Every model.
One API key.
Access 200+ models from OpenAI, Anthropic, Google, Meta and more through one API. One key, one bill, one place to manage it all. Served from Singapore.
OpenAI
Anthropic
Google
Meta
Mistral
DeepSeek
Qwen
Kimi
xAIFast, reliable,
and close to home
One integration that stays quick, stays up, and handles every kind of request.
Low latency
The gateway runs in Singapore, so responses come back fast for users in Singapore.
Automatic fallback
If a provider drops, requests move to a healthy route for the same model, instantly.
A gateway in every region
If your users are elsewhere too, every call goes to the gateway nearest them.
Every modality
Text, image, speech, transcription and embeddings, all through one connection.
Give your AI assistant
200+ models
to work with
Not every job needs your best model. Labelling thousands of rows, summarising a folder of notes, pulling fields out of invoices: your assistant can hand the work to a fast, cheap model through Roar AI and bring back the results, on the same account and the same bill.
Running a job needs an assistant that can run code, such as Claude Code, Cursor or Codex.
Picks the right model
Your assistant sees every model on your account, with what it can do and what it costs, and picks a fast, cheap one for bulk work.
A key with a cap
It makes a key that can only call models, with a spending cap, so a long job cannot run away with your budget.
One API for all of them
Every model answers on the same OpenAI-compatible API, so trying another one is a one-line change.
Every model’s
own features
Thinking, tools, JSON, files, search and caching. Each model’s strengths work through the same API, so switching models never means rewriting your code.
Reasoning
Turn a model’s thinking up or down with one setting, and read its reasoning where the model shares it.
Tool calling
Let a model call your functions, several at once if it needs to, in the same format on every model.
Structured outputs
Ask for JSON that matches your schema and get it back in exactly that shape, on the models that support it.
Images and PDFs
Send a screenshot, a photo or a PDF along with the prompt, to the models that can read them.
Web search with sources
Let Claude or GPT look things up as it answers, with citations back to where each fact came from.
Prompt caching
A long prompt you send again is billed at the cheaper cached rate, automatically. Set your own cache points when you want control.
Priority
Faster answers for a premium when someone is waiting, on the models that offer it. Charged only when it is served.
Batch
Send thousands of requests as one file and collect the results within 24 hours, for about half the usual price.
Not every model does everything. The model pages list what each one supports.
Every rate,
published up front
One price per model, shown before you call it. It does not move with what a provider charges us, so the bill you expect is the bill you get.
AnthropicOpenAIGoogleDeepSeekGive it to your team.
Keep the reins.
What leaves, what it costs, and who can do what.
Nothing sensitive leaves by accident
Personal data and secrets can be caught before a call reaches an outside model: redacted, blocked, or sent to a private one.
A ceiling you set
Caps per key, project or account, counted from real usage. The tightest one wins.
Keys that only do one job
Scope every key, so a prototype cannot reach what production can.
Your call history, readable
See what was sent and what came back, encrypted at rest. Turn it off for your organisation if you would rather nothing was kept.
More than chat,
on the same key
The same calls you already write, on every model we carry.
Chat completions
/v1/chat/completionsEmbeddings
/v1/embeddingsImages
/v1/images/generationsSpeech
/v1/audio/speechTranscription
/v1/audio/transcriptionsBatches
/v1/batchesFiles
/v1/filesModels
/v1/modelsQuestions about the AI Gateway
What is an AI gateway?
One address and one key that reach models from many AI companies. Instead of an account, a card and a contract with each of them, you send every request to the gateway and it passes it to the model you asked for. Ours also holds one route per model, so a model behaves the same way every call.
How is Roar AI different from other LLM gateways?
Most LLM gateways pick who serves each request on price, per call, so the same model can accept a setting on one call and reject it on the next. We pick one route per model and hold it, with others behind it for when one fails, so a model behaves the same way every time.
Where does the gateway run?
In Singapore, so calls from Singapore stay close to home. There are gateways in Colombo and Bangalore as well.
Do I need to change my code?
If you use the OpenAI SDK, two lines: the address and the key. Every model we carry answers on the same API, for chat, embeddings, images, speech and batch.
Which models can I use?
200+ from OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek and more, including a few open-source models that run on our own hardware in Colombo.
What happens if a provider goes down?
The request moves to a healthy route for the same model, automatically. It never quietly switches you to a different model.
How do I pay for it?
One account and one bill for every model. Each model’s rate is published on the pricing page, and it does not move with what a provider charges us.
Change two lines
and try it
Sign up and get a key in a minute. Bring the code you already have.