · | · |
| 49 | Allen Institute for AI | Olmo 3.1 32B Think | 7.1* | 10 | · | · | · |
| 50 | Perplexity | R1 1776 | 6.4* | 5 | · | · | · |
| 51 | Celeris | Celeris-1 | 6.3 | 1 | · | · | · |
| 52 | Microsoft | Phi-4 Mini Instruct | 6.3* | 4 | #3 | · | #30 |
| 53 | AI21 Labs | Jamba 1.7 Large | 6.1* | 7 | · | · | · |
| 54 | Reka AI | Reka Flash 3 | 5.6* | 2 | · | · | · |
| 55 | Swiss AI Initiative | Apertus 70B Instruct | 5.1* | 2 | · | · | · |
| 56 | Snowflake | Arctic Instruct retired | 5.4* | 1 | · | · | · |
| 57 | OpenChat | OpenChat 3.5 (1210) retired | 5.3* | 1 | · | · | · |
| 58 | Databricks | DBRX Instruct retired | 5.3* | 1 | · | · | · |
| 59 | ElevenLabs | Media models | · | · | · | · | #1 |
| 60 | Video Rebirth | Media models | · | · | · | #2 | · |
| 61 | Fal | Media models | · | · | #35 | #3 | · |
| 62 | Cartesia | Media models | · | · | · | · | #4 |
| 63 | Black Forest Labs | Media models | · | · | #13 | #5 | · |
| 64 | Vimoo | Media models | · | · | · | #5 | · |
| 65 | Inworld | Media models | · | · | · | · | #7 |
| 66 | PixVerse | Media models | · | · | · | #7 | · |
| 67 | Skywork AI | Media models | · | · | · | #7 | · |
| 68 | SpeechifyAI | Media models | · | · | · | · | #9 |
| 69 | Alibaba-ATH | Media models | · | · | · | #10 | · |
| 70 | Fish Audio | Media models | · | · | · | · | #10 |
| 71 | BreezeBlue | Media models | · | · | · | · | #11 |
| 72 | KlingAI | Media models | · | · | #52 | #11 | · |
| 73 | VUI Labs | Media models | · | · | · | · | #12 |
| 74 | Vidu | Media models | · | · | #54 | #13 | · |
| 75 | Sand.ai | Media models | · | · | · | #14 | · |
| 76 | Soniox | Media models | · | · | · | · | #19 |
| 77 | Luma Labs | Media models | · | · | #20 | #43 | · |
| 78 | Krea | Media models | · | · | #21 | · | · |
| 79 | Gradium | Media models | · | · | · | · | #22 |
| 80 | Ideogram | Media models | · | · | #22 | · | · |
| 81 | Lightricks | Media models | · | · | · | #22 | · |
| 82 | Smallest.ai | Media models | · | · | · | · | #23 |
| 83 | Murf AI | Media models | · | · | · | · | #24 |
| 84 | Runway | Media models | · | · | · | #24 | · |
| 85 | Recraft | Media models | · | · | #25 | · | · |
| 86 | async | Media models | · | · | · | · | #28 |
| 87 | Newport AI | Media models | · | · | #29 | · | · |
| 88 | Pika Art | Media models | · | · | · | #30 | · |
| 89 | Creatify | Media models | · | · | · | #31 | · |
| 90 | HiDream | Media models | · | · | #31 | #53 | · |
| 91 | Pruna AI | Media models | · | · | #45 | #32 | · |
| 92 | Resemble AI | Media models | · | · | · | · | #38 |
| 93 | Hithink | Media models | · | · | · | · | #40 |
| 94 | Reve | Media models | · | · | #48 | · | · |
| 95 | Maya Research | Media models | · | · | · | · | #53 |
| 96 | Midjourney | Media models | · | · | · | #55 | · |
| 97 | Kokoro | Media models | · | · | · | · | #56 |
| 98 | Bria | Media models | · | · | #58 | · | · |
| 99 | Rime | Media models | · | · | · | · | #60 |
| 100 | Deepgram | Media models | · | · | · | · | #62 |
| 101 | Hume AI | Media models | · | · | · | · | #63 |
| 102 | Meituan | Media models | · | · | #63 | · | · |
| 103 | VectorSpaceLab | Media models | · | · | #66 | · | · |
| 104 | Bytedance | Media models | · | · | #67 | · | · |
| 105 | Boson AI | Media models | · | · | · | · | #69 |
| 106 | Bland AI | Media models | · | · | · | · | #70 |
| 107 | Leonardo.Ai | Media models | · | · | #84 | #70 | · |
| 108 | Moonvalley | Media models | · | · | · | #73 | · |
| 109 | Krafton | Media models | · | · | · | · | #79 |
| 110 | Zyphra | Media models | · | · | · | · | #80 |
| 111 | OpenVoice | Media models | · | · | · | · | #83 |
| 112 | Neuphonic | Media models | · | · | · | · | #85 |
| 113 | KugelAudio | Media models | · | · | · | · | #86 |
| 114 | Coqui | Media models | · | · | · | · | #89 |
| 115 | StyleTTS | Media models | · | · | · | · | #93 |
| 116 | Noiz | Media models | · | · | · | · | #95 |
| 117 | MetaVoice | Media models | · | · | · | · | #96 |
| 118 | Deep Cogito | No Intelligence Index published | · | 1 | · | · | · |
Source: Artificial Analysis (artificialanalysis.ai). Want every language model and setting, sortable? Open the model explorer →
How we choose
Benchmarks start the choice. Your cases finish it.
