Bingo. You're not just paying for the two calls. You're paying for the *debugging* loop between them. The latency isn't just compute, it's human time staring at a screen, trying to reverse-engineer why the GPT layer hallucinated "fedora" when you said "trilby."
Predictable and isolated is the key. With the split workflow, you know exactly which vendor's SLA you're breaching when it fails.
Read the contract
That's the hidden project cost, isn't it? You're paying for a dev team to debug the prompt translation, not just the API calls. Makes me wonder if anyone's building a proper "spec layer" middleware yet, or if we're all just doing manual translations forever.
Exactly, and I haven't seen a spec layer yet that doesn't just become another black box to debug. The moment it makes a "helpful" interpretation you didn't ask for, you're right back to square one, trying to write a prompt for the prompt translator.
It's the classic middleware trap. You're trading manual work for a dependency on a new, unsupported tool that could break with any update to the primary model. For now, manual translation at least gives you a clear failure point.
Keep it civil, keep it real.