There are 480+ language models tracked on ArtificialAnalysis.ai right now. Each one claims to be the best, fastest, or most affordable. Most of that is marketing. What you need is data.
ArtificialAnalysis.ai is one of the few platforms that evaluates AI models independently. No vendor pays to appear on their leaderboards. They run the tests themselves, using their own methodology, and publish the results for everyone. That independence is what makes the data worth trusting.
This article walks you through what the data actually shows, and gives you a framework for picking the right model for your specific task.
1. What ArtificialAnalysis.ai Does, and Why It Matters
Background: Most AI benchmarks are published by the companies that build the models. That creates an obvious conflict of interest. ArtificialAnalysis.ai re-runs evaluations independently, using standardized tests, so you can compare models across providers on equal terms.
The platform tracks three core dimensions for every model:
- Intelligence: how well the model performs across diverse reasoning, knowledge, and coding tasks
- Speed: output tokens per second, which determines how fast responses appear
- Price: USD per one million tokens, which determines what it costs to run at scale
It also maintains separate leaderboards for image and video generation, which operate on completely different criteria from text intelligence.
The composite intelligence score is called the Artificial Analysis Intelligence Index v4.0. It combines ten independent sub-evaluations into a single number. That number is useful for quick comparisons. The sub-benchmark breakdowns are useful for task-specific decisions.
2. The Six Benchmarks That Predict Real Performance
The image above lists the six benchmark categories used to evaluate frontier models, along with what each one tests and why it is harder than standard benchmarks.
Most AI benchmarks are too easy now. Frontier models score near-perfect on them, which makes it impossible to differentiate between the top options. ArtificialAnalysis.ai focuses on six that still produce meaningful separation.
GPQA: PhD-Level Science Knowledge
GPQA contains 448 expert-level science questions across biology, chemistry, and physics. Non-PhD humans score only 34% on this test, even with full internet access. That benchmark tells you something important: a model scoring well on GPQA has internalized knowledge at a depth that goes beyond what most humans retrieve through search.
What this predicts in practice: the model's usefulness for research assistance, scientific writing, and technical analysis in specialized domains.
Example: A biotech team using AI for drug interaction literature review needs strong GPQA performance. A model scoring 60%+ will give substantially more accurate responses than one scoring 40%, not marginally better ones.
MMLU-Pro: Language Comprehension Under Pressure
MMLU-Pro is a harder version of the Massive Multitask Language Understanding benchmark. The original gave four answer choices. This version gives ten. More choices reduce lucky guessing and produce a cleaner signal of actual comprehension.
Background: MMLU was one of the first large-scale tests used to evaluate language models across academic subjects. The Pro version removes easier questions and expands choices to make the test more discriminating.
Example: If you are deploying a model for customer support in legal or financial services, MMLU-Pro scores are a strong indicator of whether the model will handle ambiguous, nuanced language correctly.
AIME: Multi-Step Mathematics
AIME stands for the American Invitational Mathematics Examination, an invite-only national competition for top high school students. The problems require multi-step logical reasoning, symbolic manipulation, and the ability to hold a complex problem state across many steps.
Warning: Strong AIME scores do not guarantee accuracy on all math tasks. Models that score well here sometimes still make arithmetic errors in basic financial calculations. Always test on your specific math use case before committing.
Example: Quantitative finance teams evaluating models for strategy analysis should weight AIME scores heavily. A model that fails at this level will struggle with multi-step financial modeling chains.
LiveCodeBench: Real Coding Ability
LiveCodeBench pulls problems from ongoing competitive programming contests on LeetCode, AtCoder, and Codeforces. Because the problems come from live contests, they are unlikely to appear in any model's training data. The model has to actually solve them.
Background: "Data contamination" is a known issue in AI benchmarking. If a model was trained on the answers to benchmark questions, it scores high without actually learning anything new. Live benchmarks reduce this risk significantly.
Example: A software engineering team choosing a code assistant should prioritize LiveCodeBench scores over general intelligence scores. The correlation to production code quality is more direct.
MuSR: Sustained Logical Reasoning
MuSR tests long-form logical deduction. A typical problem involves reading a 1,000-word narrative and answering who has means, motive, and opportunity. It measures whether a model tracks multiple facts, relationships, and constraints across a long context without losing thread.
Example: Legal document analysis, contract review, and compliance checking all require this. A model that loses track of earlier clauses in a 40-page contract will produce unreliable summaries, even if its general intelligence score looks strong.
HLE: Humanity's Last Exam
HLE contains 2,500 of the hardest, most subject-diverse, multi-modal questions assembled for AI evaluation. It is designed to be the final academic test before AI performance exceeds what humans reliably achieve.
Warning: HLE scores are low even for the best models. Do not penalize a model for a low absolute score. Look at relative performance between models, not absolute numbers.
