Artificial Analysis Llama 3.3 Instruct 70B provider comparison (speed, TTFT, price ranks), fetched 2026-10-16

Fetched 2026-10-16. Output speed medians (tok/s): Groq 305.7, SambaNova 288.5, Google Vertex 149.9, Azure 132.2, CoreWeave 92.9, Scaleway 83, Together AI 79, Parasail 77, Novita 45, DeepInfra (Turbo, FP8) 15.2. TTFT medians (s, first chunk): Google Vertex 0.68, CoreWeave 0.82, Groq 1.00, Scaleway 1.33, Together AI Turbo 1.50, Novita 1.77, SambaNova 2.06, DeepInfra 2.14, Azure 2.28. Amazon Bedrock is listed with context 128k but speed/latency cells blank. Blended price ranks (USD per 1M, not split in/out on the extracted table): DeepInfra (Turbo, FP8) 0.12, Novita 0.16, Parasail (FP8) 0.17, Scaleway 1.05 (up to 8.6x spread). Split prices stated in the FAQ: DeepInfra input $0.10 / output $0.32, Novita input $0.14 / output $0.40. Context windows on the comparison table: Groq/Novita/Parasail/Together/DeepInfra 131k, SambaNova/CoreWeave/Azure/Google/Bedrock 128k, Scaleway 100k. Price cells in the comparison grid itself rendered as '--' (client-side); the numeric price claims used here are the FAQ prose, not the blank grid.