How scores are built
This site does not average every published number into one ranking. Quality Score, Value Score, Reliability, and Evidence are separate scores with their own inputs. A missing source is omitted from that score. It is not treated as zero.
Quality Score v2
Quality Score is this site's answer-quality ranking. It requires Correct and Incorrect shares from AA-Omniscience. Abstains are neither right nor wrong. They sit outside both counts.
Pre-normalization weights are 0.50 answerQuality, 0.20 Intelligence, 0.075 Coding, and 0.075 Agentic. If a component is missing, its weight drops from both the numerator and the denominator. With all four present, the 0.50 answerQuality weight is 0.50 / 0.85, or 58.8% of the total.
Artificial Analysis Intelligence, Coding, and Agentic scores overlap. This site does not apply a source-overlap correction, so Quality Score is not an independent-evidence average. See the Artificial Analysis methodology for its benchmark aggregation.
precision = Correct / (Correct + Incorrect); excessWrong = max(0, Incorrect − Correct); answerQuality = clamp((0.6 × Correct + 0.4 × precision) × (1 − excessWrong × 0.75)); intelligence = 0.85 × clamp(AA Intelligence / 65) + 0.15 × clamp(Epoch general-capability percentage) when safely matched, otherwise AA Intelligence only; Quality Score = (Σ available weight × component / Σ available weight) × 100, with weights 0.50 / 0.20 / 0.075 / 0.075 on answerQuality, intelligence, Coding/80, Agentic/60; missing index drops from both sumsIntelligence component
The Intelligence component is 85% normalized AA Intelligence plus 15% Epoch General ECI, but only when Epoch uniquely matches the normalized family or name and the release date. Tied ECI values share a midpoint rank. If Epoch is missing, ambiguous, or mismatched, the component uses AA Intelligence alone.
With all Quality components present, Epoch's maximum effective contribution to Quality Score is about 3.5% (15% of the 0.20 Intelligence weight after renormalization). This site consumes published General ECI. It does not re-aggregate Epoch's raw benchmarks.
Epoch AI
Epoch estimates a general capability score from 50 or more benchmarks using an IRT-style logistic fit of capability, difficulty, and slope. Eligible settings for the same name and same release date are aggregated by taking the best score on each benchmark. A model needs at least four benchmark scores to enter that estimate.
Value Score v1
Value Score estimates capability per token dollar. It requires Intelligence, Correct, and blendedPrice. It does not use Epoch or Evidence. Changing the graph X-axis to Cost per task does not change Value Score. Missing Incorrect is treated as 0.
capability = clamp(Intelligence / 65); reliability = clamp(Correct × (1 − Incorrect × 0.35)) with missing Incorrect = 0; affordability = 1 / (1 + log10(1 + max(0, blendedPrice) × 8) × 0.45); Value Score = (0.6 × capability + 0.4 × reliability) × affordability × 100Reliability v1
Reliability is correct minus incorrect on a 0-100 scale. It requires Correct and Incorrect. Abstain is not an input. The same ranking idea as the AA-Omniscience Index, on a different scale.
Reliability = clamp((correct − incorrect + 1) / 2, 0, 1) × 100; requires finite correct and incorrect; abstain is not an inputEvidence Score
Evidence is a separate optional cross-source score. It does not contribute to Quality Score or Value Score. Matches are exact only. Each board is converted to a percentage. Software family averages the observed DeepSWE, FrontierCode, CursorBench, and Terminal-Bench percentages. General uses the Artificial Analysis percentage. Truthfulness uses BullshitBench. Preference uses LMArena WebDev.
The criteria for it to appear is at least four exact sources across at least three families.
The score is the average of the available family values. A missing family is omitted, not zero.
Blended price
Blended price is USD per 1M tokens using a 7:2:1 mix of cache-hit, input, and output where the source or component prices permit. The feed prefers Artificial Analysis's blended field, computes the same mix when components exist, and may retain a legacy 3:1 source fallback. Lower is better.