How it works
From one request to a verified saving.
Change one line in your app. Parezen routes each request, records what it cost and what it would have cost, and proves the gap on your provider bill.
Early access. The X-ray is planned and not live yet.
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.parezen.com/v1",
apiKey: process.env.PAREZEN_API_KEY,
});Every request goes to the model that does the job.
The gateway reads the route and task from your key and headers, checks the policy, and sends the request to the cheapest model that still meets your quality bar. Every decision is recorded.
Requested model Cheaper model that passed your quality bar
Routing decisions
Sample data
- Support repliesgpt-4.1 gpt-4.1-mini26% lower
- Docs searchclaude-sonnet claude-haiku22% lower
- Nightly evalgemini-2.5-flash gemini-2.5-flash-lite42% lower
- Coding reviewgpt-4.1 gpt-4.1-mini6% lower
Illustration on sample data. Percentages come from the sample month. Routing is planned, not live.
Read, find, switch, count.
Planned. A sample walk-through of one saving.
- 1
Read the bill
Parezen reads the usage you already pay for and splits it by feature, model and team. You see which part of the product spends what.
Spend by feature, one month
- 2
Find the cheaper path
For each feature it looks for a smaller model, a cache or a shorter prompt that does the same job, then tests it on a sample of real requests before anything changes.
Support replies
Now
gpt-4.1
$8,200
Candidate
gpt-4.1-mini
$6,100
Shadow test on real requestsPasses your bar - 3
Switch, inside your rules
A change goes live only if it passed the test and your policy allows it. Budgets and hard stops still hold. You can review, enable or ignore each one.
Routing policy, support route
- Right-size support repliesShadow test passed
- Monthly budget hard stop$9,000 for this route
- Apply changes automaticallyOff: you approve each one
- 4
Count it against your baseline
Every saving is the gap between what the same work cost before and what it costs now, at the same prices. It becomes verified when the month matches the provider bill.
Savings ledger line
Baseline$8,200−Actual$6,100=Gap$2,100- Estimated
- Realised
- Verified
Verified when the month matches the provider bill within 2%. Only verified savings are billable.
Three levels of proof
A saving climbs one step at a time. A month outside the 2% tolerance is shown as disputed and never billed.
- Estimated
What the audit expects
From your usage export and published prices. Shown in the X-ray and the calculator. Never billed.
- Realised
What real traffic saved
Measured request by request after a change goes live: the baseline model's cost minus what you paid, at the same price version.
- Verified
What your provider billed
The month's ledger matches the provider invoice within 2%. Only verified savings count toward the savings share.
Nothing switches until it passes your bar.
Anything that can change answer quality goes through the same four steps.
Shadow test
The cheaper model answers a copy of real requests. Your users only ever see the original answer.
PlannedReview
Quality and cost side by side against the bar you set. Every review is recorded with who, when and why.
PlannedEnable
Turn the switch on for one route. Automation runs only where your policy says it may.
PlannedWatch and roll back
Quality and spend stay on watch. One click puts the route back on the model you asked for.
Planned
Your servers run the traffic. Parezen keeps the books.
Hybrid by default. The two sides talk only through signed, outbound-only sync.
Data plane
- The gateway container, next to your app
- Your own Postgres, or the bundled one
- Your provider keys
- Request logs, and prompts only if you turn them on
- Budget hard stops, enforced locally
Control plane
- Dashboard, onboarding and the free audit
- Price book, policies and licences
- Savings ledger and bill match
- Billing connectors and invoices
- Never your prompts, responses or provider keys
Startups that would rather not run anything can use Hosted Quickstart: change one URL and bring your own keys.
Designed to stay out of your way.
Rules the product is built around. None of it is live yet.
Never blocks your traffic
A billing or licence problem switches to pass-through. Requests keep flowing while optimisation pauses.
Hard stops stay local
Budgets you set are enforced in your gateway, so they keep working even if Parezen is unreachable.
Metadata only by default
Tokens, model, cost and latency. Prompt and response capture stays off unless a project turns it on.
Price cuts are yours
When a provider lowers a price, your baseline is re-priced too, so the cut never lands in our fee.
Want the full rules? How savings are verified
See where your AI spend goes.
Try the free AI bill X-rayEarly access. The X-ray is planned and not live yet.