Delivered. Check your inbox for the letter.
DeepSeek in Claude Code, the setup.
Cheap models · Better systems
AI Operators — watch me run this stack live →
The setup below is one wire in a bigger machine. Inside AI Operators I show the whole thing — my content engine, my funnels, and the real numbers, live every week.
01 · Get a key
Create an API key at platform.deepseek.com and load a few dollars. Pricing: $0.14 per million input tokens, $0.28 output — and cache hits drop to $0.003 (a 98% discount that ranks #1 of all 101 models tracked by Artificial Analysis).
02 · Two environment variables
export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic export ANTHROPIC_API_KEY=<your deepseek key> claude
That's the whole setup — DeepSeek ships an Anthropic-compatible endpoint, documented here. Open a new terminal without the variables and you're back on regular Claude.
03 · The gotchas
- • It's a reasoning model — long thinking on hard tasks. Great for real work, slower for quick edits.
- • 1M-token context, but verbose output (about 2x the median model) — cheap tokens, just more of them.
- • Keep your real Claude setup for the tasks where frontier polish matters; route the volume work to the cheap endpoint. Cheap models, better systems.
- • Model quality check: it scores 50 on the Artificial Analysis Intelligence Index — #3 of 101. Median is 25. See the live benchmarks.
This is one wire in a bigger machine. If you want to see how I run cheap models inside systems that make money — my content engine, my funnels, the numbers — AI Operators is where I show everything.