Skip to content
PromptUnit.ai

One line of code. 40-70% lowerAI bill.

Swap your base URL. PromptUnit routes each call to the cheapest model that handles it. Full cost visibility from day one. No other code changes.

Try live demo

No setup · 30 seconds · No API key needed

Works with

OpenAIAnthropicGoogle GeminiGroqDeepSeekMistralTogether AIPerplexityxAICohere
Question 1 of 4:

What is this?

An AI proxy that reduces your costs automatically. Sits between your app and your AI provider, no code changes required.

See it in action

Before you connect anything.

Run a real prompt.

Watch how we choose the model.

See your cost drop instantly.

Try the demo

The problem

You're overpaying for AI.

And it doesn't show up anywhere obvious.

40–70%

of AI spend is wasted

Your app sends every call to GPT-4o. Sentiment checks, summaries, simple Q&A — a $0.30/M model handles those just as well. The difference goes straight to OpenAI.

$0

cost visibility by feature

OpenAI sends you one number. No breakdown by endpoint, feature, or call type. You can't cut what you can't see.

1 line

to fix all of it

Not a rewrite. Not a new SDK. Change base_url and you get full visibility, automatic routing, and real savings from the first call.

Before you enable anything

See your exact savings before routing goes live.

Connect in minutes. Your traffic runs in observation mode. We show you the numbers. You decide when to flip the switch.

  • Real-time cost breakdown

    Every API call logged by model, feature, and cost, the moment it happens.

  • Savings forecast before routing is on

    See exactly how much you would save. No risk, no routing until you click.

  • Routing decisions explained

    Know why each request was routed where, not just that it was.

Try the live demo, no API key needed
app.promptunit.ai / dashboard
connected · acme-prod

This month's AI spend

$273

down 43% vs $480/mo baseline

Saved this month

$207

43% of your bill

Total API calls

48,200

this month

What-if Analysis

You could be saving this much. Routing is off.

ConservativeBalancedAggressive

Current monthly cost

$480

without optimization

With PromptUnit

$273

balanced routing mode

You save

$207/mo

43% of your bill

Recent Routing Decisions

Why each request was sent to the model it was.

live
Task typeRequested → Routed toSaved
classification
gpt-4ogpt-4o-mini
+$0.00420
summarization
gpt-4oclaude-haiku-4-5
+$0.00780
extraction
gpt-4ogpt-4o-mini
+$0.00310
coding_complex
gpt-4o
-
Observing · 2 days leftday 12
Watching your traffic
We log every call and shadow-classify it. Zero routing, zero risk.
Calls logged412,807
Routable share71.4%
Forecast savings$5,440 / mo
Enable Routingone click · reversible
After you enable
days 15+
Projected savings: $5,440 / mo
Same SDK, same endpoints, same quality. Lower bill starting with the next request.
Calls routedlive
Median added latency41ms
You pay-

How it works

Up and running in minutes

01

Connect

Add your provider API keys and swap your base URL. No SDK changes, no refactoring.

Takes ~5 minutes
02

Analyze

Watch your spend dashboard populate in real time. See cost broken down by model, feature, and user segment.

Data from first call
03

Save

Enable smart routing. We route each request to the cheapest model that clears your quality bar.

Savings from day one

Built for production

Routing is the start.
This is what comes after.

Most proxies tell you which model ran. PromptUnit tells you which feature is overpaying, warns you when quality drops, and stops runaway spend before it reaches your provider.

Feature Attribution

Tag API calls with x-promptunit-feature and see exactly what each part of your product costs: calls, spend, savings, and save rate per feature. No other proxy gives you this breakdown.

customer-support
12,400$18.2041%
document-summary
3,100$31.5067%
search-assist
8,800$4.4028%

Quality Regression Alerts

PromptUnit scores every response and monitors per-task quality in real time. When a model starts degrading on your specific traffic, you get an email before your users notice. No other proxy does this automatically.

Quality alert

classification dropped to 58% avg quality

38 calls in the last 4 hours

Alerts fire at most once per 12 hours per task type. Rate-limited, not noisy.

Circuit Breaker Controls

Set hourly and daily spend limits. Define a spike threshold. Choose to auto-downgrade or block on anomaly. Configured in your dashboard, enforced at the proxy layer before any cost hits your provider.

Hourly limit$25 / hr
Daily limit$200 / day
On spikeDowngrade model
StatusActive, no blocks today

All three included. No add-ons, no enterprise gate.

Pricing

We only make money when you save.

No subscription. No flat fee. We take 20% of what we save you. Nothing else.

Performance-based pricing
20%of verified savings

That's it. No subscription, no flat fee. We earn only when you save.

No savings. No charge. Ever.

