Artificial Analysis Llama 3.3 Instruct 70B provider comparison, long-prompt medians, fetched 2026-10-17

Fetched 2026-10-17. The page renders two different precisions of the same measurement and they do not agree, so this source records the per-provider JSON embedded in the page (the byPromptType.long block: long-prompt median output speed and median time to first chunk) rather than either rendered rounding. Long-prompt medians, output tok/s and first-chunk seconds: Groq 305.75 / 0.997, SambaNova 286.27 / 2.055, Google Vertex 153.59 / 0.678, Azure 130.91 / 2.278, CoreWeave 92.83 / 0.821, Scaleway 82.58 / 1.310, Together AI 80.37 / 1.532, Parasail (FP8) 77.16 / 2.482, Novita 45.34 / 1.769, DeepInfra (Turbo, FP8) 14.89 / 2.146. Amazon Bedrock is in the provider list with price $0.72/$0.72 but has no performance block. Against the 2026-10-16 capture of this same page, three long-prompt medians moved by more than rounding: SambaNova 288.5 to 286.27 tok/s (TTFT unchanged at 2.06s), Google Vertex 149.9 to 153.59 tok/s (TTFT unchanged at 0.68s), DeepInfra 15.2 to 14.89 tok/s (TTFT unchanged at 2.15s). Groq 305.7/1.00, Azure 132.2/2.28, Novita 45/1.77, Together 79/1.50 and Scaleway 83/1.33 are unchanged once the rendered rounding is undone, so those records are not touched. The page's own summary prose disagrees with its comparison table: the summary says SambaNova 286.3, Vertex 153.6 and DeepInfra 14.9, while the table still prints 288, 150 and 15. The embedded JSON is the finer-grained source and is what the numbers above come from. Split prices in the same JSON, unchanged from 2026-10-16: DeepInfra $0.10/$0.32, Novita $0.135/$0.40, Groq $0.59/$0.79, SambaNova $0.60/$1.20, Vertex $0.72/$0.72, Together $1.04/$1.04, Scaleway $1.05/$1.05.