AI IQ

Home About

How smart is your AI model, really?

AI IQ intelligently estimates the IQ's of popular AI models

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 pull in the performance ratings of models on objective industry-standard benchmarks
  2. We map each benchmark score to an implied IQ based on informed assessments of its difficulty
  3. We combine implied IQ scores into a single estimated IQ for each model
AI Models by EQ
AI Models by EQ
Each model's estimated EQ plotted on a standard normal IQ distribution

How AI EQ estimates emotional intelligence

  1. We pull in each model's EQ-Bench 3 Elo score and Arena Elo score
  2. We map each Elo score to an estimated EQ using calibrated piecewise-linear scales
  3. Anthropic models receive a 200-point Elo penalty on EQ-Bench to correct for family bias (judged by Claude)
  4. The composite EQ weights Arena 50% and EQ-Bench 50% (or uses one if only one is available)
Effective Cost
IQ vs Effective Cost
Each model's estimated IQ plotted against its token-efficiency-adjusted cost per 1M output tokens

Effective cost & isoquant curves

Effective cost adjusts each model's raw price by its token efficiency. Some models use fewer tokens to achieve the same result. We measure equivalent-token usage across benchmarks (ARC, SWE-bench, HLE, and others), then compute a geometric-mean multiplier relative to the median. A model that consistently uses fewer tokens gets a lower effective cost.

Isoquant curves are lines of equal preference. Models on the same curve give the same tradeoff between the selected metric and cost. Use the dropdown to switch between IQ, the four dimension IQs, or any of the 10 individual benchmarks. Use the slider to set your preference: slide left if the metric matters more, right if cost matters more. The curves reshape to reflect your tradeoff — models above and to the right of a curve are strictly better.

IQ Cost ?x:y = x cost halvings are required to justify a drop of y IQ points
Response Time
IQ vs Response Time
Each model's estimated IQ plotted against its response time in seconds

Speed & isoquant curves

Response time measures how long each model takes to return a complete answer, in seconds. Faster models appear further left on the chart.

Isoquant curves show lines of equal preference between the selected metric and speed. Use the dropdown to switch between IQ, the four dimension IQs, or any of the 10 individual benchmarks. Use the slider to set your tradeoff: slide left if the metric matters more, right if speed matters more. Models above and to the left of a curve are strictly better at that preference level.

IQ Speed ?x:y = x response time halvings are required to justify a drop of y IQ points
IQ Methodology
Each benchmark's hand-calibrated mapping from raw score to implied IQ
$$\text{IQ} = \tfrac{1}{4}\!\left(\text{IQ}_{\text{Abstract}} + \text{IQ}_{\text{Math}} + \text{IQ}_{\text{Prog}} + \text{IQ}_{\text{Acad}}\right)$$

10 benchmarks, 4 dimensions

Each benchmark score is mapped to an implied IQ via a hand-calibrated anchor curve. Benchmarks are grouped into four dimensions; each dimension IQ is the average of its benchmarks. The composite IQ is the mean of all four dimensions.

Hard, ungameable benchmarks retain high IQ ceilings. Easier or data-contaminated benchmarks have compressed ceilings to limit their influence.

Each model's composite IQ plotted against each dimension IQ, with linear fit and R²
Abstract Reasoning (D1)
Mathematical Reasoning (D2)
Programmatic Reasoning (D3)
Academic Reasoning (D4)

How dimensions relate to composite IQ

Composite IQ is the simple average of four dimension IQs: Abstract Reasoning, Mathematical Reasoning, Programmatic Reasoning, and Academic Reasoning.

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.

EQ Methodology
Each benchmark's hand-calibrated mapping from Elo score to implied EQ
$$\text{EQ} = 0.5 \cdot \text{EQ}_{\text{EQ-Bench}} + 0.5 \cdot \text{EQ}_{\text{Arena}}$$

2 benchmarks, 1 composite

Each Elo score is mapped to an implied EQ via a hand-calibrated anchor curve. EQ-Bench 3 measures emotional understanding in challenging roleplays. Arena Elo reflects broad conversational ability as judged by human preference.

When both scores are available, EQ is an equal-weighted composite: 50% Arena (human-judged) and 50% EQ-Bench (AI-judged). When only one is available, that value is used directly.

Anthropic family-bias adjustment

EQ-Bench 3 is judged by Claude (an Anthropic model), which creates potential scoring bias in favor of Anthropic models. To correct for this, we subtract a 200-point Elo penalty from the EQ-Bench component for all Anthropic models before mapping to implied EQ. The Arena component is unaffected since it uses human judges.

Each model's composite EQ plotted against each benchmark's implied EQ, with linear fit and R²
EQ-Bench 3
Arena Elo

How benchmarks relate to composite EQ

Composite EQ is the equal-weighted average of two benchmark EQs: EQ-Bench 3 (AI-judged emotional understanding) and Arena Elo (human-preference ranking).

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