AIVory Smart Inference · Routing

Early Access

Same AI. Cheaper bill.

Q: Is Smart Inference compatible with OpenAI?

Yes. It implements the OpenAI /v1/chat/completions endpoint. Change your base_url to https://smart.aivory.net/v1, use your AIVory API key, and keep the same SDK, model names, streaming, and tool-calling code.

Q: How do I migrate from OpenAI?

Change one line. Set base_url to https://smart.aivory.net/v1 and use your AIVory API key. Model names, streaming, tool calling, and JSON mode all work identically.

We reroute every request to the cheapest inference provider — or spin up a spot GPU for you. Fully OpenAI-endpoint compatible: swap one line, keep your SDK, prompts, and streaming. Credits from $10, no subscription.

Get early access See how easy it is

Why it works

Providers cut prices. You're paying like they don't.

One URL change routes every call to the cheapest endpoint running right now.

Price spread across providers

The same model costs 2–5× more on one API than another. We find the cheapest one live.

Cheaper via AIVory

Not a one-time saving — the router finds the cheapest endpoint on every single request.

Line of code

Swap your base_url. Keep the SDK, prompts, tests, retries, streaming handlers. All of it.

What it is

An AI inference router that finds the cheapest provider for every request.

Smart Inference is a routing proxy that sits between your application and more than ten inference providers. Every time you make an API call, the router scores every available endpoint by cost, latency, and availability, then forwards the request to the cheapest one that meets the quality bar. The result comes back in the same format your code already expects.

Behind the scenes, the router checks live pricing from providers like Together AI, DeepInfra, Fireworks, Groq, Cerebras, and AWS Bedrock. It also monitors spot GPU capacity from RunPod, Vast.ai, Crusoe Cloud, and Azure Spot. When a new price drop appears, the router picks it up within seconds.

The endpoint is fully OpenAI-compatible. It implements /v1/chat/completions with support for streaming, tool calling, JSON mode, and vision. You keep your existing SDK (Python, TypeScript, or curl), your model names, and your prompts. The only change is a single line: point your base_url at https://smart.aivory.net/v1.

Pricing is pure pay-as-you-go. Buy credits from $10, set an optional monthly spend cap, and pay the routed provider's token rate. There is no subscription, no per-seat fee, and no markup on the token price. Every response includes headers showing the cost, the provider that served it, and the other candidates the router considered.

How it works

Three steps. That's the whole onboarding.

No new SDK to learn. No model aliases to memorise. Low enough to try without a procurement call.

Join early access

Request early access, add a card, and buy credits from $10. Set a monthly spend cap (default $100, change it anytime).
balance: $10.00 · cap: you decide
Swap base_url

Point your existing OpenAI SDK at our URL. Keep the same model names, prompts, tools, streaming handlers.
base_url = "https://smart.aivory.net/v1"
We route every call

Each request lands on the cheapest provider that can serve it right now. If one fails, we try the next. You see a receipt per request.
X-SI-Cost · X-SI-Score · X-SI-Candidates

Smart Inference pricing

Live prices. Straight from the router.

Pulled from our live catalogue, sorted cheapest first.

These are the cheapest per-model prices our router sees right now. When cheaper capacity comes online, the number moves. Get early access to see per-request routing and historical trends.

Llama 4 Maverick $0.33−45% $0.27

DeepSeek V3.2 $0.22−56% $0.28

Gemma 4 31B $0.07−42% $0.05

Qwen 2.5 72B $0.20−44% $0.16

Savings = difference to the most expensive provider serving the same model.

Model not listed? Browse the full catalogue or launch your own GPU for any open-weight model.

Price comparison

How Smart Inference pricing compares.

Snapshot of what the router sees right now. The live table above updates in real time; this is the crawlable version.

Model	Params	AIVory (input)	Priciest provider	Savings
Llama 4 Maverick	400B	$0.33/M	$0.60/M (Azure AI)	45%
DeepSeek V3.2	671B	$0.22/M	$0.50/M (DeepInfra)	—
Gemma 4 31B	31B	$0.07/M	$0.12/M (Together AI)	—
Qwen 2.5 72B	72B	$0.20/M	$0.36/M (Together AI)	—

These are real per-million-token input prices. The router picks the cheapest provider live, so the rate you actually pay is often lower than this snapshot. All models are open-weight and served through the same /v1/chat/completions endpoint. No proprietary models are resold — if you need GPT-4o or Claude, call those vendors directly and use Smart Inference for everything else.

