Spotify developed a method to sharply reduce Claude Code token usage, Techzine reported on Saturday.

The core idea is to hand off repetitive tasks such as file reading and predictable code generation to cheaper AI models through its in-house developer platform portal. Tests showed this cut Claude token consumption by an average of 90 percent, Techzine reported.

According to the report, the system Spotify developed centers on agents known as "AiKA mode." These agents run in a temporary runtime. Developers specify the model, instructions and MCP tools to use by mode. There is no need to manage separate infrastructure or API keys.

Spotify engineer Dmitry Mazmanov (디미트리 마즈마노프) tested the approach on a Java monorepo. He set up 2 modes: a "bulk reader" that reads multiple files and a "code writer" that writes predictable code snippets. Both modes use Google Gemini 2.5 Flash as the worker model that handles the actual tasks. The portal can also be switched to other models.

Mazmanov said, "Agents usually do not need to think and only perform input/output (I/O) functions." He said, "Because it can still be costly even if an AI model handles this work, we separated I/O tasks from the expensive Claude so that frontier models are used only for tasks where inference is actually needed."

Spotify stressed that worker models are not suitable for debugging, architectural decisions or core reasoning tasks. Delegated analysis does not always provide the line-level accurate information Claude needs when modifying code. Delegated tasks typically take 10 to 30 seconds, and the portal has a 30-second limit per call. The bulk reader and code writer modes were released through AiKA. Developers can reuse or modify them to fit their projects.