Model Quality and Precision
Compare models and understand quality signals
Every model page on Token Hub carries independent evaluation signals so you can choose with data instead of vibes: the Artificial Analysis Intelligence Index, a public benchmark suite, and per-model speed and latency medians.
Signals on each model page
| Signal | What it tells you |
|---|---|
| Intelligence rank | The model's position across the tracked frontier, from the AA Intelligence Index (0–100 composite). |
| Benchmarks | Task-level scores: GPQA (science), HLE, MMMU-Pro (multimodal), LiveCodeBench (code), AIME (math), and more. |
| Output speed | Median output tokens/second — predicts end-to-end latency for long generations. |
| Time to first token | Median TTFT — dominates perceived latency for short answers and chat. |
Comparing models programmatically
bashcurl -s https://tokenhub.coality.io/api/public/models/:id | jq '{id, intelligence_index, aa_rank, benchmarks, perf}'
Quality vs. price
The pricing table on each page includes a blended 3:1 rate (three input tokens to one output token) for quick cost comparisons. In practice, the best value is usually one tier below the absolute frontier — check whether a flash-tier variant of the same family meets your quality bar at a fraction of the price. Benchmark bars make that trade-off visible at a glance.
Precision and quantization
Token Hub is a routing gateway: model weights, precision (bf16/fp8), and serving stacks are managed upstream and are not user-configurable. What you can rely on is that every request to the same model id behaves consistently regardless of which healthy route served it — the gateway treats route-level variance as a reliability concern, not yours.