Concepts
Data model & evidence
A few conventions run through every endpoint. Understand these once and the whole API reads clearly.
Entities
| Entity | What it is |
|---|---|
| Model | A canonical model identity (e.g. gpt-oss-120b), normalized across every provider's dialect. |
| Provider | Who actually serves inference (Fireworks, Together, a cloud). Ranked on the provider trust board. |
| Aggregator / Router | A marketplace that routes across providers (OpenRouter, HuggingFace, 0G, Novita, Requesty). Ranked on the aggregator board. |
| Offer | One (model × provider × router) price point, with prices, context window and an evidence tag. |
Evidence tags
Every offer says how we know its numbers. This is the backbone of TKX's credibility.
| Tag | Meaning |
|---|---|
| official | Vendor's published list price (rate card). |
| api | Read from the provider's live public API. |
| probe | Measured by a TKX probe (real TTFT / TPS / uptime). |
| self | Self-reported — always down-ranked, never treated as truth. |
Canonical model IDs
Providers name the same model a dozen ways. TKX normalizes them to one canonical id so a single model row aggregates every provider serving it. Use the canonical id as the path parameter for model detail.
Blended price
To rank a model by a single number we use a blended price weighted toward output, which dominates real cost:
blended = (3 × input_per_1M + output_per_1M) / 4Tiers: live vs listed
Aggregators carry a tier: live means TKX ingests their real data now; listed means tracked but not yet integrated (shown, never scored on fabricated numbers).
Provenance
i
Every collection run is auditable without database access at /data/provenance.json — per-source timestamps, HTTP status and payload SHA-256.