Over the past few months I have been trying to understand how to use LLMs cost-effectively on OpenRouter. AI bills can quickly become excessive. And with agentic coding, code quality can deteriorate as costs go sky-high.
Using spec-driven development—where one writes a spec, and then has an agent implement it in chat—can lower costs. In many of my experiments, costs were between 1 and 7 cents to evaluate a Markdown prompt (up to 180 lines) and generate a function. This is significantly less expensive than a coding harness like OpenCode—less context is needed. Besides generating code from specifications, using cheaper models (cost per million tokens) is beneficial. For example, GLM 5.3 Flash and MiMo-V2.6 Flash were much cheaper for code generation.
• GPT-6-Sol evaluated a 180-line prompt for 5 cents, while Kimi-K3 evaluated a shorter prompt for 32 cents—the difference is that GPT generated fewer thinking tokens.
• GLM 5.3 Flash and MiMo-V2.6 Flash had such low per-token costs that they could generate output for a tiny fraction of a cent.
• GPT-6-Sol, because it was more frugal with thinking tokens, was cheaper than Kimi-K3; the cheapest options were GLM 5.3 Flash and MiMo-V2.6 Flash.