If you're looking to optimize AI costs, understanding OpenRouter’s per-model pricing structure and the real differences in token costs is essential. OpenRouter doesn’t simply offer “cheap per 1M tokens”—its pricing varies significantly based on the model, provider, and whether tokens are input or output. Quick key points: 🧠 Why the pricing structure is complex Token-based billing: Input (prompt) and output (completion) tokens are charged at different rates. Wide price variation across models: Some models are free, and even identical models can have different rates depending on the provider. Pay-as-you-go: No minimum usage required—you only pay for what you use, lowering the barrier to entry. 💡 Representative examples deepseek/deepseek-r1 (based on a specific provider): Input: $0.70 per 1M tokens Output: $2.50 per 1M tokens → Output costs are significantly higher than input, so costs rise sharply as generation volume increases. 🆓 Free and open options OpenRouter offers free models, which may come with usage limits or priority restrictions—but still serve as practical free alternatives. The DeepSeek R1 family has been reported as accessible via a free tier, making it a powerful choice for testing or low-scale applications. 🧪 Practical tips Check per-model rates on OpenRouter’s pricing page: It displays real-time rates for each model/provider, making it the most accurate source for cost comparison. Choose models based on use case: For high-output tasks (e.g., long text generation), prioritize models with low output costs; for context-heavy tasks, favor models with high input efficiency. Consider prepay credits or volume discounts: For high token usage, prepay credits and volume discounts can significantly reduce overall costs. In short, OpenRouter isn’t a flat-rate system—it’s a layered, model- and provider-specific pricing marketplace. To maximize cost efficiency, make it a habit to directly compare rates across individual model and provider pages.