Running local models on Apple Silicon is 3 times more expensive than OpenRouter calls

Core conclusion

On May 17, 2026, developer Will Angel published a hotly discussed article "Apple Silicon costs more than OpenRouter", which proved through detailed data comparison: the actual amortized cost of running a local large language model on the M5 MacBook Pro is actually 2-3 times more expensive than calling the cloud model through OpenRouter.

Key Points

• Event Time: 2026-05-17
• Core findings: Apple Silicon local inference costs about $1.5-$4.8 per million tokens, and OpenRouter’s similar model only costs $0.5-$1.5
• Affected objects: All individual developers, independent entrepreneurs, and content automation teams who run AI models locally
• The biggest revelation: Hardware accelerated depreciation is an invisible killer, but electricity costs are the lowest

Background: A big discussion about the cost of "local vs. cloud

The data released by Will Angel on his personal blog attracted widespread attention from the Hacker News community (48+ points that day). He is using the latest M5 Max MacBook Pro (64GB memory, official website price $4,299), running the Gemma 4 31B model - a model with performance close to the Anthropic Claude Sonnet level.

He made detailed calculations from three dimensions: power cost, hardware depreciation, and inference speed:

Electricity Cost: 50-100W power consumption, calculated as $0.18/kWh, only $0.009-$0.018 (about 1 cent) per hour Hardware depreciation: $4,299 equipment is amortized based on 3-10 years of life, $0.05-$0.16/ hours Inference speed: Gemma 4 31B is about 10-40 tokens/s on M5 Max

Key impact analysis

Adaptation suggestions: When should I use which one?

This is not a "black and white" issue. Choose according to your actual scenario:

Select local inference (Apple Silicon) scenario

• Privacy Sensitive: Processing of customer data, business secrets, and personal privacy content
• Latency sensitive: requires millisecond response (local inference is slow but network latency is lower)
• Offline/Mobile Scenario: No stable network connection
• Development and Debugging: Frequent parameter adjustment and experimental testing (pay-as-you-go payment will accumulate)

Select cloud inference (OpenRouter, etc.) scenario

• Mass Production: Large-scale content generation, data processing
• High throughput requirements: A large number of requests need to be processed simultaneously
• Zero upfront investment: Idea verification stage, don’t want to bet on $4,000+ hardware
• Multi-model comparison: OpenRouter supports dozens of models to switch on demand

Mixed strategy (recommended)

• Local model for daily development/debugging
• Use OpenRouter for production-level batch tasks
• When using n8n + AI Agent to build automated workflows, set up intelligent routing: simple tasks go locally, and complex tasks go to the cloud.

Further reading

This topic has sparked a lot of discussion, and I recommend the following content to get more perspective:

• Apple Silicon costs more than OpenRouter — Original text
• Hacker News — A collision of community opinions
• I don't think AI will make your processes go faster — Alternative thinking on AI efficiency

Tool entry

The tools and services covered in this article include: OpenRouter, Apple Silicon, Claude, n8n, DeepSeek, Gemini. The above platforms have detailed guides on WayToClawEarn.

Internal link guidance

• Newbie wants to build an AI automated workflow? Read the detailed tutorial first: AI Agent Tools 2026 Complete Tutorial: 5 Tools to Build an Automated Pipeline in 30 Minutes
• Want to know how to use n8n + AI to automate content and make money? Refer to this real case: I used n8n + OpenAI to build AI content automation website: a complete review of monthly income $4,500
• Want to replace costly API calls with local models? Check out this tutorial: How to run local AI models on M4 Mac with LM Studio: A complete 30-minute tutorial