One request format
Send text, vision, audio, and video through the same interface.
Platform
Ship the product.We'll handle the route.
One request for chat, image, audio, and video. If the first model fails, the next one answers. You can see which.
01 / Unified API
The request stays. The model can change.
The fields stay the same when you change providers. You are not writing a new client for each of them.
- Same body for every call
- The reply keeps the same fields
- Text, image, audio, video, embeddings
- The provider stays out of your client
Same request
const response = await client.chat({
model: "gpt-4.1",
messages: [{ role: "user" }]
});
Serving OpenAI. Route 1 of 3.
Model
gpt-4.1
Latency
180ms
Tokens
816
02 / Smart Routing
Send it where the job belongs.
Cheaper, faster, or a model you named. The call your app makes does not change.
- Cheaper, faster, or a named model
- Only the models you allowed
- The request does not change
Policy
Selected route: GPT-4.1 mini
Claude Sonnet 4380ms
- Cost
- ₹22.00
- Availability
- 99.2%
- Capability
- High
GPT-4.1 mini190ms
- Cost
- ₹7.40
- Availability
- 99.9%
- Capability
- Medium
Gemini 2.5 Flash240ms
- Cost
- ₹4.10
- Availability
- 99.4%
- Capability
- Medium
03 / Automatic Fallbacks
If one drops, the next answers.
A timeout or a provider error moves the call down the list you set. The caller still gets one reply.
- A list, in the order you set
- Timeouts and provider errors
- One reply, one request id
- The caller does not retry
- You see which model stepped in
- It stops at the first one that answers
A request stays healthy, is sent to Claude Sonnet 4, times out, triggers a fallback, is accepted by GPT-4.1 mini, and is delivered as one response.
04 / Observability
See which model answered.
Latency, errors, and which provider served the call. One place, not a tab for every vendor.
Observability
Demonstration data
Requests
12,480
+3.4% vs prior
Success rate
99.2%
−0.1 pt
Latency
214ms
+11ms vs prior
Estimated spend
₹15,624
This window
Request volume
24 hours
00:0004:0008:0012:0016:0020:0024:00
Providers
- OpenAI38%
- Anthropic27%
- Google21%
- Other14%
Models
- GPT-4.1 mini34%
- Claude Sonnet 429%
- Gemini 2.5 Flash22%
- Other15%
Recent fallback
12:01
Claude Sonnet 4 timed out.
Served by GPT-4.1 mini
- Latency and success rate
- Spend by model
- Fallback events
- Provider breakdown
05 / Enterprise Controls
Tell each team which models they may use.
Workspaces, regions, and a cap on usage. When someone changes the policy, that change is kept.
- Model access by workspace
- Region pins
- Usage limits
- Audit of policy changes
- Production and staging keep separate rules
- A model stays off until you allow it
- Fallback can be off for a workspace
- The cap is a monthly amount
Workspace policy
Frontend preview. Nothing here is saved.
Workspace
Region
Monthly usage limit
What your app no longer has to do.
Policy routing
Choose a model by cost, latency, capability, or availability.
Provider fallback
Move a request to the next healthy model when one fails.
Usage analytics
See latency, success rate, and spend for every routed request.
Workspace controls
Limit models, regions, and usage from one policy layer.
Structured output
Keep one response shape when the provider behind the route changes.
Questions
Questions, plainly.
The gateway, the route, and what happens when a model does not answer.
General
Routing
Pricing
Support
Build the product.Leave the route to us.
The models are listed. The request is the same for each of them.