How a single API router can unify OpenAI, Claude and Gemini costs for small teams

A developer building RAG and agent features in Python found that token cost management consumed roughly a third of their workload, prompting them to route all AI requests through a single API layer instead of managing three separate vendor SDKs and keys. Using one router consolidates credentials, usage tracking, and per-model pricing into a single interface, though it adds an extra network hop and delays access to the latest vendor-specific features. The developer learned a costly lesson when an evaluation harness quietly ballooned from 2.4 million to 19 million tokens per night after a configuration change, ultimately accounting for 41% of total token spend. They now recommend routing commodity tasks like summarisation and classification through the cheapest available model, while calling vendor SDKs directly only for exclusive features such as Anthropic's prompt caching or Gemini's large context window. The core advice is to meter every component of an AI pipeline — including eval suites — and to derive live pricing from the same API being called rather than relying on vendor pricing pages that change without notice.