Token cost

Every API provider prices in tokens, almost always quoted per million tokens (per-MTok). Input tokens (your prompt + context) are cheaper than output tokens (the model's response), usually by 3-5x. The asymmetry exists because output is the bottleneck — a model can ingest a million-token prompt in seconds but generate output one token at a time on expensive accelerator memory. Mid-2026 ballpark pricing: Anthropic Claude Opus 4.7 = ~$15/$75 per M input/output, Sonnet 4.6 = ~$3/$15, Haiku 4.x = ~$0.80/$4. OpenAI GPT-5 = ~$10/$30, GPT-5 mini = ~$0.40/$1.60. Google Gemini 3 Pro = ~$1.25/$5, Gemini 3 Flash = ~$0.10/$0.40. xAI Grok 4 = ~$5/$15. Open-weight via inference providers (DeepSeek V3 on Together, Llama 4 on Fireworks) often runs $0.20-$1 per M tokens combined, which is why cost-sensitive apps route to them. The hidden costs to watch: prompt caching can drop input cost by 90% on cache hits — Anthropic and OpenAI both support it, but you have to architect prompts to be cache-friendly. Reasoning models bill thinking tokens as output tokens, so an o-series or extended-thinking call that outputs 500 visible tokens may have generated 5,000 billable thinking tokens. Batch APIs are 50% off but async only. For agentic apps, the dominant cost is usually the long input from re-sending tool-call history every turn — RAG with a vector DB beats stuffing everything into context for cost reasons too.