The monthly LLM bill lands, it's higher than expected, and now someone has to reconstruct four weeks of changes to figure out why. By the time you find the answer, the waste has already cost a month of runway. The fix isn't a better post-mortem process — it's not needing one, because the regression got caught in week one instead of week four.
That's what a 30-minute weekly cost review does. It's not a new tool or a new dashboard. It's a five-part ritual, run every Monday, that catches drift while it's still cheap to fix.
1. Monthly reviews let a regression run for up to four weeks before anyone notices. Weekly reviews catch it in days.
2. The agenda is five checks — total spend, top cost drivers, anomalies, cache hit rate, budget-versus-actual — each roughly five minutes.
3. The ritual only pays for itself once LLM spend is a real line item; under $1,000/month, skip it.
Why Monthly Is Too Slow
A cost regression is rarely dramatic. It's a system prompt that grew by 400 tokens over three sprints. It's a cache hit rate that slid from 30% to 14% because a new feature started injecting a timestamp into the cached prefix. It's a new endpoint that quietly defaulted to the frontier model because nobody set the routing rule. None of these show up as an alert — they show up as a slightly higher number that's easy to write off as "usage growth" until someone actually looks.
At a monthly cadence, that slightly-higher number gets four weeks to compound before it's even reviewed, and another cycle to actually fix. A weekly review compresses that to days: catch it Monday, ship the fix by Wednesday, confirm it in next Monday's numbers.
The 30-Minute Agenda
Five checks, run in this order, against whatever cost dashboard your team uses:
1. Total spend vs. last week
The single top-line number. Is it flat, up, or down — and does that match what you'd expect from traffic growth alone? A spend increase that outpaces traffic growth is the signal that something below the surface changed.
2. Top 3 cost drivers by feature
Not total spend by model — spend by feature. If the ranking shifted from last week, ask why. A feature that jumped from #5 to #2 in cost rank is worth five minutes of investigation before it becomes a habit.
3. New anomalies or spikes
Any single day or single endpoint with a spend spike outside the normal range. This is where runaway loops and leaked keys show up first — the incidents that cost $9K-$47K when nobody's watching daily are visible as a spike days before they're visible as a monthly line item.
4. Cache hit rate trend
A slow decline here is one of the quietest regressions there is — it doesn't fail loudly, it just costs more every week. Check it explicitly; it won't surface on its own in a total-spend view.
5. Budget vs. actual, by team
If more than one team shares the LLM budget, this is where a FinOps or engineering-management view earns its place — which team is tracking to plan, which is trending over, and whether that's expected (a launch) or not.
Open action items
Whatever came up in the first four checks gets a one-line owner and a target date. The review isn't complete until every anomaly has a name attached to it.
Want the exact dashboard setup?
Preto's weekly view maps directly to this five-part agenda — spend trend, top drivers, anomalies, cache rate, and team budgets in one place.
Start Your Weekly Cost Review — FreeOpen Preto every Monday. 5 dashboards. 30 minutes. Thousands saved.
What a Single Review Session Typically Catches
The specific dollar figure varies by team size and spend, but the shape is consistent. A team spending $10,000-15,000/month on LLM APIs that hasn't run a structured review before typically finds one or two of these in the first session:
| Finding | Typical Monthly Cost | Time to Fix |
|---|---|---|
| A feature quietly defaulted to the frontier model | $800-2,500 | Same day (routing config) |
| Cache hit rate degraded, unnoticed for weeks | $400-1,200 | 1-2 days (prompt restructure) |
| A new endpoint shipped with an unbounded output length | $300-900 | Same day (set max_tokens) |
| A retry loop on a flaky downstream dependency | $200-800 | 1 day (backoff + circuit breaker) |
Stack two or three of these — which is typical for a first review on a team that's never done one — and $4,000/month in caught waste is a realistic, not exceptional, outcome. The second and third reviews find less, because the obvious drift has been swept out. That's the point: the review's job shifts from "find the big waste" to "stop new waste from accumulating," which is a permanently smaller but permanently necessary job.
Why the Weekly Cadence Is the Part That Matters
The five checks above aren't complicated — any team could run them monthly, or even quarterly. What makes the weekly cadence non-negotiable is what happens between reviews if you don't run it that often. The same pattern shows up across teams that hold onto their savings: a cut that isn't protected by ongoing measurement drifts back up, quietly, until the next audit finds it again — except now it's a bigger number and a longer investigation.
The documented worst cases make the point starkly: a Stanford lab burned $9,200 in 12 hours on a forgotten token, a production multi-agent loop ran 11 days and $47,000 before anyone noticed. Neither of those needed a 30-minute review to prevent — they needed someone looking at the numbers more often than once a month. The weekly review is the cheapest insurance against becoming the next version of that story.
Frequently Asked Questions
How often should a team review LLM API costs?
Who should be in the weekly LLM cost review?
What should a weekly LLM cost review actually check?
Isn't a 30-minute meeting overkill for most teams?
Run Monday's review in Preto.
Spend trend, top drivers, anomalies, cache rate, and team budgets — the five checks in one dashboard, so the review takes 30 minutes instead of a morning stitching together spreadsheets.
Start Your Weekly Cost Review — FreeFree forever up to 10K requests. No credit card required.
