Change the base URL. Keep the SDK, the code and the response shape.
One endpoint, every provider, one bill
Point it at a different base URL and the model name picks the provider. No adapter per vendor, no second SDK, no key of your own scattered across four services — and a per-request log that says what each call actually cost.
Your first request
curl https://ai-gateway.piskes.com/v1/chat/completions \
-H "Authorization: Bearer $PISKES_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-4o-mini",
"messages": [{"role": "user", "content": "Say hello in five words."}]
}'The shape of it
Everything else an AI gateway sells is a dashboard you open once.
Chat completions, embeddings and the model list, in the request and response shapes the official SDKs already send and parse. Streaming is server-sent events terminated by [DONE], the same as you handle today. The integration is a base URL and a key, which is the whole reason a gateway is worth having rather than a per-vendor adapter you maintain.
Context window, capabilities and per-million input and output price for every model, readable before you have an account.
Model, provider, tokens both ways, cost, latency and status. Charted by day and by model, and not a rollup — the row is kept.
Provider credentials live in one service. Your key spends credit and cannot mint itself a replacement or read your billing.
128K context. OpenAI’s newest, same request body as the rest.
02128K context, images in, tools out.
03The cheap one worth defaulting to while you are still measuring.
0416K context, for the classification job that does not need more.
05Runs on Cloudflare’s own network. No per-token charge at all.
06Embeddings through the same endpoint and the same key.
Bring your own key
Store your own provider key and the request goes out on it. You pay your provider directly and this gateway charges 5% of what the call would have cost on credits — the routing, the log and the one API, without buying the tokens twice.
Against the two things you would otherwise do
| A client per provider | A self-hosted proxy | Piskes AI Gateway | |
|---|---|---|---|
| Adding a fifth model | A new SDK and a new adapter | A config change and a deploy | A different string |
| Where the provider keys live | In four services | In your secret store | In one service, never a browser |
| Per-request cost visibility | Four invoices, monthly | Whatever you built | A row per call, charted |
| Who is on call for it | You | You | Nobody on your side |
| Using a contract you already have | Yes | Yes | Yes, at 5% |
| Per-seat charge | n/a | n/a | None |
Pricing
Per-seat pricing on an API is a charge for having colleagues.
Most teams
Pay as you go against the model’s own per-token price plus a platform markup.
$0.01 per credit
For a team that already has a provider contract worth keeping.
5% of what the call would have cost
Prices in USD. Credit is bought in advance and spent per request; nothing renews and nothing expires on a monthly cycle. The exact per-model numbers live on gateway.piskes.com/models — this page does not restate them, so the two can never disagree.
Questions
A base URL and a key. The request bodies, the response shapes and the streaming format are the ones your client already sends and parses, which is what "OpenAI-compatible" has to mean to be worth anything.
The request is refused with a specific code before it reaches a provider, so you get an error you can branch on rather than a silent failure or an unexpected invoice.
Whatever the gateway currently holds a key for, listed with prices on a public page that is generated from the same registry the router uses. It is not a brochure list.
Yes. Every completion writes a row with its tokens, its cost, its latency and its status, and the dashboard charts them by day and by model.
No, and saying so is cheaper than an asterisk. The smallest top-up is $5, which is a lot of small completions.
That is the whole evaluation, and it takes about as long as reading this page. The model list and the docs are open without an account; the key takes a minute after you have one — same Piskes sign-in as every other app on the platform.