Devstral 2 2512 vs Mistral Small 4

A token is the unit an AI model processes: usually between half a word and a full word. Rule of thumb: 1,000 tokens ≈ 750 English words. A 20-word sentence is about 26 tokens; a 300-word email is around 400. Models charge for input tokens (your prompt) and output tokens (their answer) separately.

When your app sends the same system prompt on every request (e.g. "You are a customer support agent for Acme…"), providers store it once and reuse the computation. They only charge full price for the new part (the user's question). OpenAI, Anthropic and Google give 50% to 90% off cached tokens. The calculator above accounts for this via a cache-hit % slider.

Chatbot: GPT-5.4 Mini or Claude Haiku 4.5 for latency; DeepSeek V3.2 for cost. Code: Claude Opus 4.6 and GPT-5.4 still lead the benchmarks, with o3 and DeepSeek R1 for hard reasoning. Document analysis: you need a large window — Gemini 3.1 Pro, Claude Opus or Llama 4 Maverick. Use the Model Finder above for a personalised pick.

Five levers: (1) cache system prompts (saves 50-90%), (2) batch API for non-urgent requests (50% off), (3) smaller model for simple tasks and flagship only when needed (complexity router), (4) trim prompts and cap output tokens, (5) fine-tune or use RAG instead of re-sending the same context every call. At Letbrand we audit each lever and typically cut costs by 40-80%.