AI coding at home without going broke
Points and comments are a snapshot, not live.
Three approaches to affordable AI coding at home: self-hosting, API rentals of open models, or frontier subscriptions, with hybrid strategies offering best value.
Self-hosting requires steep upfront costs and only pays off with long-running tasks, as home-runnable models lag frontier labs. Renting open source models via API (like OpenRouter) avoids hardware investment and lock-in while maintaining flexibility. Frontier subscriptions from OpenAI and Anthropic offer roughly $2800 API value for $400/month but hit usage ceilings quickly for all-day agent workflows.
Optimal strategy combines the last two: use frontier models for planning and specification, then deploy cheap open models for implementation tasks. With spec-driven development, a single engineer can output roughly what a twenty-person team produces monthly for around $1000.
What commenters are saying
DeepSeek's direct API dominates the thread as a cost-optimized alternative. Users report spending $10 over weeks using DeepSeek V4 Flash at $0.14 per million input tokens versus Claude's $15 per million, with V4 Pro benchmarked at 80-90% of Opus quality while costing roughly $35 per million tokens versus $250. One commenter notes DeepSeek's metered, non-subscription model suits burst coding workflows better than fixed monthly plans.
A secondary discussion questions self-hosting power costs versus human brain efficiency, with rough estimates suggesting a human + HVAC costs roughly 1.5x a frontier model's power draw per workday, making the AI comparison context-dependent on actual task parallelism and human baseline costs.