Most leaderboards are benchmark runs. NemoRouter's rankings are aggregated from real customer traffic — what actually gets shipped, what costs what, what fails when. Live data, hourly refresh, anonymized at the source.
Sample preview
Across the NemoRouter network
Tracked for ranking eligibility
Volume, latency, cost, quality, trend, reliability
P50/P95/P99 windows refresh on the hour
Sample data only. Live rankings ship once the network has aggregated enough traffic to be statistically meaningful — currently 20+ models on Google Vertex AI, with Google (Gemini, Imagen, Veo) — Anthropic, OpenAI, and AWS Bedrock shipping next.
Rankings only matter when they reflect production reality. Each ranking below ships with the same methodology page — open dataset definitions, refresh cadence, and statistical caveats.
Top 20 models ranked by 30-day token throughput across the NemoRouter network. The volume tells you what teams are actually shipping with — not what they pinned for a benchmark run.
Sortable matrix combining speed (TTFT, tokens/sec), quality (independent eval scores), cost efficiency ($/1M tokens), and reliability (success rate, error class).
Treemap of provider token share, refreshed in real time. Watch which provider gets called most — and where customer traffic shifts after a price drop or a new model release.
Dollar-per-million-tokens compared across the catalog, normalized for context length and tool-use overhead. The ranking that should drive routing-strategy decisions.
7-day sparklines showing which models are gaining or losing share — the early-signal page for which provider to evaluate next.
P50/P95/P99 measurements per model and provider, refreshed hourly. These are observed numbers from production traffic — not vendor-reported claims.
The integrity of the ranking is the integrity of the data source. Here is exactly what we use.
All rankings are aggregated from real production calls flowing through NemoRouter. We never log prompt content for ranking purposes — only aggregate counts, latency, cost, and provider/model identity.
Cost data comes from the x-nemo-response-cost header on every call — provider-authoritative, never estimated. The same number that hits your credit ledger drives the ranking.
Quality scores supplement throughput rankings using public eval results (HumanEval, MMLU, GSM8K, MT-Bench). We disclose every source on the methodology page.
Launch list · ~weekly drops
One email when the data crosses the statistical-significance threshold. Then nothing until something interesting happens.