Analyze
We read your agents' real logs and price every call against a live model index. You see spend per agent, per model — the breakdown no one else gives you.
scan · admin api · proxyno signup · no card · no keySpend less
Spend less
on AI.
We read what your agents actually run, price every call, and show you where the money goes. Then we show you exactly what to switch.
$ npx whatmodel scan → your real audit in 15s
What you'd savelive estimate
Estimated recoverable / month
$2,992/mo
≈ 37% of spend — $35,907/yr
model-downshiftsemantic cacheprompt compression
built for the agents you already run
OpenClaw·Hermes·CrewAI·AutoGPT·LangGraph·AutoGen·or your own loop
545
models indexed and priced
4
categories of waste we detect
15s
to your first real audit
0
keys to see your first number
the problem
Your tenth AI bill is the scary one.
Agents fan out across models, retry, re-prompt, and quietly burn opus tokens on work haiku could do. The spend is real, per-agent, and invisible — no dashboard tells you which agent wasted what.
We tie every dollar to the agent that spent it — the part nobody else does automatically.
claude-code · main loop$3,140
claude-code · subagents$2,980
cursor · autocomplete$1,210
custom rag agent$ 680
this month$8,010
how it works
Analyze. Optimize. Monitor.
One managed service — what RDS is to Postgres. Run the scan or paste your keys once; we run the cost ops from there.
Optimize
We find the cheapest model that does the same job, the repeats worth caching, and the prompts safe to trim — and show you exactly what to change.
model-downshift · cache · prompt-compressionMonitor
Drift and regression detectors watch every model you run. The moment quality slips, we flag it and tell you what to revert.
drift · regression · alertssee exactly what we'd switch
The dashboard is the evidence.
agent reasoning stepmodel-downshiftsemantic cachesaved $0.0497
tool-call summaryprompt compressionsaved $0.0121
draft generationmodel-downshiftsaved $0.0312
why managed
Managed. Not DIY.
The other tools sell you the parts. We sell you the result. Same gap as self-hosted Postgres vs Amazon RDS — and the same reason teams stop running their own.
| DIY tools | whatmodel | |
|---|---|---|
| Instrumentation | wrap every SDK, plumb traces | paste a key, we read what's already there |
| Dashboards | wire your own metric → chart | built-in audit + activity feed |
| Routing rules | JSON config, deploy, hope | a tested swap plan, scoped to your policy |
| Regression watching | alerts you forget to read | we flag regressions and tell you what to revert |
by the numbers
Every call. Every model.
545
models indexed, repriced from upstream
every call
priced against the index, per scan
4
savings categories detected per audit
15s
to your first audit — $0, no keys
manifesto
Neutral. By design.
We don't sell a model. We don't pick favourites. We watch what your agents actually do, price it honestly, and point you to the cheapest equivalent for the day you ran it. Your keys stay yours. The recipe stays private. The savings are yours to keep.
Read the manifesto →design partners
Built with the first builders.
We are taking on a small group of early design partners to shape whatmodel Capture around real agent stacks. You run the free audit, we sit with your numbers, and your feedback steers what we build next.
Become a design partner →free audit first — no keys, no commitment
pricing
One flat fee. Priced to your spend.
A predictable monthly fee for managed cost and performance ops — your tier is set by how much AI spend we manage, not by a usage meter. The audit is always free.
Free Audit$0— we scan your stack and show you what you'd save before you pay a cent.
Starter
$99/mo
up to ~$2k/mo in AI spend
Pro
$399/mo
up to ~$20k/mo in AI spend
Scale
$1,499/mo
up to ~$200k/mo in AI spend
Enterprise
custom
$200k+/mo — your data, your tenancy
Request early access →your tier moves with your spend — no surprises, no usage meter
for the agents you buildgive this to your agent
Reach us from anywhere.
Don't touch a terminal. Paste this into the agent you already run.
Drop it into Claude Code, Cursor, Cline, or OpenClaw. It runs the scan, reports your per-agent cost audit, and — only when you say OK — pushes it and hands back your claim link. Read-only, metadata-only. We advise on what to change — your live traffic stays untouched.
Agent prompt
You are helping me audit my AI/agent costs with whatmodel. Do this in order, and stop for my OK before the final step: 1. Run this exactly, from my project root: npx whatmodel scan It is read-only and metadata-only: it reads model identifiers, token counts, latency and cost — never my prompt or response content. Do not change any config, do not edit any files, and never repoint a base URL or API endpoint at whatmodel (whatmodel advises; it does not route my live traffic). 2. Summarize the result back to me in plain language: which agents/stacks it found and the per-agent / per-model cost breakdown. Do not invent numbers — only report what the scan output. 3. Then STOP and ask me to confirm. Only after I explicitly say OK, run: npx whatmodel scan --push and give me back the claim URL it prints so I can open my audit. If the scan reports nothing or errors, tell me what it said — do not retry with extra flags.
Prefer the native tool path? @whatmodel/mcp is live:
claude mcp add whatmodel npx @whatmodel/mcpnpx whatmodel scan
CLI
audit any stack from the terminal
@whatmodel/mcp
MCP server
add whatmodel to your agent — it runs the cost scan as a tool
claude mcp add whatmodel npx @whatmodel/mcpAgent Skills
skills
ship a whatmodel skill to your agent
/llms.txt
llms.txt
machine-readable site map for agents
docs mcp
docs MCP
live docs as an MCP tool
See where your
money goes.
No signup. No card. Run the scan, see your real numbers, keep them if you like what you see.