Your AWS bill is boring now. Your OpenAI bill is where the surprises live.

#ai #aws #monitoring #openai

Here's a pattern I keep seeing: a team ships an AI feature, usage grows, and three months later someone in finance asks why the Anthropic or OpenAI line item tripled. Nobody can answer immediately, because nobody was watching it the way they watch EC2 or RDS.

Cloud cost tooling has spent a decade building dashboards for compute, storage, and network. AI API spend is still mostly untracked — teams find out what happened after the invoice arrives, not before.

I added a free audit for exactly this gap. Here's what it checks and why each one matters.

Why AI spend behaves differently from cloud spend

Infrastructure waste is usually static — an idle EC2 instance costs the same whether anyone's using it or not. AI spend is dynamic: it scales with usage, model choice, and prompt design, all of which change weekly as a product evolves. A team can be doing everything "right" in January and burning 5x the necessary spend by March, purely because nobody revisited the model routing decision made at launch.

That volatility is exactly why this needs its own checklist instead of being bolted onto a generic cloud audit.

The 12 checks, grouped by category

Model Selection

Using the same (often frontier) model for every task, including ones a cheaper model handles equally well
Deprecated model versions still in production, which frequently carry legacy pricing
No tiering logic — no routing between "cheap enough" and "needs the best model available"

Caching

No response caching for repeated or near-identical prompts
No prompt caching for long, reused system prompts

Spending Controls

No monthly spend cap set on the provider account
No max_tokens limit, so output length is effectively unbounded
No alerting when spend crosses an expected threshold

Architecture

Dev and test environments calling the same paid production endpoints
Individual request calls where a batch API would cut cost significantly

Attribution

No way to tell which feature or team is driving spend
No token usage monitoring or trend visibility over time

Each of these maps to a specific, checkable thing in your codebase or provider dashboard — not a vague "optimize your AI costs" suggestion.

What the free report shows

Same format as the existing cloud cost audit: a spend score, estimated monthly and annual recoverable cost, and the flagged issues sorted by impact. No credentials requested — it's a self-assessment, not a scan of your actual account.

Here's what one flagged issue looks like in the preview:

Using deprecated model versions
Old model IDs carry legacy pricing that runs meaningfully higher than current equivalents for the same or better output.

The paid version (Blueprint, same price as the existing Cost Blueprint) adds the exact fix for each flagged issue — model routing code, caching implementation with the right API parameters, spend cap configuration steps, batch API migration code. Written to be implementable the same afternoon, not "here are some ideas."