GPU Marketplace

Spot GPUs. Deploy in one click.

Live spot pricing from multiple providers. Self-host open-weight models with a single click, or let Smart Inference manage the rented capacity for you behind the same OpenAI-compatible endpoint.

Pre-baked images for Llama, Mixtral, Qwen, Stable Diffusion — boot in ~90 seconds.
Cheapest hourly rate across providers, refreshed every 30s.
Pay per-second, max-price guard rails, auto-shutdown on idle.

Browse marketplace Read the docs

PROVIDERS

app.aivory.net Snapshot

GPU Provider $/hour 24h Δ Stock

B300288GB RunPod $8.189/hr — Available Deploy

MI300X192GB RunPod $2.348/hr — Available Deploy

B200180GB RunPod $6.478/hr — Available Deploy

H200141GB RunPod $4.708/hr — Available Deploy

A100 80GB80GB Crusoe $1.534/hr — Available Deploy

H10080GB RunPod $2.348/hr — Available Deploy

L40S44GB Vast.ai $0.569/hr — Available Deploy

The swap

Your existing code. Cheaper bill.

If you've already written an OpenAI-compatible call, you've already written this one.

Point base_url at us. That's the migration. Your SDK, retry logic, streaming handlers, tool-calling schemas — untouched. The response shape is identical. Only the bill changes.

from openai import OpenAI

client = OpenAI(
    base_url="https://smart.aivory.net/v1",
    api_key="sk-aivory-...",
)

resp = client.chat.completions.create(
    model="llama-3.3-70b",
    messages=[{"role": "user", "content": "Ship it."}],
)
print(resp.choices[0].message.content)

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://smart.aivory.net/v1",
  apiKey: process.env.AIVORY_API_KEY,
});

const resp = await client.chat.completions.create({
  model: "llama-3.3-70b",
  messages: [{ role: "user", content: "Ship it." }],
});

console.log(resp.choices[0].message.content);

curl https://smart.aivory.net/v1/chat/completions \
  -H "Authorization: Bearer $AIVORY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama-3.3-70b",
    "messages": [{"role": "user", "content": "Ship it."}]
  }'

That's the whole migration. Get early access and try it →

FAQ

Common questions.

Which AI API is cheapest?

It changes by the minute. Smart Inference checks every provider in real time and routes to whichever is cheapest right now. Open-weight models like Llama 4 Maverick can be served for as little as $0.33 per million input tokens versus $0.60 on Azure AI.

Is Smart Inference compatible with OpenAI?

Yes. It implements the OpenAI /v1/chat/completions endpoint. Change your base_url to https://smart.aivory.net/v1, use your AIVory API key, and keep the same SDK, model names, streaming, and tool-calling code.

How much does Smart Inference cost?

Pay-as-you-go credits from $10. You pay the routed provider's token rate with no markup on the token price. There is no subscription and no per-seat fee.

What models are supported?

All major open-weight models including Llama 4 Maverick, DeepSeek V3, Qwen 2.5, Gemma 4, Mixtral, Phi, and Mistral. The router checks availability and pricing across more than ten providers for every request. Browse all models with live pricing →

Can I also rent a GPU instead of using the API?

Yes. The GPU Marketplace lets you deploy spot GPUs by the second for self-hosting any model.

What happens if a provider goes down?

Automatic failover. If the cheapest endpoint errors, the router retries on the next cheapest candidate. Response headers (X-SI-Cost, X-SI-Score, X-SI-Candidates) show which provider served the request.

How do I migrate from OpenAI?

Change one line. Set base_url to https://smart.aivory.net/v1 and use your AIVory API key. Model names, streaming, tool calling, and JSON mode all work identically.

Is there a free tier?

Credits start at $10. There is no free tier, but there is also no minimum spend, no subscription, and no commitment. Buy credits when you need them.

Pricing

Pay as you go.

Buy credits from $10. Spend them however you like—no subscription, no seats, no commitment.

You never pay more than the on-demand rate for the same model. Prices update in real time.

Llama 3.3 70B · priciest provider 10M output tokens / month

~$8.80

Llama 3.3 70B · via AIVory Same model, same volume, cheapest provider

~$2.80

Same model, same request. The router picks the cheapest provider live, every time.

No subscription
No seat fees
Set your own spend cap

Get early access How billing works →

Cheaper inference. One URL. Today.

The router is live. Capacity is limited during early access — join now to lock in your spot.

Get early access Read the docs