Apple Silicon local inference costs 3x more than OpenRouter, analysis finds

A detailed cost breakdown from the Offline Agentic Coding blog reveals that running large language models locally on Apple Silicon is significantly more expensive and slower than using cloud APIs like OpenRouter. The analysis focuses on the M5 Max MacBook Pro with 64GB RAM ($4,299) running Gemma 4 31b—a model close to Anthropic Sonnet in performance. At 50–100 watts power draw and $0.20/kWh electricity, the per-hour electricity cost is roughly $0.01–0.02. However, the dominant cost is hardware depreciation: over a 5-year lifespan, the hourly hardware cost is ~$0.10, pushing total cost per million tokens to $1.50 at 10 tokens/second or $0.40 at 40 tokens/second.

In contrast, OpenRouter offers Gemma 4 31b at $0.38–0.50 per million tokens with speeds of 60–70 tokens/second—2–7x faster than local inference. The author concludes that for most use cases, especially when a human employee's salary is factored in (roughly 1,000x the token cost), using cloud services like Anthropic or OpenRouter is far more sensible. While it's impressive that consumer hardware can run near-Sonnet-level models, the economics and speed still heavily favor cloud inference.