OpenRouter provider table for meta-llama/llama-3.3-70b-instruct, re-read 2026-10-16 (per-provider price, P50 latency, throughput, uptime)

Fetched 2026-10-16, same URL as svc-src-openrouter-provider-table-2026-10-14. Per-provider rows for Llama 3.3 70B Instruct, P50 over OpenRouter's own routed traffic: DeepInfra (Turbo) $0.100/$0.320, 0.48s latency, 16 tps, 98.80% uptime; NovitaAI $0.135/$0.400, 1.05s, 44 tps, 98.40%; AkashML $0.200/$0.520 (cache read $0.100), 0.49s, 38 tps, 99.87%; Parasail $0.220/$0.500 (cache read $0.110), 0.61s, 50 tps, 99.88%; Cloudflare $0.293/$2.253, 0.73s, 40 tps, 98.11%; SambaNova Turbo list $0.60/$1.20 with a 25% off badge rendering $0.450/$0.900, 0.81s, 104 tps, 99.83%; Groq $0.590/$0.790 (cache read $0.295), 0.52s, 193 tps, 100.00%; CoreWeave $0.710/$0.710 (cache read $0.710), 0.12s, 85 tps, 100.00%; Google Vertex $0.720/$0.720 with no latency or throughput cell; Together $1.040/$1.040, no latency cell, 95.86% uptime. Page headline price is the floor route $0.10/$0.32, not the Groq row. Best-across-providers throughput widget reads 193 tok/s (Groq); best latency widget reads 0.12s (CoreWeave). Compared with the 2026-10-14 snapshot the Groq row moved from 0.40s / 233 tps to 0.52s / 193 tps, DeepInfra from 0.51s / 14 tps to 0.48s / 16 tps, Novita from 0.58s / 41 tps to 1.05s / 44 tps, SambaNova from 0.77s / 98 tps to 0.81s / 104 tps. List prices on every row are unchanged. These are OpenRouter-measured P50 numbers over its own traffic, not vendor claims and not the Artificial Analysis medians already stored on the individual provider records.