One API key.
Every leading LLM.

CI-AI Gateway gives you a single OpenAI-compatible endpoint in front of GPT-5, Claude, Gemini, DeepSeek, Qwen, Llama and Kimi. Smart routing, automatic fallbacks, response caching, BYOK and open-weight models you can run on your own GPU.

Create an API key Open the demo chat

OpenAI-compatible /v1/chat/completions · streaming · function calling · GET /v1/models

curl https://ci-ai.duckdns.org/v1/chat/completions \
  -H "Authorization: Bearer $CIAI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-flash-4.1",
    "messages": [
      {"role": "user",
       "content": "Summarize this changelog"}
    ],
    "stream": true
  }'
99.95%gateway uptime, last 90 days
140 msmedian routing overhead
41models from 9 vendors + open weights
428Btokens routed per week

What people build with it

codingreasoningagents & toolstranslationRP & personaslong contextmultimodalopen weightsself-hostEU residency

Popular models Full catalogue — 41 models →

gemini-3-flash

Google

$0.10 in · $0.40 out

  • 1M ctx
  • multimodal
  • 240 tok/s

qwen-flash-next-3.8

Alibaba · open weights

$0.06 in · $0.24 out

  • runs locally
  • MoE
  • 256k ctx

deepseek-flash-4.1

DeepSeek · open weights

$0.08 in · $0.32 out

  • runs locally
  • tools
  • MoE

gpt-5

OpenAI

$1.25 in · $10.00 out

  • vision
  • reasoning
  • 256k ctx

claude-sonnet-4.5

Anthropic

$3.00 in · $15.00 out

  • tools
  • 200k ctx
  • long tasks

qwen3.8-27b

Alibaba · open weights

$0.10 in · $0.40 out

  • runs on 24 GB GPU
  • 128k ctx
  • tools

Top models by weekly tokens

gemini-3-flash83.6B
qwen-flash-next-3.8 open71.3B
gpt-5-mini61.4B
deepseek-flash-4.1 open58.9B
gpt-548.2B

Full breakdown on the catalogue page.

How it works

1. Create a key

30 seconds in the panel, no card required: 100k trial tokens are already on the account. Scope each key by models, budget and region.

2. Change nothing

Base URL → https://ci-ai.duckdns.org/v1. Any OpenAI-compatible code, SDK or framework keeps working as-is.

3. Let the router decide

model: "auto" or the :fast / :cheap / :eu suffixes — the gateway picks the upstream from live metrics and keeps a fallback chain ready.

Built for production workloads

  • Smart routing — pick a family by quality, cost or latency; the gateway scores every upstream.
  • Fallback chains — if an upstream 503s or rate-limits, the next model answers; your code sees one stream.
  • Response cache — identical prompts served from cache for free (opt-in, per key).
  • BYOK — bring your own vendor key: you pay the vendor directly, zero gateway fee.
  • Open weights + local run — qwen, deepseek-flash, devstral and llama weights mirrored with stable checksums; compute at home, burst through the gateway.
  • EU / US points of presence — pin traffic to a region for data residency.
Inside the CI-AI edge datacenter
The eu-central edge PoP, where routing and caching run closest to the open-weight fleet.

FAQ

Is it really drop-in compatible with OpenAI?

Yes — base URL and auth header change, nothing else. /v1/chat/completions, /v1/embeddings, /v1/models, streaming SSE and tool calls follow the OpenAI spec; vendor-native extras are passed through under extra_body.

Do you train on my prompts?

Never. Prompts are forwarded to the selected upstream under zero-retention agreements and are not logged beyond 24-hour request metadata unless you enable request archiving per key.

What happens if a vendor is down?

Your key keeps working: the router moves to the next model in your fallback chain and stamps the response with an x-gateway-fallback header.

Can I compute on my own hardware?

Yes: open-weight catalogue models are mirrored in our repo and run locally (24 GB VRAM at 4-bit). Enable "burst": "gateway" and overflow peaks are served at catalogue prices.

Get started in 30 seconds

100k trial tokens, no card. One key — forty-one models.

Create a key Browse the catalogue