Methodology

Intelligence Score

Each model's quality is a weighted blend of three Artificial Analysis indices, all on a 0–100 scale:

• 50% Agentic Index — multi-step autonomous task completion (the highest-value capability).
• 40% Coding Index — code generation, debugging and refactoring quality.
• 10% Speed Score — normalized output tokens/second; secondary to quality.

Intelligence = 0.50 × Agentic
           + 0.40 × Coding
           + 0.10 × Speed

AA Speed is not exported as a standalone column in the static dataset; it is folded into WMQ at the weight above. Where AA has no profile for a model, Intelligence Score cannot be computed and the Value per Intelligence per Task shows "—".

Efficiency multiplier

Two models can post the same quality score yet differ sharply in how much compute they burn to finish real work. Artificial Analysis publishes a cost-per-task figure — the dollar cost to run its standardized agentic task — and Code Smart folds it into the Value per Intelligence per Task as a bounded multiplier. Cheaper-per-task nudges value up; pricier nudges it down. It rides on top of quality and price rather than replacing either, so it can never dominate them.

reference   = median cost-per-task across models with data
eff         = clamp(0, 100, reference / costPerTask × 50)   // median model → 50
multiplier  = 0.85 + (eff / 100) × 0.30                     // bounded [0.85, 1.15]
Value per Intelligence per Task = Intelligence-adjusted capacity × multiplier / price → normalized 0–100

The reference is self-calibrating: the median model sits at a neutral 1.0×, cheaper models reach up to 1.15×, pricier ones down to 0.85×. When a model has no published cost-per-task — the current state for every model — the multiplier is exactly 1.0 and the Value per Intelligence per Task is unchanged. Cost-per-task and the active multiplier surface on each ranking card ("—" until a value is sourced).

Model Efficiency Index

A separate ranking that answers: which model delivers the most intelligence and throughput per dollar spent on API calls? Three inputs from Artificial Analysis and OpenRouter combine into a single composite.

$/Task =
  ( 7,000 × OR_In$/1M
  + 3,000 × OR_CH$/1M
  + Tok/Task × OR_Out$/1M
  ) ÷ 1,000,000

(Intel × t/s) ÷ $/100T

Doubling speed at constant cost
and intelligence doubles the score.
Doubling cost halves it.

Intel = AA Intelligence Index (0–100). t/s = AA median output tokens/sec. $/100T = cost to run 100 representative agentic tasks at OpenRouter pricing. Full model rankings →

Token estimation

Providers disclose limits in many units — messages, requests, credits, compute units, or raw tokens. The normalization engine converts each into per-window token estimates using an 8-layer priority dispatch (one layer per limit type); the first matching layer wins. These are estimates, not guarantees.

Per-window estimates (5h / 24h / 1w / 1mo) are extrapolated proportionally with confidence decay. A vague or undisclosed limit yields null — shown as "—", never a misleading 0.

Confidence levels

Every figure carries one of five confidence levels. Confidence is never upgraded — a derived value can be no more certain than its weakest input.

observed

Read directly from the provider's official page or API.

inferred

Mathematically derived from observed figures (e.g. annual ÷ 12).

assumed

A reasonable assumption, not yet verified from an official source.

stale

Was observed, but the source is now more than 90 days old.

unknown

Could not be determined; the value is null and displays as "—".