  • Unlimited API calls proxied
  • All 10 provider integrations (OpenAI, Anthropic, Google, Groq, DeepSeek, Mistral, Together, Perplexity, xAI, Cohere)
  • Real-time cost analytics dashboard
  • Smart model routing with quality guardrails
  • Full request/response logging
  • Email spend alerts
Start free audit

No savings

No charge

Cancel anytime

No contracts

Free to start

No card needed

99.9% uptime

Auto failover

Your numbers

How much are you leaving on the table?

Drag the slider. See your estimate in seconds.

ROI Calculator

Estimate your savings

$2,000
$500$50k
Your current monthly spend$2,000
Estimated savings$800 – $1.3k
PromptUnit fee (20%)$160 – $260
Your net saving$640 – $1.0k/mo

Estimates based on observed routing patterns. Actual savings depend on your traffic mix and quality threshold. 14-day observation period shows your exact numbers before any charge.

See your real numbers free

14-day observation. No routing, no charge until you see your exact savings.

Security & Privacy

Your data stays yours. Always.

We never read, store, or retain your prompt content. We do not train on your data. Ever.

Zero prompt storage

Your prompt content is never written to disk. Not to our database, not to logs. The request passes through. The content does not.

Encrypted in transit

All traffic between your app, PromptUnit, and AI providers travels over TLS 1.3. Your keys and requests are encrypted at every hop.

Keys encrypted at rest

Your provider API keys are encrypted with AES-256-GCM before being stored. Each key gets a unique random IV. Even with database access, the keys are unreadable.

No training on your data

We never use your traffic to train models or share prompt signals across customers. Quality fingerprinting is statistical, never content-based.

One line change. Full savings.

Swap your base_url to our proxy. Keep every other line of code exactly as it is. PromptUnit routes calls between models automatically. Your SDK, error handling, and response parsing are untouched.

  • Same OpenAI SDK, no new imports or dependencies
  • Automatic failover if we're ever unreachable
  • Works with Python, Node.js, Go, Ruby, any HTTP client
  • No downtime. We never sit in your critical path
  • Privacy mode: log nothing if you prefer
  • 14-day observation before any routing change

Works with any OpenAI-compatible SDK: Python, Node, Go, Ruby

integration.py
# Before. direct to OpenAI
from openai import OpenAI
client = OpenAI(base_url="https://api.openai.com/v1")
# After. one line change, all the savings
from openai import OpenAI
client = OpenAI(
base_url="https://api.promptunit.ai/api/proxy/openai",
api_key=your_promptunit_key
)

Works with any OpenAI-compatible SDK: Python, Node, Go, Ruby

Stop guessing your AI costs.
See exactly where you're wasting.

Connect in minutes. We run in shadow mode and show you real savings data before anything changes. Activate routing when ready.

Try the demoRead the docs

14-day observation period · 5-minute setup · Cancel anytime

Why PromptUnit

Built different from every proxy you've tried

Other tools route requests. PromptUnit optimizes them. Before, during, and after.

CapabilityPromptUnitLiteLLMOpenRouterCloudflareHelicone
Smart routing✅Partial✅❌❌
Prompt compressionunique✅ManualPartial❌❌
Token inflation defenseunique✅❌❌❌❌
Dialect translation✅✅PartialPartial❌
Prompt efficiency scoringunique✅❌❌❌❌
Circuit breaker✅PartialPartial❌Partial
Aligned pricing model✅❌❌❌❌

Based on publicly available documentation as of April 2026. Features reflect default capabilities without custom implementation.

Layer 14Only PromptUnit

Prompt Compression

TF-IDF compression removes redundant tokens before the request leaves your server. Savings happen before you're billed. Not instead of billing.

Layer 19Only PromptUnit

Token Inflation Defense

Detects malicious prompts designed to artificially inflate your token count and bill. The only proxy that treats this as a security problem, not just a cost problem.

Layer 15Only PromptUnit

Prompt Efficiency Advisor

Scores every prompt 0–100. Tells you not just what you spent, but why you overspent and which features to fix. This turns your dashboard into an action plan.

Advanced Intelligence

Algorithms that get smarter with every request

Five compounding intelligence layers that run on top of routing. They improve automatically as traffic grows. A new entrant starts at zero. You don't.

Layer 23Intelligence

Prompt Complexity Classifier

Scores every prompt across 8 axes before a single token is sent. Detects reasoning depth, constraint density, and code indicators. Routes simple requests to cheap models without burning tokens on the routing decision itself.

Layer 24Cost

Semantic Request Cache

Fingerprints incoming requests using normalized content hashing. Returns a cached response when an equivalent request was seen recently. Zero API cost. Hit rate compounds with volume.

