Apple M5 local running model is 3 times more expensive than OpenRouter: how to choose local vs. cloud

Core conclusion

On May 17, 2026, an actual measurement analysis from William Angel detonated Hacker News (146 votes hot list). The author used detailed data to prove: **The comprehensive cost of running the AI ​​model locally on the M5 Max MacBook Pro ($4,299) is about 3 times that of using the OpenRouter cloud API, but the speed is only 1/3. **

The implications for AI practitioners are straightforward:

• Daily development/content production: Cloud API (OpenRouter) is the better choice - cheaper and faster
• Data Sensitive Scenario: Local inference is expensive, but has privacy and security advantages
• Scale Deployment: Electricity is very cheap, and hardware depreciation is the real big factor - don’t just count electricity

Key Points

• Time of incident: 2026-05-17 -Affected objects: AI content producers, independent developers, automation teams
• Core change: The "hidden cost" of local AI inference is revealed through data - hardware depreciation is the main force

Background and trigger events

William Angel is the author of the "Offline Agentic Coding" series. In this third installment of his article, "Apple Silicon costs more than OpenRouter," he calculates the combined costs of running Gemma 4 31B natively (with performance close to Anthropic Claude Sonnet), using an M5 Max MacBook Pro (64GB RAM, $4,299) as a benchmark.

The analysis dimensions include: electricity cost, hardware depreciation, inference speed (tokens/second), and are directly compared with the API pricing of similar models on OpenRouter.

Key Impact (by Dimension)

Adaptation suggestions

Based on this data, we can optimize our own AI tool usage strategy:

• Content production scenario: Prioritize the use of cloud APIs such as OpenRouter. Taking content automation workflow as an example, when processing hundreds of thousands of tokens every day, the cloud solution can save $500-$2,000 every year.
• Local Alternative: If you already have an M4/M5 Mac, you can use tools like LM Studio to run smaller models locally for offline testing and prototyping
• Hybrid strategy is the best: Put complex tasks of large models in the cloud, and put simple reasoning (such as summary and classification) of small models locally, balancing cost and speed.
• Cost Monitoring: Use OpenRouter's API usage statistics function to regularly check token consumption and expenses

Task List

• Evaluate the average daily token consumption of the current AI workflow
• Comparing actual annual cost of local inference vs OpenAI/n8n/OpenRouter
• For privacy-sensitive tasks, consider running it locally with LM Studio

Professional interpretation: Why hardware depreciation is the real cost trap

Many people think that running an AI model locally "only costs electricity" - this is the biggest misunderstanding. This article states:

• Electricity cost: M5 Max is only $0.02 per hour at full load, and only $0.48 at full capacity for a day.
• Hardware depreciation: $4,299 of equipment is amortized over 5 years, depreciating $0.10 per hour (5 times the electricity bill)

What really determines the cost-effectiveness of local inference is not the laws of physics, but your hardware usage. If your MacBook is only running AI inference 10% of the time, and the remaining 90% is writing documents, then the actual cost of "effective inference" per hour skyrockets 10 times.

Related extended information

• Apple Silicon costs more than OpenRouter —
• HN （146 ）
• lmstudio.ai — AI

Tool entry

When thinking about AI tool selection, OpenRouter provides an API platform that aggregates multiple models. ChatGPT, Claude, Gemini, etc. all have cloud-available versions. If you need a local solution, LM Studio is a good choice. For automated workflows, n8n can concatenate different AI API calls to achieve flexible hybrid strategies.

Internal link guidance

• Want to learn how? Watch: How to run local AI models on M4 Mac with LM Studio: A complete 30-minute tutorial
• Real case: He Built an AI Automation Stack with Claude + n8n — $4K to $12K/mo in 6 Months
• Advanced reading: DeepSeek V4 vs Claude Code: 90% Cheaper, Same Quality