Inference Provider Metrics

Reading & caveats

Price (× floor)

Blended price at an assumed 10:1 input:output mix, divided per model by the cheapest endpoint's blend. 1.0× = that provider is the floor. Log scale, clamped at 8×.

Context fidelity

Endpoint context ÷ the largest context any provider serves for that model. Low values are cut-down deployments (8K–262K of a 1M-context model). A capacity signal that complements the measured speed index.

Reliability

1-day uptime per endpoint, clamped below at 80%. Medians in scatter panels; full endpoint spread in curves and ECDFs.

Speed index

Endpoint p50 throughput ÷ the fastest endpoint p50 for that model, so 100% = the fastest host of that model, comparable across models. OpenRouter-measured over a trailing 30-minute window sampled once daily at pull time — routed production traffic, not a clean-room benchmark. Raw tok/s and p90 latency appear in tooltips and the table. Provider medians are suppressed when fewer than 3 endpoints and under half the catalog carry throughput data; endpoints without throughput are omitted from speed panels only.

Curves

In endpoint-level panels each provider's endpoints are sorted by the x-metric and connected — a profile across their catalog, not a time series.

First parties

Google, Amazon Bedrock, Azure, Anthropic, Alibaba appear via their own or licensed models; their 1.0× price index reflects single-source pricing power, not cheapness.