One API key for every AI model
Point your app or coding tool at Zumik instead of OpenAI. Every request goes to whichever model can do that job well for the least money — across OpenAI, Anthropic, Google, xAI and more — on one key and one bill. Your code does not change.
Create an account
Email or Google, GitHub, Microsoft. About a minute, no card needed yet.
Add $5 in credits
Pay-as-you-go from there. No subscription, and you can cap your spend.
Change the URL
Point your app or coding tool at Zumik and use your Zumik key. Done.
Why it costs less
Two kinds of waste, both invisible on your provider bill.
Nothing here is a trick or a discount we negotiated. It is money most teams spend without meaning to, and it shows up on every provider invoice the same way.
Paying a top model to do easy work
Most apps pick one strong model and send everything to it — the hard reasoning and the one-line classification alike. Zumik reads each request and sends it to the cheapest model that will still get it right, and records which model answered and why.
How the choice is madePaying twice for the same context
Agents re-send the same system instructions, tool definitions and history on every single call. Providers will charge you full price for that again unless the request is arranged so their cache can match it. Zumik arranges it, and measures how much of the repeat actually got reused.
See where the money goesWho this is for
If any of these sound like you, Zumik pays for itself.
You run a coding agent
Cline, Roo Code, Continue, Aider and anything else with an OpenAI-compatible setting. Long sessions re-send enormous context - that is exactly the bill Zumik cuts.
Read moreYou are building an AI feature
One integration instead of an account, key and rate limit per provider. Swap models without shipping code, and see per-request cost from day one.
Read moreYour inference bill is growing
Find out how much of it is repeated work before you rewrite anything. The check is free and needs no payment method.
Read moreLook before you commit
Prices and speeds we measured ourselves
Every page below is built from our own traces, not a vendor deck. Check the numbers before you sign up for anything.
All models and prices
Every model Zumik can route, with what it costs and how fast it answers.
OpenCost calculator
Enter your usage, see what you would pay.
OpenBest model for the job
Ranked by cost, speed, and how well they hold up in long agent runs.
OpenCompare two models
Side by side on price, speed, and context length.
OpenSpeed and caching benchmarks
Numbers we measured ourselves, not vendor claims.
OpenScore your workload
Answer six questions and see how much repeat spend you could recover.
OpenPrompt layout checker
Spot the ordering mistakes that quietly make you pay full price twice.
OpenUsage trends
What we see across the agent workloads we have measured.
OpenFor developers
A drop-in API, and a deeper one when you want it.
/v1 — the drop-in one
Start hereChat Completions, Responses, Embeddings and Models, byte-for-byte OpenAI-shaped, with streaming relayed live from the provider. Every response tells you what routing saved you in the Agent-Savings-Micros header.
r = client.chat.completions.create(
model="auto", # Zumik picks the model
messages=[{"role": "user",
"content": "Why did CI fail on main?"}],
)/v2 — the deeper one
OptionalStore the context you re-send every call once, then refer to it by id. Branch a conversation, replay a past run, or prove a deletion. Nobody needs this on day one.
curl https://api.zumik.ai/v2/artifacts \
-H "Authorization: Bearer zk_live_..." \
-d '{"artifact_type":"policy",
"content":"Run the linter before commit."}'
# => { "id": "art_01JY…" }Everything else
The rest of what Zumik does.
How Zumik works
The path a request takes, start to finish, in plain English.
Routing across providers
One key for OpenAI, Anthropic, Google, xAI and Fireworks, with automatic failover.
Stop paying twice for context
Store the instructions and context you re-send every call, and pay for them once.
Find your wasted spend
See how much of your bill is work you already paid for.
Run a free check
A free scan of your workload. No payment method needed.
Model names that never drift
Ask for auto and get a recorded, repeatable choice every time.
Delete with proof
Remove your data and get a receipt showing exactly what went.
Run it on your own hardware
Only when the numbers show self-hosting genuinely beats the providers.
Zumik is not a chatbot, a model of its own, or a GPU cloud. It is the layer that decides where your requests go and stops you paying for the same work twice. When the model you are already using is the right answer, it will tell you to stay put.
Frequently asked
Start here if you are not sure what this is.
What is Zumik?
One API key that works with every major AI model. Instead of holding an account with OpenAI, Anthropic, Google and the rest, you send your requests to Zumik and it forwards each one to whichever model can do that job for the least money. You get one key, one bill, and one place to see what you spent.
What do I have to change to use it?
Two settings: the URL your code or coding tool sends requests to, and the API key. Zumik accepts the same requests OpenAI does and returns the same responses, so nothing else in your app changes. Most people are done in under five minutes.
Do I need to know which model to pick?
No. Use the model name "auto" and Zumik chooses per request - a cheap fast model for easy work, a stronger one when the request needs it. You can always name a specific model instead.
How much does it cost?
You pay per request out of prepaid credits, starting at $5. No subscription, no seats, no monthly minimum. Each request costs the published price of the model that answered, and on managed routing that lands under what you would pay the provider directly.
Why is it cheaper than going direct?
Two reasons. Zumik picks a cheaper model when a cheaper model is good enough, and it stops you paying full price twice for the instructions and context your agent re-sends on every single call. Across the agent workloads we have measured, about half of a typical request ends up billed at the cheaper cached rate.
Which providers can it reach?
OpenAI, Anthropic, xAI, Google Gemini and Fireworks AI directly, with wider coverage behind them. If one has an outage, Zumik fails over to another model that can do the same job rather than returning an error.
Is my data used for training?
No. Prompts are passed to the provider answering the request and nothing else. You can delete stored context whenever you like and get back a receipt showing exactly what was removed.
Who is this not for?
If you want something to chat with in a browser, Zumik is not that - it is plumbing for software you are building. And if you send a handful of requests a month, the savings will not be worth the five minutes of setup.
Five minutes and five dollars
Create an account, add credits, change one setting. If Zumik does not save you money you have lost an afternoon coffee.