Most teams discover expensive requests on invoice day — weeks after the deploy that caused it. CostObs puts a dollar sign on every span, so you know what each endpoint costs before your CFO does.
Who this is for
Your LLM calls cost $0.38 each. Your search calls cost $0.003. Only one shows up in traces.
GPT-4 and Claude API calls have 100–1000× more cost variance than traditional compute. You need cost visibility at the span level, not just the invoice level.
typical signal: LLM Inference is 90% of request cost
Some customers are 20× more expensive to serve than others. You’re subsidizing them.
You have a pricing model based on seats or usage. But you don’t know which customers are profitable to serve. CostObs lets you see cost per customer_id in traces.
typical signal: customer_id: enterprise-04 costs $12/day to serve
You have 12 microservices, OTel already deployed, and no idea which one is expensive.
You ship fast. Cost only becomes visible when finance escalates. By then, root cause is weeks old. CostObs turns cost into a real-time signal alongside latency and errors.
typical signal: POST /api/export is 91× more expensive than avg
The gap
What you have
Traces show what’s slow.
Latency is a first-class signal. P99 goes up — you get paged. Errors spike — you see it in dashboards. Slow spans glow red in Jaeger.
What’s missing
Nothing shows what’s expensive.
A new feature ships. Fast. No errors. But it added a new LLM call per request — turning a $0.003 endpoint into a $0.42 one. Nobody notices until invoice day.
The consequence
Cost bugs compound silently for weeks.
By the time finance flags the number, the engineer has moved on, the deploy is months old, and root cause is manual archaeology. CostObs makes cost a deploy-time signal.
The aha moment
endpoint cost · GET /api/reports/export
$0.136
per request — 91× your average. 89 users hit it yesterday. That’s $12/day, $360/month for a feature 3% of users actually use.
deploy regression · PR #482 · merged 2h ago · by @sarah
+$4,200/mo
POST /api/checkout went from $0.0018 to $0.0058 per request. At 3,400 calls/day, that’s $4,200 extra this month for one merged PR.
How it works
Add one exporter
Point your existing OTel Collector at CostObs. No SDK changes. No code deploys. No restarts.
Set service prices
Tell us roughly what each service costs. One-time config. Directional numbers are enough — relative cost is what matters.
See cost in every trace
Cost appears as a column on every span. Every endpoint, every scenario, every request — in real time.
Get alerted on regressions
Deploy cost regressions fire to Slack within minutes. Link to the trace, the culprit span, and the estimated monthly impact.
Live demo
This is real data from an AI chat endpoint. LLM inference is 90% of the cost. Click any span to inspect.
Total duration
2.1s
Total est. cost
$0.4200
Spans
5
Cost
$0.0010
Duration
20ms
% of request
0.2%
Cost
$0.0100
Duration
120ms
% of request
2.4%
Cost
$0.0200
Duration
80ms
% of request
4.8%
Cost
$0.3800
Duration
1.90s
% of request
90.5%
Cost
$0.0001
Duration
15ms
% of request
<0.1%
Alerts
Every deploy is automatically benchmarked. Cost regressions fire to Slack before the engineer closes their laptop.
Cost regression detected
deploy #1847 · triggered by @sarah
POST /api/checkout is 3.2× more expensive after your last deploy.
Before deploy
$0.0018
After deploy
$0.0058
Δ monthly
+$4,200
Most likely cause: discount-svc: getEligibility — new span introduced in this deploy, adds 68ms and 1 extra DB call to every checkout request.
→ View example trace from deploy #1847Deploy regression · order-service
PR #482 increased avg cost of POST /api/orders by 40%. New DB join examining 3× more rows. Estimated impact: +$1,800/mo.
2 hours ago · @sarah · deploy #1845
Cost spike · /api/ml/classify
Avg cost per request up 3× in 30 minutes. No deploy detected — probable upstream model latency increase.
31 minutes ago · no deploy associated
Cost improvement · /api/search
Index added by @marcos reduced avg cost by 62%. Saving $180/day at current volume.
Yesterday · deploy #1841 · resolved
Why now
LLM calls don’t cost cents. They cost dollars.
A GPT-4 call in a hot loop can cost more than an entire serverless function per day. AI features introduced 100–1000× cost variance that didn’t exist two years ago.
100–1000×
cost variance introduced by LLM endpoints
Eng teams are now accountable for spend.
Post-2022, “optimize later” stopped being acceptable. CFOs are asking VPs of Eng for line-item attribution. Most engineering orgs don’t have an answer.
$62B
wasted annually in cloud overspend (Gartner)
The instrumentation layer is already deployed.
OTel is the default. It’s in Kubernetes, every major APM vendor, cloud-native runtimes. The telemetry pipeline CostObs needs is already there — we just add cost to it.
~80%
of new cloud-native apps use OTel (CNCF 2024)
Accuracy
Directional. Not billing-grade. That’s enough.
CostObs computes cost = duration × configured price/ms per service. The absolute number might be off by 30–50%. That’s fine. The goal isn’t to replace your cloud bill — it’s to tell you which endpoint is 10× more expensive than your average, and whether that changed after a deploy.
Relative cost is what makes decisions. If /api/export is 91× your average endpoint cost, that’s true regardless of whether the absolute number is $0.136 or $0.098.
What’s accurate
Relative cost between endpoints is reliable — if A shows 10× more than B, that ratio is real. Cost trends over time are reliable — if cost goes up 3× after a deploy, that signal is real regardless of absolute calibration.
What’s approximate
Absolute dollar amounts depend on how accurately you configure service prices. CPU cost uses request allocation, not actual utilization. DB costs are duration-based, not query-complexity-based.
How to improve accuracy
Connect AWS CUR / GCP Billing Export (Team tier) for nightly reconciliation. Estimated vs. actual typically lands within ±10% after reconciliation.
Early feedback
“We’d been running that export endpoint for 8 months and had no idea it cost $0.14 per call. We fixed it in an afternoon. It paid for a year of CostObs.”
“The deploy regression alert fired before I’d closed my laptop. PR reverted in 12 minutes. That cost spike would have compounded for weeks without CostObs.”
“I showed the dashboard to our VP of Eng on day one. She asked me to roll it out to every team. We didn’t have to sell it internally — the trace view does that for you.”
30 min
median time to first cost insight
$0
code changes required to start
1 day
to catch a cost regression before finance does
Pricing
The trace view and dashboards are free forever. You shouldn’t have to pay just to see what your system costs. Upgrade when you want to be told about cost regressions before you notice them yourself.
The upgrade trigger
You’re on free tier. A deploy ships Friday at 5pm. Cost goes up 3×. You don’t find out until Monday when you check the dashboard.
That’s when teams upgrade to Pro — when they want the alert to fire at 5:02pm instead.
Free
forever · no card · no expiry
Pro
per workspace · 14-day trial · cancel anytime
Team
workspace · or custom enterprise pricing
Get started
Connect your OTel Collector in 30 minutes — or upload a trace JSON right now. No account needed to try the demo.
Free forever for up to 3 services No credit card required Works with your existing OTel stack