02
Your cases decide
An evaluation harness runs the shortlist on your own data and tasks, so quality is measured, not assumed.
03
Route by the job
The strongest model where the stakes are high, faster and cheaper models where they are enough, with cost and latency set before launch.
04
Re-run when the frontier moves
When a better model ships, we re-run the harness and switch if the numbers say so. A configuration change, not a rebuild.
About the data
Data from Artificial Analysis, an independent firm that benchmarks AI models and providers, using its website for language models and its data API for media arenas. Last refreshed ; we check twice a day.
Intelligence is the Artificial Analysis Intelligence Index. Speed is the output tokens per second published in AA’s website model measurements. Price is a 3:1 blend of input and output price per million tokens, in US dollars. Media rankings are arena Elo scores from Artificial Analysis’s data API. Artificial Analysis’s website also ranks video editing and controlled-voice speech; those rankings are not shown here. Charts show each model family at its highest-scoring reasoning setting. The directory counts individual models and reasoning settings, including models without a published Intelligence Index.
Models Artificial Analysis marks as deprecated are labelled retired. They stay in the release-date history, the record table and the explorer, but are not chosen as a lab’s flagship, today’s top score, or a price or speed frontier model, unless a lab has nothing current. New labs, models, settings and benchmarks appear as Artificial Analysis publishes them, and a new media arena appears once Artificial Analysis lists it on its website and serves it through its data API; if a source fails a check, the page keeps its last good data and says so.
The scores and prices are Artificial Analysis’s; the selection, charts, and commentary are ours. Model and lab names appear as Artificial Analysis lists them. Charts, rankings, and scores marked as ours are by Promise & Proof, based on data from Artificial Analysis; they do not represent the views or endorsement of Artificial Analysis.
Questions
Questions, answered.
Which AI labs and models do you work with?
We are model agnostic: we can build with any lab on this page, closed or open-weight: Anthropic, OpenAI, Google, Meta, Mistral, and the open-weight labs, among them DeepSeek, Alibaba’s Qwen, Kimi, Z AI, and MiniMax, plus the image, video, and speech specialists. No single vendor gets a default; the job decides.
How do you choose a model for a job?
Independent benchmarks narrow the field to a shortlist. Then an evaluation harness tests the shortlist on your own cases, so quality is measured rather than assumed. Price and speed set the tier: the strongest model where the stakes are high, faster and cheaper models where they are enough.
What happens when a better model ships?
We re-run the evaluation harness and switch when the numbers say so. Integrations go through a model-agnostic layer and MCP servers that connect your systems once, so a switch is a configuration change, not a rebuild.
Where does the data on this page come from?
From Artificial Analysis, an independent firm that benchmarks AI models, using its website for language-model measurements and its data API for media arenas. We refresh it twice a day and draw it in our own charts. The scores and prices are theirs; the selection and the commentary are ours.
What is the Intelligence Index?
The Artificial Analysis Intelligence Index combines a set of independent evaluations, covering reasoning, knowledge, coding, and agentic work, into one score. Where a model runs at several reasoning settings, this page shows its best-scoring setting.
Next
Tell us the job. We will show you which model fits it, and prove it on your cases.