Saw a post on HN today about fine-tuning an LLM to write documentation like it's 1995. The idea is neat — train a model on your old docs, get new ones in the same style. But it got me thinking about the infrastructure side of things.
If you're fine-tuning, you're probably touching at least three different APIs. OpenAI for GPT-4 base, maybe Claude for comparison runs, DeepSeek for the cheap bulk processing. Each one has its own billing cycle, its own key management, its own rate limits.
I used to keep a spreadsheet. Seriously. Five API keys, five dashboards, five invoices at the end of the month. The worst part wasn't even the money — it was the context switching. You're in the middle of debugging a prompt, and suddenly you hit a rate limit on one provider. Now you're scrambling to switch keys mid-thought.
That's what pushed me to build AIOpenCloud. One key, one bill, models behind it just work. DeepSeek v4 for the heavy lifting, GPT-4 when I need it, Claude for the nuanced stuff. The routing happens at the gateway level — I don't think about it anymore.
The fine-tuning post also reminded me: when you're experimenting with multiple models, the cost adds up fast. Having a single gateway with transparent pricing means I can actually track what's happening. No surprise bills on the 1st.
If you're running multi-model workflows, how do you manage the keys? Still got that spreadsheet?
→ aiopencloud.xyz
Top comments (0)