AI IQ

Subscribe

The AI Intelligence Leaderboard

Estimating the intelligence of every major AI model

AI Models by IQ
AI Models by IQ
Each model's estimated IQ plotted on a standard normal IQ distribution

How AI IQ estimates model intelligence

  1. We archive source captures from public benchmark leaderboards and extract only source-backed values
  2. We map each benchmark score to an implied IQ using calibrated difficulty curves
  3. We group scored benchmarks into seven dimensions: abstract, mathematical, scientific, app building, production engineering, computer use, and reliability
  4. We conservatively fill missing benchmark and dimension estimates only inside the scoring pipeline
  5. Every derived IQ averages all seven scored dimensions, so missing coverage cannot make a model look better by omission
IQ vs Effective Cost
IQ vs Effective Cost
Each model's estimated IQ plotted against effective cost per 1M I/O Tokens (sticker price × blended usage multiplier).
IQ 1:1 Cost

Effective cost & iso-curves

Effective cost on the X-axis is sticker price for 1M I/O Tokens × token usage multiplier. 1M I/O Tokens means 1M input tokens plus 1M output tokens, priced at the model's published rates.

Iso-curves trace lines of equal preference for IQ versus cost. The slider weights quality vs cost: center is 1:1, drag toward Cost to make cost matter more, or toward IQ to make quality matter more. Models above and to the right of a curve are strictly better.

Frontier IQ Over Time
Frontier IQ Over Time
X = release date. Y = estimated IQ. Provider step-lines connect each provider's flagship frontier checkpoints over time.

Tracking frontier progress

Each dot is a model with a known release date and a derived IQ estimate. Models are positioned left-to-right by release date, so the chart shows how the frontier changes over time rather than just where models rank today.

Provider-colored lines connect each lab's flagship frontier checkpoints. Codex, mini, nano, flash, coder, and smaller open-weight variants are omitted so the chart tracks each lab's main offering rather than every SKU.

This view is most useful for spotting whether a new release is actually ahead of its direct predecessor, or whether source coverage and conservative imputations are shaping the comparison.

Mathematical Reasoning IQ
Mathematical Reasoning IQ
Each model's Mathematical Reasoning IQ plotted on a standard normal IQ distribution

What it measures

Multi-step quantitative reasoning, from competition problems to research-level proofs.

FrontierMath Tier 4 FrontierMath Tier 1-3 AIME ProofBench MathArena
Scientific Reasoning IQ
Scientific Reasoning IQ
Each model's Scientific Reasoning IQ plotted on a standard normal IQ distribution

What it measures

Graduate-level reasoning across the natural sciences and applying scientific knowledge to hard problems.

Humanity's Last Exam CritPt SciCode GPQA Diamond
Abstract Reasoning IQ
Abstract Reasoning IQ
Each model's Abstract Reasoning IQ plotted on a standard normal IQ distribution

What it measures

Fluid problem-solving on novel puzzles a model cannot have memorized — abstracting patterns from just a few examples.

ARC-AGI-2 ARC-AGI-1 ARC-AGI-3
App Building IQ
App Building IQ
Each model's App Building IQ plotted on a standard normal IQ distribution

What it measures

Turning product and design prompts into usable apps, front-end experiences, and full-stack prototypes.

Arena.ai WebDev DesignArena Frontend DesignArena Full Stack Vibe Code Bench
Production Engineering IQ
Production Engineering IQ
Each model's Production Engineering IQ plotted on a standard normal IQ distribution

What it measures

Coding fluency, repository repair, debugging, testing, and long-horizon engineering execution.

SWE Marathon FrontierCode Diamond SWE-Bench Verified SWE-Bench Pro DeepSWE SWE-rebench LiveCodeBench
Computer Use IQ
Computer Use IQ
Each model's Computer Use IQ plotted on a standard normal IQ distribution

What it measures

Agentic operation of real tools and environments — terminals, browsers, and desktop apps.

Terminal-Bench 2.0 Terminal-Bench Hard BrowseComp OSWorld-Verified Toolathlon MCP Atlas
Reliability IQ
Reliability IQ
Each model's Reliability IQ plotted on a standard normal IQ distribution

What it measures

Following instructions precisely and knowing the limits of its own knowledge instead of guessing.

IFBench AA Omniscience
Emotional Reasoning (EQ)
Emotional Reasoning (EQ)
Diagnostic Emotional Reasoning scores, excluded from Composite IQ

What it measures

A diagnostic view of emotional and interpersonal behavior. This is excluded from Composite IQ until the benchmark base becomes more rigorous.

EQ-Bench 3 Arena.ai Overall AttuneBench
IQ vs Speed vs Cost in 3D
IQ vs Speed vs Cost in 3D
3D scatter: X = response time (log, faster to the right), Y = IQ, Z = effective cost (log). Color = provider. Drag to rotate.

Three tradeoffs at once

Most charts pit two qualities against each other. This view holds all three of the practical tradeoffs in one space: how smart a model is, how fast it answers, and what it costs to run.

IQ rises on the vertical axis, faster models sit to the right, and effective cost runs back into the depth axis on a log scale. The ideal model lives up, right, and toward the front — high intelligence, quick responses, and low cost. Drag to rotate and find where each provider clusters.

IQ Methodology
Each benchmark's IQ 70-160 expected-score ladder from raw score to implied IQ
$$\text{IQ} = \tfrac{1}{7}\!\left(\text{IQ}_{\text{Math}} + \text{IQ}_{\text{Science}} + \text{IQ}_{\text{Abstract}} + \text{IQ}_{\text{AppBuild}} + \text{IQ}_{\text{ProdEng}} + \text{IQ}_{\text{Computer}} + \text{IQ}_{\text{Reliability}}\right)$$

Scored benchmarks, 7 dimensions

Each benchmark score is mapped to an implied IQ through an expected-score ladder at IQ 70, 85, 100, 115, 130, 145, and 160. Benchmarks are grouped into seven scored dimensions: Mathematical Reasoning, Scientific Reasoning, Abstract Reasoning, App Building, Production Engineering, Computer Use, and Reliability. Each dimension IQ is the average of its benchmarks.

Hard, low-gameability benchmarks can assign high IQ to modest raw scores. Easier, saturated, or contamination-sensitive benchmarks require very high scores before reaching the 145+ range.

FrontierMath Tier 4, FrontierMath Tier 1-3, ProofBench, MathArena, and AIME are included in the Mathematical Reasoning dimension and shown as standalone benchmark charts on the IQ page.

Models need at least 2 of 7 dimensions covered to receive a derived IQ. Missing benchmarks and dimensions are conservatively imputed only inside the scoring pipeline, while raw benchmark charts keep source-backed values. Every derived composite uses all seven scored dimensions.

Direct predecessor lineage is used first when it is explicit. Remaining missing dimensions use a matched lower-quartile cap based on models with similar capability across the other dimensions.

Each model's composite IQ plotted against each dimension IQ, with linear fit and R²
Mathematical Reasoning (D1)
Scientific Reasoning (D2)
Abstract Reasoning (D3)
App Building (D4)
Production Engineering (D5)
Computer Use (D6)
Reliability (D7)

How dimensions relate to composite IQ

Composite IQ is the equal-weight average of seven dimension IQs: Mathematical Reasoning, Scientific Reasoning, Abstract Reasoning, App Building, Production Engineering, Computer Use, and Reliability. Derived scores always use all seven scored dimensions, with missing dimensions conservatively filled before averaging.

R² measures how much of the variance in composite IQ is explained by each dimension. A higher R² means that dimension is a stronger predictor of overall model intelligence.