How I Cut My LLM API Bills by 40% (And Built a Free Cost Calculator)

Description and use

The author built an LLM cost calculator and a task routing strategy, cutting API bills by 40 % and simplifying multi-model management.

Problem

Cost observability was a nightmare; monthly bills changed unpredictably due to varying pricing models across providers.

Outcome

Implemented cost calculator and routing strategy cut LLM API bills by ~40% and streamlined multi-model management.

Workflow

• Build a frontend-only cost calculator to quantify token costs per model.
• Define a tiered routing strategy: Tier 1 for high-cost, high-intelligence tasks; Tier 2 for balanced tasks; Tier 3 for low-cost, high-speed tasks.
• Apply routing logic to direct 70% of agent calls to cheaper models, escalating to Claude only when confidence is low.
• Move to an API aggregator gateway to unify endpoints, handle failover, and simplify key management.

Evidence

Every time I tweaked the routing logic, my monthly bill changed in ways I couldn't predict.

Source

By routing the "dumb" tasks (which make up 70% of my agent's calls) to cheaper models, and only escalating to Claude when confidence is low, my bills dropped by nearly 40%.

Source

Using a unified endpoint means I just send one OpenAI-compatible request, and the gateway handles the routing and automatic failover.

Source