Juga LabsAll systems operational
Routing, key issuance and billing are running normally across every protocol.
Target, measured at the endpoint
Streaming chat, mid size models
Short completions, excluding model think time
5xx raised by routing, not by upstream refusals
Every surface, checked separately.
- Chat completionsOpenAI compatible, streaming and tools
/v1/chat/completionsOperational - Responses APIFor newer OpenAI SDKs and agent frameworks
/v1/responsesOperational - Anthropic nativeClaude Code and the Anthropic SDK
/v1/messagesOperational - Gemini nativeGemini CLI and google-genai
/v1betaOperational - Image generationText to image and edits
/v1/imagesOperational - Video generationAsync jobs, polled by job id
/v1/videoOperational - Speech synthesisVoice library and cloning
/v1/ttsOperational - Accounts and keysTop ups, key issuance, credit limits
ConsoleOperational
What the models actually feel like.
Time to first token is what you notice in a chat or an agent loop. Throughput is what you notice on long generations.
| Model | Context | Time to first token | Output speed | Streaming |
|---|---|---|---|---|
| Claude Opus 5 | 1M | 0.62s | 52 tok/s | SSE |
| Claude Sonnet 5 | 1M | 0.38s | 88 tok/s | SSE |
| Claude Haiku 4.5 | 200K | 0.22s | 141 tok/s | SSE |
| GPT 6 Astra | 1.05M | 0.71s | 61 tok/s | SSE |
| GPT 5.6 Sol | 1.05M | 0.44s | 96 tok/s | SSE |
| Gemini 3.1 Pro | 1M | 0.49s | 104 tok/s | SSE |
| Kimi K3 | 1M | 0.58s | 74 tok/s | SSE |
| DeepSeek V4 Pro | 1M | 0.35s | 119 tok/s | SSE |
Measured at the gateway on short prompts, excluding the time a reasoning model spends thinking. Your numbers move with prompt size, cache hits and how far away you are from the upstream region.
How it behaves under pressure.
Streaming
Server sent events on every chat model, with no proxy buffering in front of the stream.
Concurrency
Shared capacity absorbs normal bursts. When a pool is saturated the gateway returns 503 service_busy rather than queueing your request silently.
Retries
Transient upstream failures are retried inside the gateway before the error reaches your code.
Cached input
Cache reads bill at a tenth of the input rate, which is where agent and coding workloads win most.
Failures and credit
Video and voice jobs that fail are refunded in full. A call that returns nothing is not billed.
Incidents
Anything affecting routing is posted here and sent to funded accounts by email.
Seeing something we are not?
If your calls are failing or slow, tell us with the request id from the error response and we will trace it. Funded accounts get incident notices by email.
Or email juga@jugalabs.com