AI API infrastructure

AI agents make a lot of model calls.
Make them cost less.

Run Anthropic, OpenAI and xAI models through one OpenAI-compatible API. One key, one prepaid balance, per-request cost visibility, and lower pricing on eligible models. Eligibility is decided per model. Where there is a verified advantage, the direct price and the Astrodyne price are shown side by side before you send traffic.

View pricing Get API Key

For AI products and agents where one action triggers many model calls. Sign up, add funds and generate a key yourself — no sales call. Sending requests is invite-only while we roll out.

  • OpenAI-compatible
  • One API key
  • Per-request cost visibility
  • Model-level pricing
  • No platform fee
ASTRODYNE one key one balance one account your app OpenAI Anthropic xAI ORB · 03 TRAJECTORIES RELAY · 01 INTEGRATION
Model families on the launch plan One key, one balance, one account

Agent economics

One action. Dozens of model calls.

A multi-step agent turns one customer action into a long tail of model calls — planning, retrieval, reasoning, tool use, retries and verification each add inference cost. Every one of those calls gets its own request ID, tokens and charge, and they all settle against one prepaid balance.

  • Every call carries its own request ID, tokens and charge.
  • All of them settle against one prepaid balance.
  • A runaway loop stops at the funded balance, not at the invoice.

Astrodyne is the model API underneath your agent, not the agent. Planning, retrieval, tool use and orchestration stay in your application.

Astrodyne for agents →
One user action becomes many model calls on one balance Inside your application, a single user action fans out across five agent steps — planner, retriever, reasoner, writer and reviewer. Each step emits several model calls. Those calls leave your application and reach Astrodyne, where every one settles against one prepaid balance with one spending limit. Astrodyne runs the model calls; your application runs the agent. YOUR APPLICATION MODEL CALLS ASTRODYNE One user action Planner Retriever Reasoner Writer Reviewer Astrodyne One balance One limit
Each dot is a model call. Illustrative — not a measurement.

Model ecosystem

Model families on the launch plan.

One integration reaches every family below, and Astrodyne runs the model you name — never a substitute. Your console and GET /v1/models show the live models available to your account.

  • Anthropic 3 models on the launch plan FAM · 01
  • OpenAI 6 models on the launch plan FAM · 02
  • xAI 2 models on the launch plan FAM · 03

Families are listed from the launch catalogue. Astrodyne runs the model you name and does not substitute a different one.

Why Astrodyne

Model costs compound with every call. So does the advantage of buying at scale.

As your product makes more model calls, model economics become infrastructure. A small per-token difference is negligible on one request and meaningful across millions.

Large AI buyers can get better commercial terms by consolidating demand and holding direct provider relationships. Astrodyne aggregates that infrastructure behind one API, so smaller accounts can reach those economics without negotiating separately with every provider.

Direct

One relationship per provider.

Working with providers directly Four separate rows. Each row holds its own account, agreement and invoice before reaching one model — twelve relationships to reach four models. account key invoice model
  • 4 accounts
  • 4 agreements
  • 4 invoices

Astrodyne

One relationship, the same reach.

Working through Astrodyne One account, one agreement and one balance reach the same four models. Astrodyne holds the provider relationships. account key astrodyne models
  • 1 account
  • 1 agreement
  • 1 balance

You keep the account, the key and the balance. Astrodyne keeps the provider relationships.

Cost

See the difference before you build.

Eligibility is decided per model, not across the catalogue. Where Astrodyne has a verified advantage we publish both figures side by side, so you can check the arithmetic rather than take a percentage on trust.

Pick a model

GPT-5.6 Sol verification pending
Direct provider
—
Astrodyne
—
Difference —

Verified model comparisons appear here as prices are approved — a figure is published only when there is an approved Astrodyne price and a current published direct price to compare it against. How we compare prices.

  1. 1Approve a price for the model
  2. 2Read the current published direct rate
  3. 3Show both, and the difference

Integration

Change the base URL and key. Keep your code.

Keep the stack you already have. Anything speaking the OpenAI API — your client, or the model layer beneath your agent — reaches Astrodyne by changing the base URL and key.

base_url="https://api.openai.com/v1"
base_url="https://staging.astrodyne.ai/v1"
api_key=OPENAI_API_KEY
api_key=ASTRODYNE_API_KEY

Two lines. The OpenAI SDK works unchanged.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://staging.astrodyne.ai/v1",
    api_key=os.environ["ASTRODYNE_API_KEY"],
)

resp = client.chat.completions.create(
    model="YOUR_MODEL_ID",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)
ASTRODYNE_API_KEY in your environment · base URL https://staging.astrodyne.ai/v1

YOUR_MODEL_ID is a placeholder — replace it with a model ID from your console. See the compatibility guide for supported fields and known differences.

