meta-llama/Llama-3.3-70B-Instruct-Turbo

By price, meta-llama/Llama-3.3-70B-Instruct-Turbo ranks 37th cheapest of 68 on input and 19th on output. Its quality score of 78 ranks 56th of 68, and on value (quality score ÷ output price) it ranks 20th of 68. Its blended rate sits 70% below the catalog average.

In real billing terms: a support chatbot handling 50,000 turns a month (2,000 input / 500 output tokens each) costs about $110 per month on meta-llama/Llama-3.3-70B-Instruct-Turbo, while a RAG workload of 20,000 queries (8,000 input / 800 output tokens each) runs about $155. Because output is priced 1.0× input here, long-generation tasks scale cost fastest — cap output length before you optimise anything else.

A 131K context window holds roughly 175 pages of A4 text (at ~750 tokens per page), ranking 47th of 68 here. That covers typical RAG and multi-turn chat, though book-length inputs still need retrieval and chunking. Output runs at about 110 tok/s (26th fastest of 68), acceptable interactively, though long responses will feel slow.

We track 2 Together AI models, and meta-llama/Llama-3.3-70B-Instruct-Turbo's blended rate sits 15% below that provider's average. If your task tolerates a little less headroom, qwen3.5-flash (Alibaba) blends about 94% cheaper at a quality score of 82, and is worth benchmarking first. Going the other way, Hunyuan HY3 scores 87 on quality for about 89% less.

The closest model in Together AI's line-up is Qwen/Qwen2.5-72B-Instruct-Turbo ($1.20 input / $1.20 output, quality 79). Against it, meta-llama/Llama-3.3-70B-Instruct-Turbo scores 1 point lower on quality, runs 10 tok/s faster (about 10%), costs 36% less blended. When two models sit this close on price, the deciding factor is usually the quality ceiling — run both against your real prompts before picking.