01
Sign up free
No credit card. You get a PromptUnit key instantly.
Dashboard · API keys
pu_••••••••••••••••7c21Copy
Key created. No card needed.
Stop paying flagship prices for simple calls.
PromptUnit checks every AI call your app makes and sends it to the cheapest model that's just as good. Up to 70% off your AI bill.
01The waste
40–70%
lower AI bill: our estimate for an app with many simple requests. Tagging a ticket, pulling fields from an email, a two-line summary: your app sends every one of them to the model you picked for the hard ones.
2settingsto change in your app
$5freebefore you pay anything
02How it works
Set it up once. Then forget it's there.
02
Point your app at us
Two settings, wherever your app's settings live: the API address and the key.
Your app · environment variables
OPENAI_BASE_URLhttps://www.promptunit.ai/api/proxy/openai
OPENAI_API_KEYyour PromptUnit key
03
Ship as usual
Same code, same SDK, same response format. Nothing else in your app changes.
Your app · requests
04
Watch the savings add up
Every request, by feature and by model, with what you saved on it.
PromptUnit · This month (example)
Saved$0
Requests0
03The engine
Cheaper only when it's just as good.
Each call is classified across 10 signal dimensions and sent to the cheapest model that clears your quality bar. Hard tasks and sensitive topics get a stricter bar, so they stay on the model you chose or move only to one that is nearly as strong.
task: classification
asked forgpt-4o
answered by…
78%
of 32 side-by-side checks on a live app's real requests: the cheaper model's answer matched or beat the original.
−88%
cost per call on those requests.
5
providers connected directly: OpenAI, Anthropic, Google, Groq and DeepSeek. Plus OpenRouter, with your own key.
One live app, Oct 5 to 8, 2026, rated by the model the app chose · How we measured
04Setup
Pick your setup. Change two things. Done.
about 2 minutes
Two settings where your app runs
Set OPENAI_BASE_URL to https://www.promptunit.ai/api/proxy/openai and OPENAI_API_KEY to your PromptUnit key, then redeploy.
05Control room
Every request comes with a receipt.
PromptUnit.aiapp.promptunit.ai / dashboardexample account
This month's AI spend$273down 43% vs $480 baseline
Saved this month$20743% of your bill
Total API calls48,200this month
classificationgpt-4o → gpt-4o-mini+$0.00420
summarizationgpt-4o → gpt-4o-mini+$0.00750
extractiongpt-4o → gpt-4.1+$0.00070
judgmentgpt-4o → gpt-4okept
Feature attributionSee what each feature costsx-promptunit-feature
Quality alertsHear it before users doclassification dropped to 58% avg quality
Spend limitsCaps that hold$100 / hr · $500 / day by default
Nightly learningTunes itself every nightyour quality results decide what moves
06Security
Built to forget your prompts.
Model, tokens, latency, costlogged
Task type, routing decision, feature taglogged
Prompt textnever
Completion textnever
System promptsnever
End-user IDs and IP addressesnever
Keys encrypted at rest with AES-256-GCM and decrypted in memory only. TLS in transit. Never used for training. Metadata is kept while your account is open and deleted with it. No SOC 2 yet, and we say so plainly.
07Pricing
20%
You only pay out of what you save.
No subscription, no flat fee. 20% of verified savings, charged weekly, and nothing at all if there are none. Your first $5 of savings is free, no card needed.
Example: your AI bill without PromptUnit$1,200
Your actual spend through PromptUnit$380
Savings, split$164 fee · $656 yours
$500$50k
Estimated savings$800 – $1.3k
Our fee (20%)$160 – $260
You keep$640 – $1.0k/mo
Our estimate for an app with many simple requests, not a guarantee. Real savings depend on your traffic mix and quality bar. Nothing is charged before you've saved money.
08FAQ
Good questions. Straight answers.
Usually not. You change two settings wherever your app's settings live: the API address and the key. If your code sets them itself, that's two lines. If Lovable or Bolt built your app, you paste one message into its chat and it makes the change.
Apps that send plain text requests: support chat, email triage, summaries, tagging, data extraction, translation. Requests that carry tool results always stay on the model you asked for, so nothing breaks. Requests with images or tools move only where there is proof: a cheaper model that passed the side-by-side checks on that step's images, or a Claude model known to handle the same tools. Coding tools like Claude Code and Cursor send tools and tool results on almost every request, so they see little savings for now.
A request only moves to a cheaper model when that model clears your quality bar for that kind of task. Hard tasks and sensitive topics (medical, legal, money) get a stricter bar, so they stay on your model or move only to one that is nearly as strong, and crisis topics never move. If a cheaper model fails, the request is retried on the model you asked for. Routing adds a small delay to each request. The optional quality checks run after your app already has its answer, so they add no delay.
No. Requests pass through in memory to be routed. Prompts and responses are never stored. We log metadata only: which model, how many tokens, what it cost. We also keep a 2 to 3 word name for each step of your app, written once by AI on your own key. Privacy mode skips it.
Nothing breaks. Your app keeps working on the models you chose and savings pause. We email you at $2.50, at $4 and when savings pause. Add a card and savings start again, and from then on you pay 20% of what we save.
Yes. Every account has hourly and daily limits, $100 and $500 by default, and you can change them. When one is reached, requests stop until the window resets, so a runaway loop can't keep spending.
Make your next AI bill the smaller one.
Two settings to connect. Savings from the first request. First $5 of savings free.