case 01verbatim
“Our OpenAI bill was $1,200 last month. I spent two hours trying to figure out what changed between October and November. Still don't know.”
Every team we talked to has had this conversation.
We're in development! Things may crash or break
v1.0 · sdk for openai · anthropic · gemini
You have an AI bill. You have no idea what's in it.
Two lines of code. Every LLM request in your app, tracked by feature, user, and model. Finally know which part of your product is burning money.
Wraps OpenAI, Anthropic, Gemini, and any OpenAI-compatible API. No architecture changes. No proxy.
LIVE · prod.us-east-1
spend.todaytotal estimated spend · today
$41.83
updated 09:04:49 AM
chat$29.0769.50%
summarize$9.4122.50%
embed$3.358.00%
requests/sec · 14.2p95 latency · 412mserrors · 0.03%
02 · the situation
Does this sound familiar?
case 02verbatim
“I added a userId tag to every API call by hand, built a spreadsheet, and checked it every Friday. That's how we tracked AI costs. For six months.”
Engineering time is expensive. This shouldn't be a spreadsheet problem.
case 03verbatim
“We launched a summarization feature. Two weeks later our bill tripled. Turns out we were using GPT-4 for something gpt-4o-mini handles just as well at 1/20th the cost.”
The model was right. The routing wasn't. You can't fix what you can't see.
03 · the product
From invisible API spend to operational data — in two lines of code.
server/openai.ts
1npm install @trackllm/sdk2 3import { track } from '@trackllm/sdk'4import OpenAI from 'openai'5 6const openai = track(new OpenAI(), {7 projectKey: 'proj_xxxx',8 feature: 'chat',9 userId: req.user.id10})// That's it. Every call you make through this client is now tracked — with model, tokens, cost, latency, user and feature attached.
cost · last 7 days
$519.82
chat$412.30 +22%
summarize$89.40 −4%
embed$18.12 +1%
feature
model
avg/req
chat
gpt-4o
$0.0091
summarize
gpt-4o-mini
$0.0011
embed
text-embedding-3-small
$0.0002
chat ↑ 22% vs last weekview all →
04 · how it works
Up and running before your next deploy.
step 01setup
Wrap your client
SDK wraps your existing OpenAI or Anthropic client. One import, one function call. No proxy, no architecture change, no new environment to manage.
$
step 02setup
Tag your calls
Pass a feature name and user ID per call. Optional — but this is what turns raw telemetry into the breakdown that actually tells you something.
>
step 03result
See where the money goes
Every request logged: model, tokens in, tokens out, estimated cost, latency, user, feature. The dashboard builds the breakdown for you.
→
05 · what you get
Everything you've been building in spreadsheets.
Cost by feature
See which features drive spend. Not the whole bill — the specific feature. Built a new summarizer last week? See exactly what it cost.
chat
$406
summarize
$198
embed
$62
Cost by model
gpt-4o at $0.0091/request. gpt-4o-mini at $0.0011. The same feature, two models. Now you have the data to make the switch.
modelavg/reqmonthly
gpt-4o$0.0091$412.30
gpt-4o-mini$0.0011$48.20
claude-haiku$0.0008$31.10
Cost by user
Rank users by what they cost to serve. Find outliers, detect abuse, decide if your pricing model actually covers your per-user AI spend.
usr_8a4f…
outlier$48.12
usr_2c91…
$12.04
usr_5b7d…
$9.88
Spend alerts
Set a daily or monthly threshold. Get an email or Slack message when you cross it — not when you open the invoice.
⚠ alert · daily14:22
Spend crossed $50.00 threshold · +38% vs 7d avg.
Request explorer
Every API call, searchable and filterable. Sort by cost. Filter by user. Click into a request to see tokens, latency, finish reason.
POST /chatgpt-4o$0.01921.4s
POST /chatgpt-4o$0.00880.9s
POST /sum4o-mini$0.00110.4s
Weekly digest
Every Monday: last week's cost, biggest mover, projected month-end spend. The data you used to build manually — delivered automatically.
✉ mon · 09:00
Last week: $519.82
Biggest mover: chat ↑22%
Projected month: $2,140
Biggest mover: chat ↑22%
Projected month: $2,140
06 · pricing
Start free. Pay when it's clearly worth it.
Free
no card$0/mo
For solo builders and early prototypes
- →Up to 20k requests/month
- →1 project
- →Basic cost breakdown (model-level)
- →Basic threshold alerts
- →7 days of request history
- →Email notifications only
recommended
Growth
$49/mo
For teams running AI features in production
- →Up to 150k requests/month included
- →3 projects
- →Cost by user + feature breakdown
- →Advanced alerting (per-user and per-feature thresholds)
- →Multi-channel alerts (Email, Telegram)
- →Weekly digest reports
- →Request explorer (limited depth)
- →Revenue tracking per user
- →60 days of request history
- →Team access (5 seats)
Scale
$99/mo
For high-usage teams needing deep insights
- →Up to 1M requests/month included
- →Unlimited projects
- →Full alerts system (user, feature, model-level thresholds and cost spikes)
- →Email, Telegram and Slack alerts
- →Full request explorer
- →Session & prompt logging ( coming soon )
- →Revenue tracking per user
- →180 days of request history
- →Unlimited seats
- →CSV export
- →Priority support
Both plans include a weekly digest email, SDK for all major providers, and no changes to your application architecture.
07 · the why
Built from real pain.
We interviewed 40+ engineering teams about their AI cost management. 38 of them were doing it manually — spreadsheets, console logs, or not at all.
The 2 who weren't had built their own internal tooling. That tool is what this is.
hn · sep 2024
“We used GPT-5 for phone number extraction for six months before we noticed.”
▲1,693 upvotes·412 comments·"this is me"
Two lines of code. Know where your AI money goes.
The first event usually arrives in under a minute. No proxy. No architecture change. No monthly invoice surprise.
terminal
1npm install @trackllm/sdk08 · faq
Real questions, real answers.
Most questions fall into four buckets: latency, privacy, scope, and how this differs from observability tools. Real answers below.
Anything missing? founders@trackllm.dev
No. The SDK wraps your existing OpenAI/Anthropic/Gemini client in-process. Your requests go directly from your server to the model provider. We never sit in your request path. If FluxGate goes down, your app keeps shipping tokens.