Risk

Nothing here is hard to undo.

Astrodyne sits in your request path. These are properties of the implementation, not promises about it.

  • OpenAI-compatible
  • Model you name runs
  • No prompt storage
  • Revocable keys
  • Prepaid spend limit
  • No platform fee
  • Leaving is the same two lines as arriving Point the base URL and key back at your provider. Nothing in your integration is Astrodyne-specific.
  • The model you name is the model that runs No automatic substitution, no silent downgrade. An unavailable model fails and says so.
  • Prompts and responses are not stored Request records keep metadata only — ID, model, status, tokens, latency and charge. Logs are redacted.
  • Spend cannot exceed the balance you funded Every request reserves funds before it runs, so a balance cannot go negative.
  • No subscription and no platform fee You pay for settled usage at the price shown for that model. Adding funds is the only commitment.
  • Keys are hashed, scoped and revocable The secret appears in exactly one response, at creation. Revoke it from the console without touching your code.

How security works · How we compare prices · Compatibility scope

How it works

Three steps, then you are building.

01

Connect once

Point your OpenAI-compatible client at the Astrodyne base URL with an Astrodyne key.

02

Choose a model

Send the model ID you want. Your account sees exactly the models it can call.

03

Track usage and cost

Every request returns an ID. Look it up for the model, status, tokens, latency and charge.

In the console

Every request, priced and accounted for.

Look up any request by ID and see its model, status, tokens, latency and charge, against one balance — no estimating backwards from a monthly invoice.

  • Per-request — every call carries its own ID, tokens and charge.
  • Prepaid — the balance and the spend against it, in one place.
  • Per-key — issue, scope and revoke keys without touching your code.

Capabilities

One integration instead of one per vendor.

One integration

One OpenAI-compatible API instead of a client, an auth scheme and a billing relationship per vendor.

Visibility into requests

Look up a request by ID and see its model, status, tokens, latency and charge.

Lower cost on eligible models

Where Astrodyne has a verified advantage, the direct rate and the Astrodyne rate are shown side by side, per model.

Simpler billing

One prepaid balance in US dollars. Settled usage is charged at the approved price for the model you used.

Useful for AI agents, apps, automations, prototypes and production workflows.

Cost mechanics

Why AI API costs add up.

Every product with real usage pays this arithmetic once per call. An agent pays it many times over for a single user action.

  1. One user action a click, a message, a job
  2. Many model calls plan, retrieve, check, retry
  3. Tokens in + out priced separately
  4. A cost per call set by the model
  5. Your monthly bill the sum of all of it

One user action is many model calls

A chat turn is one call. An agent that plans, retrieves, uses a tool, checks itself and retries is a dozen — from the same click.

You pay per token, not per request

Priced separately, per million tokens, and output is usually the dearer side — a short prompt with a long answer can cost more than the reverse.

Context is charged every time

System prompts and retrieved context are re-sent on every call, so a prompt you wrote once is a cost you pay on every request.

The rate is a property of the model

Identical token counts cost different amounts on different models — across one family the range can exceed ten times.

Small per-token differences compound

A fraction of a cent per call is invisible in testing and is the whole bill at volume.

So the per-call number is the one to watch — it is the figure that multiplies. Astrodyne prices per model and shows what each call cost.

Estimate a workload → What drives AI API cost →

The catalog

The launch set, tier by tier.

Every model on the launch plan. Each becomes callable when its price is approved — the catalogue in your console is always the list your key can actually reach.

Not yet available 11

On the launch plan, not yet callable. They appear in your console once they are.

GPT-5.6 Sol · OpenAI GPT-5.6 Terra · OpenAI Claude Opus 5 · Anthropic Grok 4.5 · xAI Claude Sonnet 5 · Anthropic GPT-5.1 · OpenAI Grok 4.3 · xAI GPT-5.6 Luna · OpenAI Claude Haiku 4.5 · Anthropic GPT-4.1 mini · OpenAI GPT-5.3 Codex · OpenAI

Sending requests is invite-only while we roll out. Astrodyne runs the model you name — it does not choose a different one for you.

FAQ

Common questions.

The product

What is Astrodyne?

One OpenAI-compatible API for leading models, with keys, a Playground, a per-account model catalogue, a prepaid balance and per-request cost.