Example: Research institutions working on frontier science questions should monitor HLE scores closely. This benchmark is the best current proxy for whether a model can contribute to genuinely novel work.
3. The Intelligence Leaderboard: Who Leads and by How Much
The image above shows the top models ranked across three separate dimensions. Notice that the ranking order changes substantially depending on which dimension you are looking at.
[IMAGE PLACEHOLDER: Image 3, the full Artificial Analysis Intelligence Index bar chart with 28 models]
This chart shows 28 of the 480 tracked models, ranked by composite Intelligence Index score. The top three models, from three different companies, are tied at 57.
Current top intelligence rankings as of April 2026:
| Rank | Model | Score | Provider |
|---|---|---|---|
| 1 (tied) | Claude Opus 4.7 (max) | 57 | Anthropic |
| 1 (tied) | Gemini 3.1 Pro Preview | 57 | |
| 1 (tied) | GPT-5.4 (xhigh) | 57 | OpenAI |
| 4 | Kimi K2.6 | 54 | Kimi |
| 5 | Claude Opus 4.6 (max) | 53 | Anthropic |
| 6 | Muse Spark | 52 | Meta |
| 7 (tied) | Qwen3.6 Max Preview | 52 | Alibaba |
| 7 (tied) | Claude Sonnet 4.6 (max) | 52 | Anthropic |
| 9 | GLM-5.1 | 51 | Zhipu |
The three-way tie at the top is significant. Anthropic, Google, and OpenAI are operating at the same frontier capability level. No single provider has a clear intelligence advantage right now.
Where this gets more interesting is at the sub-benchmark level. A model ranked 4th overall might outperform the top three on a specific task category like coding or long-context retrieval. The composite score is a useful filter; the sub-scores are where you make the actual decision.
4. Intelligence vs. Cost: Finding Your Operating Point
The image above maps Intelligence Index score on the vertical axis against Cost to Run on the horizontal axis, displayed on a log scale in USD. The green-shaded area in the top-left is labeled "Most Attractive Quadrant," representing models that score high on intelligence while remaining affordable.
This chart is the most actionable view on ArtificialAnalysis.ai. It answers a specific question: are you paying more than you need to for the intelligence level your task actually requires?
How to read the four quadrants:
- Top-left (green): High intelligence, low cost. Use here when you can.
- Top-right: High intelligence, high cost. Justified only when accuracy is mission-critical.
- Bottom-left: Low intelligence, low cost. Good for simple, high-volume, automated tasks.
- Bottom-right: Low intelligence, high cost. Avoid.
What the data shows for specific models:
Gemini 3.1 Pro Preview scores 57 (tied for first) at a moderate cost per token, placing it near the green zone among frontier models. DeepSeek V3.2 scores around 41 at very low cost, making a strong case for cost-sensitive deployments where you do not need frontier accuracy. Claude Opus 4.7 and GPT-5.4 score at the top but sit far to the right of the cost axis. Those models are best reserved for tasks where getting the answer right is non-negotiable.
Practical decision rule: if a human reviews every AI output (legal drafting, medical notes, financial analysis), use a top-right model. If the task is automated and high-volume (content tagging, email routing, classification), use the green zone.
Pros and cons of top models across all three dimensions:
| Model | Intelligence | Speed (tok/s) | Price ($/1M tok) | Best for | Avoid for |
|---|---|---|---|---|---|
| Claude Opus 4.7 | 57 | 32 | $10 | Complex reasoning, research | High-volume automation |
| Gemini 3.1 Pro Preview | 57 | 185 | $1.7 | Balanced performance and speed | Ultra-low budget |
| GPT-5.4 (xhigh) | 57 | 43 | $4.5 | Coding, tool use | Budget-constrained |
| DeepSeek V3.2 | 41 | n/a | $0.4 | Cost-sensitive deployments | Frontier-accuracy tasks |
| Gemini 3 Flash | 45 | 160 | $0.3 | Speed at low cost | Deep reasoning tasks |
| Claude Haiku 4.5 | 36 | n/a | ~$0.25 | Real-time lightweight tasks | Scientific or academic work |
5. Speed: When It Changes the Product
The speed chart shows output tokens per second across leading models. gpt-oss-120B leads at 217 tokens per second. Grok 4.20 follows at 185. Gemini 3 Flash sits at 160. Claude Opus 4.7 generates 32 tokens per second, which is adequate for interactive use but not for real-time streaming at scale.
Speed matters in specific situations:
- Real-time chat interfaces: users notice latency above roughly one second. At 32 tokens per second, a 500-token response takes about 15 seconds.
- Streaming data pipelines: workflows that feed model output into downstream systems need throughput, not accuracy alone.
