I audited my SaaS stack across 10 platforms. Found a $6.46 phantom cost, a 67% error rate, and a bill way lower than I thought.

I was about to panic when PostHog said my SaaS was burning $6.46/week on embeddings. Turns out it was tracking paid pricing for usage that was fully under Cloudflare's free tier. Zero actual dollars. That kicked off a full audit across every platform I pay for.

## TL;DR

- Real monthly infra cost for aiapplyd: about $35 to $55 after Browserbase cancels next month. Not $104 like my dashboard suggested.

- Hidden stuff my tools weren't telling me: a phantom $6.46/7d tracking artifact, a 67% error rate in my enrichment pipeline, and a BYOK Gemini key routing $5/mo outside my "official" AI budget.

- If you're at 5 paying users on $100+ infra, do the audit. You'll find money and bugs.

## What is aiapplyd

AI-powered job application platform. ATS scoring, auto-apply, interview prep, resume builder, tracking. Solo founder, TypeScript, Cloudflare Workers, D1, Hono, TanStack Start. Five paying users right now. That number stings to type but it's the honest state of things.

## Why I audited

Two triggers, both uncomfortable.

First, my ChatGPT referral traffic collapsed this week. I'd been riding ChatGPT as my #2 channel for a month. Woke up to a red line on my chart. Made me want to know what I could cut if things got leaner.

Second, the math on 5 users at roughly $104/mo infra is $20.84 per user just to keep the lights on. That's not CAC, that's raw infra burn per existing customer. It felt wrong. I wanted to know if that number was real or inflated by dashboard lies.

## The method

Nothing fancy. Playwright MCP driving through every platform dashboard I could think of. Gmail receipt search for anything that renewed quietly. PostHog HogQL queries to cross-check what my own app was logging. Ten platforms total.

The goal was one row per platform: actual cash going out, not rate card estimates.

## The numbers

Claude Max personal sub $200. Resend Pro $20. Browserbase Developer $20 to $0 after May 12 downgrade. Cloudflare free tier $0 with $5 lifetime spend. OpenRouter prepaid credits about $3 to $5 on $35 lifetime. Anthropic API Haiku auto-topup $18 one-off for Stagehand browser. Firecrawl $19 one-off. PostHog free tier under 1M events. Stripe $0 outgoing platform fee. GCP Gemini BYOK about $5 routed via OpenRouter.

Project infra total about $50/mo now, $35 to $55/mo after Browserbase downgrade.

Claude Max is the single biggest line item on my personal card, but it's my coding IDE subscription, not infra. Stripped out, the project itself runs lean.

## The three hidden findings

### 1. Phantom $6.46/7d embedding cost

PostHog's LLM costs dashboard showed Workers AI embeddings as a paid line item. Added up to $6.46 over a rolling 7 days. Looked like it was going to be my fastest-growing cost.

I pulled the actual Workers AI billing page. $0.00. The embeddings model I'm using is inside the Cloudflare free tier bucket for my tier. PostHog was applying public rate card pricing to events my provider never billed me for. Tracking artifact. The dashboard lied and I almost made a product decision based on it.

Fix: adjusting my PostHog event properties so it doesn't apply dollar per token on free-tier paths. Filing with PostHog because I doubt I'm the only one.

### 2. BYOK Gemini leaking outside the official AI budget

I routed one subset of AI calls through a user-provided Google AI Studio key via OpenRouter BYOK. It worked. But it meant about $5/mo was sitting on a Google Cloud bill I don't open often, completely invisible to my OpenRouter total spend number.

Not a big number, but the lesson is: BYOK routes will silently bifurcate your AI accounting. Tag them, total them, or kill them.

### 3. 67% error rate in enrichment

This one I'm embarrassed by. I run an enrichment pipeline that scrapes and normalizes data for one internal job. HogQL showed 67% of invocations were failing and retrying. The user-facing flow was fine because of fallback logic. But I was paying for browser time, AI tokens, and retries on a loop that mostly died.

Root cause: a selector change on an upstream site three weeks ago plus an overly strict schema parse on my side. The fallback was so graceful I never noticed the underlying failure rate. Cost me an estimated 40 dollars of Browserbase and AI time over the last month, bundled into the totals above. Also a chunk of latency my users silently ate.

## What changed as a result

Kicked off 10 parallel agents to stream events into a single cost and reliability Telegram channel. Daily roll-ups so I don't have to log into 10 dashboards.

Set hard monthly caps on every prepaid provider. If I blow through, the service degrades instead of silently 10x-ing.

Weekly review of HogQL error-rate queries on anything that calls a paid API. Fallbacks are great, but fallbacks without alarms are just hidden bills.

Browserbase from $20/mo Developer down to Free tier. I was under the session limit anyway, I just hadn't looked.

## Lessons

1. Your observability dashboard is not your bill. They're two different systems with different assumptions. Reconcile them at least quarterly.

2. Free-tier usage gets priced on rate cards by default in most analytics tools. Either override the pricing field or ignore the total.

3. Graceful fallbacks are the enemy of error visibility. If a pipeline can silently retry, it will, and you will pay for every attempt.

4. BYOK is convenient but splits your cost truth across accounts. Consolidate or tag rigorously.

5. Audit the stack is not a one-time task. I'm adding it to my monthly cadence because a $200 personal sub can hide a lot.

## Honest ask

IH, anyone else running on roughly $50/mo infra and finding weird tracking artifacts in analytics or billing dashboards? I'd love to compare notes on where the reported number diverged from the bill. Specifically interested in PostHog LLM cost tracking, Cloudflare Workers AI free-tier accounting, and Browserbase session metering.

Also if you're a solo founder under 10 paying users and want to swap infra-cost spreadsheets, drop a comment. I'll share the Playwright MCP script I used to scrape all 10 dashboards.

Full build log and the phantom-cost deep dive going up on https://aiapplyd.com/blog this week.