Layer 25Quality

Multi-Model Consensus

Detects high-stakes requests (medical, legal, financial, infrastructure) and runs dual cheap-model verification. If they agree, returns consensus. If they diverge, escalates to a flagship. Flagship quality, cheap-model price.

Layer 26Intelligence

Cross-Customer Quality Oracle

Aggregates anonymized quality signals across all platform traffic to build a real-world per-model, per-task performance index. Every request across all customers trains it. No single customer can build this.

Layer 27Intelligence

Adaptive Threshold Learning

Watches implicit feedback signals and automatically adjusts your quality threshold over time. The longer you stay, the more personalized routing becomes. Switching cost grows automatically.

The data flywheel

Every routing decision, quality signal, and cache hit feeds back into the system. Your routing gets more accurate the more you use it. OpenAI can't build this. Open-source proxies won't.

Features

Everything your AI stack needs

Routing, observability, failover, and cost optimization. Everything any AI-powered app needs, without building it yourself.

Smart Routing

10-dimension task classification routes every call to the cheapest model that clears your quality bar. No more paying GPT-4 prices for GPT-4o-mini tasks.

Only PromptUnit

Prompt Compression

TF-IDF compression removes redundant tokens from your prompts before they leave your server. Savings happen before you're billed, not after.

Only PromptUnit

Token Inflation Defense

Detects and blocks token inflation attacks, malicious prompts designed to bloat your bill. The only proxy with a security story, not just a cost story.

Only PromptUnit

Prompt Efficiency Scoring

Scores every prompt 0–100 for efficiency. Tells you not just what you spent, but why you overspent and which features to fix first.

Multi-Provider

OpenAI, Anthropic, Google, Groq, DeepSeek, Mistral, Together, Perplexity, xAI, and Cohere. One proxy. Switch providers without touching application code.

Zero Code Change

Swap your baseURL to api.promptunit.ai. Your existing OpenAI SDK, response parsing, and error handling all continue to work exactly as before.

FAQ

Frequently asked questions

PromptUnit is an AI inference proxy that sits between your app and your AI providers (OpenAI, Anthropic, Google Gemini, Groq, DeepSeek, Mistral, Together AI, Perplexity, xAI, and Cohere). It automatically routes each request to the cheapest model that meets your quality bar, with zero code changes required. You change one line (your base URL), and PromptUnit handles the rest: routing, failover, cost tracking, and quality validation. You pay 20% of what we save you. If we save nothing, you pay nothing.
We maintain a continuously-updated benchmark of model outputs across task categories. When a request comes in, we classify the task type and route to the cheapest model whose benchmark score meets your configured quality threshold. You can tune the threshold per route or globally.
Your app keeps running. Our SDK has built-in automatic failover: if PromptUnit is ever unreachable, requests are instantly rerouted directly to your provider (OpenAI, Anthropic, etc.) with no action required on your side. You lose the optimization and savings during that window, but your users never see an error. We also send proactive email alerts the moment an incident is detected, and a follow-up when it is resolved. Current uptime: 99.9%.
No. The Quick Start integration is a single line change. Swap your base URL to api.promptunit.ai and you are live in under a minute. For production environments we recommend our SDK (npm install @promptunit/sdk) which adds automatic failover: if we are ever unreachable, your requests go directly to your provider transparently. Both options use the same OpenAI-compatible interface. No other code changes needed.
Yes. Your requests pass through our servers for routing and optimization. We log metadata only: token counts, model names, latency, and cost. Prompt content and completions are never stored, never written to disk, and never used to train models. What you see in the dashboard is cost and usage data, not your prompts.
Yes. From the dashboard you can switch between two logging modes. Standard mode logs all metadata: tokens, cost, model, task type, and feature tags, which powers your full dashboard analytics. Privacy mode logs only token counts and cost. Feature names and task classifications are never stored. Routing works identically in both modes. The classification still happens in memory, it just isn't written to disk.
For every request we compare two numbers: what you actually paid using the routed model, and what you would have paid using the model you originally requested. The difference is your gross saving for that call. We charge 20% of the total gross saving each billing cycle. All prices come directly from official provider rate cards and are updated regularly. You can verify any number in your dashboard, which shows a per-request cost breakdown.
Simple: we take 20% of what we save you. No subscription, no flat fee, no hidden charges. During the first 14 days we observe only. No routing, no changes. After the observation period, routing goes live and we charge you only after we have already saved you money. No savings means no charge.
OpenAI, Anthropic (Claude), Google (Gemini), Groq, DeepSeek, Mistral, Together AI (Llama 3.3/3.1), Perplexity Sonar, xAI (Grok 3), and Cohere. All 10 use OpenAI-compatible endpoints, so you can connect any of them in the onboarding flow and PromptUnit routes across all of them automatically.