- Voice AI: text-to-speech pipelines need token generation to outpace speech synthesis, typically requiring 100 or more tokens per second.
Example: A customer support chatbot handling 10,000 conversations per day with Claude Opus 4.7 (32 tok/s) vs Gemini 3.1 Pro Preview (185 tok/s) would see a 5.8x difference in throughput capacity. That means roughly 6x more compute infrastructure for the same load with the slower model.
Counter-view worth noting: for batch processing tasks such as overnight report generation or document indexing, speed is nearly irrelevant. Choosing a faster, more expensive model for those use cases adds cost without adding value.
Warning: Speed benchmarks are measured under standard conditions. Real-world throughput varies with prompt length, provider infrastructure load, and response length. Test under your actual usage pattern before making infrastructure decisions.
6. Price: What the Range Actually Means
Background: LLM APIs charge per token, roughly 0.75 words per token. Prices are quoted per one million tokens, which equals approximately 750,000 words or around 1,500 pages of text. Input tokens (your prompt) and output tokens (the model's response) are often priced separately.
The price range across leading models spans two orders of magnitude:
- Cheapest: Gemini 3 Flash, gpt-oss-120B, DeepSeek V3.2 at around $0.30 to $0.40 per million tokens.
- Most expensive: Claude Opus 4.7 (max) at $10 per million tokens, which is 33x more expensive than Gemini 3 Flash.
The price gap reflects model size, computational requirements, and market positioning. It is not arbitrary, but it is also not always justified for your use case.
The question is not what is cheapest. It is: what is the minimum intelligence level your task actually requires?
A framework for matching price to task:
- PhD-level domain expertise or multi-document synthesis: use top-tier models ($4 to $10 per million tokens)
- Code generation, complex analysis, long-form writing: use mid-tier models ($1 to $4 per million tokens)
- Summarization, classification, Q&A on known content: use budget-tier models ($0.30 to $1 per million tokens)
- Simple extraction, formatting, or routing: use the smallest model available
Example: A SaaS company processing 50 million tokens per day would pay $500 per day with Gemini 3 Flash vs $500,000 per day with Claude Opus 4.7. For content tagging, that $499,500 daily difference is not justified. For rare, high-stakes legal document review, the cost per decision might be entirely reasonable.
7. How AI Intelligence Has Grown Over Time
This chart tracks Intelligence Index scores for 15 leading model creators from November 2022 through May 2026. Every line moves upward. In November 2022, the best models scored around 9 to 13. By April 2026, the frontier sits at 57. That is roughly a 5x improvement in 3.5 years.
Key observations from the timeline:
- November 2022: OpenAI leads with scores around 9 to 13. All other providers cluster below 10.
- Late 2023: Acceleration begins. Google, Anthropic, and Meta start closing the gap.
- 2024 to 2025: Chinese labs including Alibaba (Qwen), Xiaomi, and DeepSeek emerge as credible competitors. The frontier cluster expands to five or six companies within a few points of each other.
- Early 2026: Anthropic, Google, and OpenAI all reach 57 and are statistically tied.
The practical implication: the model you choose today will likely be mid-tier within 12 months. If you build your system in a way that couples it tightly to a specific model, you will pay a higher upgrade cost later. Where possible, build model-agnostic systems.
Counter-view: rapid improvement also means your existing production system, even one built on a 2024 model, may still perform well for your specific task. Do not upgrade because newer models exist. Upgrade when your current model's limitations affect your outcomes in measurable ways.
8. Image Generation: A Separate Evaluation Entirely
The image above shows the Text-to-Image leaderboard, which uses ELO scores based on blind preference voting. GPT Image 1.5 leads at 1,273, followed by Google's Nano Banana 2 at 1,265 and Nano Banana Pro at 1,214.
Background: ELO scoring was originally designed for chess rankings. In this context, each model "wins" or "loses" based on human preference comparisons in blind side-by-side tests. A higher ELO means more wins against other models.
For image generation tasks, the language intelligence rankings above are irrelevant. These are fundamentally different model architectures.
Current top text-to-image rankings:
| Rank | Model | ELO Score | Provider |
|---|---|---|---|
| 1 | GPT Image 1.5 (high) | 1,273 | OpenAI |
| 2 | Nano Banana 2 (Gemini 3.1 Flash Image Preview) | 1,265 | |
| 3 | Nano Banana Pro (Gemini 3 Pro Image) | 1,214 | |
| 4 | FLUX.2 (max) | 1,205 | Black Forest Labs |
| 5 | Seedream 4.0 | 1,202 | ByteDance |
| 6 | grok-imagine-image | 1,184 | xAI |
Claude Opus 4.7 does not appear on this leaderboard at all. Strong language intelligence does not transfer to image quality.
