Cut API Costs by Routing Tasks to Cheaper Models

The math

A DeepSeek Flash call costs a fraction of what Claude Opus charges per token. Most day-to-day coding tasks — renaming variables, writing boilerplate, updating configs — do not need a frontier model.

With claude-multi you get two (or more) commands in your terminal. Each one launches a full Claude Code session, but pointed at a different provider. Pick the right one for the job and watch your bill drop.

Setting up the split

After installing claude-multi, open the TUI:

claude-multi

Create a budget instance using the DeepSeek template. This gives you a claude-budget command that routes through DeepSeek’s API at lower per-token cost.

Then create a power instance using the GLM template. This gives you claude-power for the harder tasks.

# Routine work -- cheap and fast
claude-budget "add error handling to all fetch calls"

# Hard problems -- frontier model
claude-power "redesign the data layer to support offline-first"

Both instances share the same Claude Code interface. Same /loop, same skills, same plugins. The only difference is which provider handles the request.

What you keep

Because each instance is still Claude Code under the hood, you lose nothing:

• All your MCP servers still work
• Plugins and skills carry over if you enable auto-sync
• Conversation history stays isolated per instance
• You can still use /compact, /loop, and every built-in command

When to use which

Tips

• Use auto-sync during setup so both instances share the same plugins and skills
• Check claude-multi TUI to see cost-relevant config per instance under Instance details
• Add a third instance for MiniMax if you want another price-performance tier