Frontier Vs. Local: When Each Makes Sense

A FRONTIER MODEL is one of the largest, most capable models available, hosted by a company (Anthropic's Claude, OpenAI's GPT, Google's Gemini). You pay per use or per subscription. They're the smartest models you can access, and they're updated frequently.

A LOCAL MODEL is one you run on your own hardware. Open-weights models like Llama, Qwen, DeepSeek, Mistral, and others. You pay for the hardware once. After that, you pay nothing per query.

Frontier strengths:

• Smartest at hard reasoning, novel problems, complex code.
• Largest context windows.
• Best at "I don't know exactly what I want, help me figure it out."
• No hardware to manage.

Frontier weaknesses:

• Costs add up. $200/month is real money.
• Your data leaves your machine. Sometimes a dealbreaker.
• Subscription required. If your internet is down, so is your coding.
• Rate limits and quotas.

Local strengths:

• $0 marginal cost. Once you have the hardware, you can use it all day.
• Your data never leaves the room.
• No internet required.
• You own the stack.

Local weaknesses:

• Hardware costs upfront ($600-$10,000+ depending on what you run).
• Smartest local models are still meaningfully behind the smartest frontier models.
• You manage updates, drivers, model weights, and the inference server.
• Smaller context windows on most local setups.

The hybrid approach (what most working people end up at):

USE FRONTIER for: hard architecture decisions, complex codebases, debugging weird issues, anything where being right matters more than saving $0.50.

USE LOCAL for: routine refactors, boilerplate, simple data transformations, agent tasks that run all day, anything involving data that shouldn't leave the building.

The rule of thumb: if you'd hire a senior engineer for this task, use frontier. If you'd hire an intern, use local. If you can describe the task in three sentences and it has a clear right answer, use local.

A worked example: I use Claude Code (frontier) for new-feature work and debugging. I use a local Qwen model on a Mac for the agentic tasks that run on a schedule (the "every morning, summarize yesterday's data" crew). The combined cost is the Claude subscription plus electricity.