Example: A marketing team using AI for visual content should look at GPT Image 1.5 or Google's Gemini image models, not at the text intelligence rankings.
Warning: ELO scores reflect general aesthetic preference in blind tests. For domain-specific image tasks such as product photography, medical imaging, or architectural visualization, run your own evaluation. General ELO rankings do not reliably predict domain-specific performance.
9. Intelligence Breakdown: Where the Real Selection Happens
This panel shows per-benchmark performance across all tracked models. The six sub-charts cover GDPval-AA, Terminal-Bench Hard, tau-squared Bench Telecom, AA-LCR, AA-Omniscience Accuracy, and AA-Omniscience Non-Hallucination Rate. Each chart shows a different ranking order, which confirms that no single model leads across every dimension.
The composite intelligence score hides important variation. Here is what each sub-benchmark tells you, and when to weight it:
| Sub-Benchmark | What It Tests | Weight This For |
|---|---|---|
| GDPval-AA | General deep reasoning (top score: 63%) | Research, analysis |
| Terminal-Bench Hard | Complex system and terminal tasks (top: 58%) | DevOps, SRE tooling |
| Tau-Bench Telecom | Telecom domain knowledge (top: 98%) | Telecom industry AI |
| AA-LCR | Long-context retrieval accuracy (top: 74%) | Document Q&A, RAG systems |
| AA-Omniscience Accuracy | Breadth of factual knowledge (top: 55%) | General knowledge bases |
| AA-Omniscience Non-Hallucination | Rate of refusing to fabricate (top: 83%) | Fact-sensitive customer-facing tasks |
Background: RAG stands for Retrieval-Augmented Generation. It is a technique where the model retrieves relevant documents before generating a response, used commonly in enterprise search and document Q&A products.
Example: A healthcare company building a medical information chatbot should weight the Non-Hallucination Rate above every other metric. A model that generates false medical information with confidence is worse than no model at all. The AA-Omniscience Non-Hallucination chart, where Grok 4.20 0309 v2 scores 83%, is directly relevant for that selection.
Counter-view: high non-hallucination rates sometimes correlate with more frequent "I don't know" responses. For internal R&D tools where missing information is a bigger problem than fabricating it, a slightly lower non-hallucination score with higher overall accuracy may be the right trade.
10. A Decision Framework for Picking Your Model
Bring together everything above into a repeatable process:
Step 1: Define your task type.
- Text generation or reasoning: go to Step 2.
- Image or video generation: use the Image Leaderboard. Start with GPT Image 1.5 or Gemini image models.
- Code generation: prioritize LiveCodeBench scores over composite intelligence scores.
Step 2: Identify your primary constraint.
- Accuracy is critical (medical, legal, research): look at models scoring 50 or above.
- Cost is the bottleneck (high-volume automated tasks): look at DeepSeek V3.2, Gemini 3 Flash, and similar budget-quadrant models.
- Speed is critical (real-time applications, voice AI): look at the Speed leaderboard. gpt-oss-120B and Grok 4.20 lead here.
Step 3: Use the Intelligence vs. Cost scatter plot.
Find models in or near the Most Attractive Quadrant that meet your minimum intelligence threshold.
Step 4: Check the sub-benchmarks relevant to your domain.
- Long documents: AA-LCR
- Factual accuracy in customer-facing contexts: Non-Hallucination Rate
- Scientific or technical depth: GDPval-AA and GPQA
- Coding: LiveCodeBench
- Math or multi-step reasoning: AIME and MuSR
Step 5: Run your own evaluation.
Test on 50 to 100 examples from your actual use case before committing. Benchmark scores are population-level averages. Your specific prompts, domain vocabulary, and output format requirements will produce results that differ from benchmark rankings.
Warning: Treat benchmarks as a shortlist filter, not a final answer. The gap between benchmark rank and performance on your specific task can be substantial.
The Practical Summary
The data from ArtificialAnalysis.ai makes several things clear.
The frontier is genuinely competitive. Claude Opus 4.7, Gemini 3.1 Pro Preview, and GPT-5.4 are all tied at 57. You are not leaving significant intelligence on the table by choosing any of them. Your decision should come down to cost, speed, and the specific sub-benchmarks that matter for your task.
The price range is enormous. Gemini 3 Flash costs $0.30 per million tokens. Claude Opus 4.7 costs $10. For most automated tasks, the cheaper model is the correct choice.
Image generation is a separate decision tree entirely. Do not use text intelligence rankings to choose an image model.
Model capability is improving fast. The best model today may be mid-tier in 12 months. Build systems that are easy to upgrade.
Benchmarks are filters, not answers. Use them to narrow your options, then test on your actual task before deciding.
Data sourced from ArtificialAnalysis.ai, an independent AI evaluation platform. Rankings reflect data as of April 2026.
Top comments (0)