Public beta is live  ·  $500 in free credits for early teams

The infrastructure layer
for your AI stack.

infras sits between your app and every LLM API you call — logging every token, routing every request, and cutting your AI bill, all without touching your code.

$ pip install infras-gateway # works with your existing OpenAI SDK

Drop-in replacement for the OpenAI SDK. One line to integrate. <2ms proxy overhead.

app.infras.app/overview

Overview

Last 30 days ▾
Gross spend
$1,284.50
▲ 12.4% vs last month
Saved by infras
$612.30
▼ 32.3% net spend
Requests
1.24M
18.9k / day avg
Avg latency
412ms
+2.1ms proxy overhead
Daily spend Gross Saved
Spend by model
gpt-4o
41%
claude-3.7
27%
gemini-2.5
18%
deepseek-v3
9%
other
5%
Cache hit rate 31.8%
ModelRouteTokensLatencyCostStatus
gpt-4o/v1/chat/agent2,418388ms$0.0112cache · 92% off
claude-3.7/v1/chat/summarize6,102841ms$0.0293200
gemini-2.5-flash/v1/embed/batch48,920156ms$0.0008200
gpt-4o-mini/v1/chat/classify312204ms$0.0002200
claude-3.7/v1/chat/agent9,3381,204ms$0.0441retried → 200

One gateway · every major model

OpenAIAnthropicGeminiMistralLlamaDeepSeek
Product

Everything between your app and the models.

infras gives your team full visibility and control over AI spend, latency, and reliability — from a single drop-in gateway.

Full cost visibility

Every request logged with tokens, cost, latency, and errors. Break down spend by feature, team, model, or customer — in real time, not on your monthly invoice.

Smart routing

Send each request to the cheapest model that can handle it. Simple work goes to fast small models; hard reasoning escalates automatically. Cut spend up to 40%.

Semantic caching

Repeated and near-duplicate requests are answered instantly from cache. Users see the same quality, you pay a fraction of the tokens.

Automatic failover

Provider outage or rate limit? Requests reroute to a healthy provider in milliseconds and retry with backoff — your users never see the 500.

Spend guardrails

Set per-team and per-feature budgets with hard caps and soft alerts. The days of a runaway loop torching your credit card are over.

AI-powered insights

Gemini-powered summaries explain your bill in plain English: what spiked, why, and exactly which routing rule would have prevented it.

How it works

Live in three steps. No code rewrite.

infras speaks the OpenAI API, so the SDK you already use keeps working — you just change one URL.

01

Point your SDK at infras

Swap the base URL in your existing OpenAI, Anthropic, or LangChain setup. That's the whole integration.

02

We log and route every call

Each request is metered, cached, routed to the best provider, and retried on failure — with ~2ms of added overhead.

03

Watch the savings stack up

The dashboard shows spend, cache hit rate, and reliability per feature and model — and suggests routing rules as it learns.

# before: client = OpenAI()
from openai import OpenAI

client = OpenAI(
    base_url="https://gateway.infras.app/v1",
    api_key="infras_sk_...",  # your real keys live in the dashboard
)

resp = client.chat.completions.create(
    model="gpt-4o",  # or "auto" to let infras pick the cheapest capable model
    messages=[{"role": "user", "content": "Explain vector databases"}],
)
// before: new OpenAI()
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://gateway.infras.app/v1",
  apiKey: "infras_sk_...", // your real keys live in the dashboard
});

const resp = await client.chat.completions.create({
  model: "gpt-4o", // or "auto" to let infras pick the cheapest capable model
  messages: [{ role: "user", content: "Explain vector databases" }],
});
# Point any OpenAI-compatible request at the gateway
curl https://gateway.infras.app/v1/chat/completions \
  -H "Authorization: Bearer infras_sk_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "auto",
    "messages": [{"role": "user", "content": "Explain vector databases"}]
  }'
Pricing

Pay for what you route. Nothing else.

Start free, scale as your AI usage grows. Every plan includes the full gateway, dashboard, and caching.

Hobby

$0 /month

For side projects and early experiments.

  • 50k requests / month
  • Full cost dashboard
  • Basic caching
  • Community support
Start free

Enterprise

Custom

For companies with compliance and scale needs.

  • Unlimited requests
  • Self-hosted or VPC deployment
  • SSO, audit logs, PII redaction
  • 99.99% SLA & dedicated support
Talk to us
FAQ

Questions, answered.

Does the proxy add latency to my requests?
The gateway adds roughly 1–3ms of overhead per request. Responses stream straight through — caching and routing decisions happen at connection time, so model latency is untouched.
Which providers are supported?
OpenAI, Anthropic, Google (Gemini via Vertex AI), Mistral, DeepSeek, Meta Llama, and any OpenAI-compatible endpoint. If your provider isn't listed, the generic adapter usually works out of the box.
Do you store my prompts and completions?
You choose. Metadata (tokens, cost, latency, status) is always logged; request bodies are optional, off by default on paid plans, and can be PII-redacted before storage. Your data is never used to train models.
Can I self-host infras?
Enterprise plans include a self-hosted or VPC-isolated deployment on your own Google Cloud project, with the same dashboard and control plane.
How is "auto" routing different from just picking models myself?
With model: "auto", infras classifies each request's complexity and routes it to the cheapest model likely to succeed — then learns from retries and feedback. Teams typically save 30–40% versus hardcoding one flagship model everywhere.

Stop guessing what your AI costs.

Point your first request at infras in under five minutes.

Get started free

No credit card · 500 free credits · works with your existing SDK