Is the API OpenAI-compatible?

Yes. Point an OpenAI-compatible client at the Astrodyne base URL with your key. The compatibility guide lists supported fields and known differences.

Can I use Astrodyne for AI agents?

Yes — it is the case the pricing and visibility are sharpest for. Your agent runs in your application; Astrodyne is the model API it calls, and every call gets its own request ID, tokens and charge.

Does Astrodyne replace my agent framework?

No. Astrodyne does not plan, retrieve, call tools, hold memory or orchestrate anything — it is the model API underneath whatever does. If your framework speaks the OpenAI API, point it at Astrodyne and keep it.

Which models can I use?

The ones available to your account, listed in your console and at GET /v1/models. What you see there is what your key can call.

Does Astrodyne pick a model for me?

No. You name the model and Astrodyne runs that model. It never substitutes a different one.

Is streaming supported?

Yes, over server-sent events, through the same Chat Completions endpoint.

Can I test in the browser?

Yes. The Playground sends real requests through the same path your code uses, with a key held in memory only.

Cost

How does Astrodyne reduce AI API costs?

Astrodyne aggregates demand and maintains the provider relationships behind one API. That lets eligible accounts access commercial terms normally associated with larger buyers, without negotiating each provider relationship themselves. Availability and pricing advantages vary by model.

Is Astrodyne cheaper than buying directly from model providers?

Only where Astrodyne can show it. A price comparison is published for a model when there is both an approved Astrodyne price and a current published direct rate to compare against; where that is not available, the pricing page says so rather than implying a saving.

How does Astrodyne help control agent API costs?

Every call carries its own charge, so you can see which steps of an agent run are expensive rather than inferring it from a monthly total. Spend cannot exceed the balance you funded, so a runaway loop stops at the limit. And where a model has an approved Astrodyne price below the vendor's published direct rate, both figures are shown side by side.

How do I calculate LLM token costs?

(input tokens ÷ 1,000,000 × input rate) + (output tokens ÷ 1,000,000 × output rate), multiplied by the number of requests. The LLM cost calculator does this for you.

How can I reduce OpenAI API costs?

Send fewer input tokens by trimming system prompts and retrieved context, cap output length where the task allows, and use a smaller model for work that does not need a larger one. Per-request cost visibility shows which calls actually dominate the bill.

How can I reduce Claude API costs?

The same levers apply: less input context, capped output, and the smallest model that does the job. Because input and output are priced separately, a long answer can cost more than a long prompt.

Can I estimate my AI API spend before switching?

Yes. Enter your typical input and output token counts and expected request volume in the calculator. Where a verified price and a published direct rate both exist for a model, the two are shown side by side.

Gateways and alternatives

What is an AI gateway?

A single API in front of multiple model providers, handling authentication, request routing, billing and usage visibility so you do not integrate each provider separately.

What is an LLM gateway?

The same idea, usually said of text models specifically. The terms are used interchangeably and there is no formal distinction.

Can I use OpenAI, Claude, Gemini and other models through one API?

That is the purpose of the gateway: one OpenAI-compatible surface across model families. Which specific models your key can call is account-specific and listed in your console and at GET /v1/models.

How does Astrodyne compare with OpenRouter?

Both are hosted gateways offering many models through one API. Compare them on deployment, billing model, whether model selection is explicit or automatic, and whether you can see the cost of an individual request. We do not publish claims about another product's current behaviour.

How does Astrodyne compare with LiteLLM?

The structural difference is deployment and billing: Astrodyne is hosted and holds one prepaid balance, where a self-hosted proxy is run by you against your own provider accounts and invoices.

Pricing and access

How does billing work?

You top up a prepaid balance in US dollars and are charged for settled usage at the price shown for that model. Every request's charge is visible on its own.

How do the lower prices work?

Eligible models are priced below the vendor's published direct rate. Eligibility is per model, and we publish a comparison only when there is an approved price and a current published rate to compare against.

Do you say which vendors run the models?

No. Customer-facing surfaces show a neutral catalogue and sanitized request data.

Who is it for?

Developers, small teams and agencies who want one integration and one bill instead of one of each per vendor.

Can I sign up today?

You can create an account and explore the console. Sending requests is invite-only while we roll out.

Your next step

Your agents will make a lot of calls.
Make the model layer cost less.

Build your AI product or agent against one OpenAI-compatible API, choose the models each workload needs, and see the cost of every request. Sending requests is invite-only while we roll out.

Get API Key Read the Quickstart