free model health board

live TPS, latency, and failure rate — the metrics the SI dashboard doesn't show.

The Surplus Intelligence marketplace exposes pricing, liquidity, and 24h volume for each model — but not TPS, wall-clock latency, or per-provider failure rates. So we measure them ourselves. Every request routed through surp feeds this board, and the resulting health scores feed back into routing: cheap-and-flaky loses to slightly-pricier-but-reliable.

composite health ranking

Models are ranked by a composite 0-100 score: reliability (60%) × speed (30%) × throughput (10%). Green ≥ 70, amber 40-69, red < 40.

methodology

• Window: rolling 1 hour (configurable via SURP_HEALTH_WINDOW).
• TPS: requests / elapsed window seconds.
• p50/p95: percentile of successful-request latencies (interpolated).
• Failure rate: non-ok responses / total responses.
• Health score: (1 - failure_rate) × 60 + speed_norm × 30 + min(1, TPS) × 10, where speed_norm = (2000 - p50) / 1950 clamped to [0,1].
• Pruning: samples older than 24h are deleted automatically.

free-to-paid conversion

How many free-tier users have upgraded to a paid combo, and which one they chose. This is the metric that tells us whether the free tier is doing its job — bootstrapping usage — without permanently subsidizing freeloaders.

top paid combos chosen by free users

what this enables

• Smarter free routing: the free pool can prefer reliable models, not just the absolute cheapest.
• Provider accountability: sellers who degrade get visible — not hidden behind a static offer count.
• Capacity planning: TPS trends reveal which providers are scaling and which are stagnating.
• Community good: this data doesn't exist anywhere else. We can publish it as a public good.

related: free models & live budget · system status · cache-aware routing · reward proposal