Side-by-side technical and economic comparison between Alibaba / Qwen's Qwen 2.5 72B Instruct — Together AI and Mistral AI's Codestral (Current Generation).
• Qwen 2.5 72B Instruct — Together AI: Classified as a Legacy Reference model (Deprecated on Together AI (Reference)).
| Metric | Qwen 2.5 72B Instruct — Together AI | Codestral (Current Generation) | Advantage |
|---|---|---|---|
| Provider Organization | Alibaba / Qwen | Mistral AI | — |
| Standard Input / 1M | $0.35 | $0.30 | Codestral (Current Generation) (14% lower) |
| Cached Input / 1M | $0.175 | $0.03 | Codestral (Current Generation) lower |
| Output / 1M | $0.40 | $0.90 | Qwen 2.5 72B Instruct — Together AI lower |
| Context Window | 128k | 256k | Codestral (Current Generation) (256k) |
| Max Generation Tokens | 8.2k | 8.2k | Qwen 2.5 72B Instruct — Together AI |
| Tokenizer Family | Qwen Byte-level BPE (~152k vocabulary) | Mistral Tekken BPE (~131k vocabulary) | — |
Opt for Qwen 2.5 72B Instruct — Together AI when your engineering requirements prioritize Alibaba / Qwen's ecosystem, specific tokenizer efficiencies (Calibrated Qwen Tokenizer (±3%)), or when your expected prompt-to-completion ratios favor its $0.35/M input rate.
Opt for Codestral (Current Generation) when looking for Mistral AI's tooling integration, specific context window depth (256k tokens), or when output generation volume favors its $0.9/M rate.