Accessing the data

The data model behind the benchmarks and how it is structured so you can pull it.

Every number on this site comes from a single typed dataset with a stable shape. The same interface backs the site today and will back a public data feed: the UI reads through a DataAdapter, so the source can change without the shape changing.

Entities#

| Entity | Key fields | | --- | --- | | Model | id, developerId, family, params, isReasoning, isOpenWeights, contextWindow, modality, effortLevels | | Developer | id, name, color | | Provider | id, type (overshoot / hosted-api / self-host-vllm), regionIds | | Region | id, name, continent, lat, lng | | Benchmark | id, category, unit, higherIsBetter | | Score | modelId, benchmarkId, value, ciLow, ciHigh, sampleN | | LatencyPoint | modelId, providerId, regionId, ttftMs, e2eMs, realtimeReady | | SpeedPoint | modelId, providerId, tokensPerSec | | PricePoint | modelId, providerId, inputPer1M, outputPer1M, perFrame | | AgentTaskResult | modelId, taskSuite, passAt1, avgCostUsd, outTokens, steps |

The composite index#

The Intelligence Index is computed, not hand-set:

index = Σ (weightᵢ × scoreᵢ) / Σ weightᵢ

over the six capability benchmarks, with published weights. Recompute it from the raw Score rows and you'll reproduce the value we display, exactly.

Adapter interface#

interface DataAdapter {
  models(): Promise<Model[]>
  scores(q?: ScoreQuery): Promise<Score[]>
  latency(): Promise<LatencyPoint[]>
  // …one method per entity
}

Swap the implementation, keep every page. A public read API is on the roadmap. Same discipline as the product: one stable shape, results you can pull and verify yourself.