inference gateway · 12 regions · SOC 2 Type II

Every frontier model,
one sane endpoint.

Point your OpenAI-compatible client at https://lumenal-ai.duckdns.org/v1 and get routing across 38 hosted and open-weight models, live-latency failover, exact-match caching and per-key budgets. No proxy spaghetti.

# same request shape you already send to OpenAI
curl https://lumenal-ai.duckdns.org/v1/chat/completions \
  -H "Authorization: Bearer $LUMENAL_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "lum-route:cheap",
    "messages": [{"role": "user",
      "content": "Summarize this changelog"}]
  }'
99.97%gateway uptime, 90 days
118 msmedian routing overhead
38hosted + open-weight models
301Btokens served / week

Why teams route through Lumenal

Latency-aware routing

lum-route:fast|cheap|smart picks an upstream from live error-rate and p50 metrics, re-scored every 15 seconds per region.

Failover chains

Upstream 503 or rate-limit? The next model in your chain answers with an x-lumenal-fallback header. Your code never retries.

Exact-match cache

Identical prompts (temperature 0) served from cache at zero cost. Cache hit-rate is shown per key in the console.

BYOK passthrough

Bring your own vendor key and pay the vendor directly — Lumenal takes 0% on BYOK traffic, routing included.

Open-weights mirror

SHA-verified mirrors of Llama, Qwen and DeepSeek weights. Run them locally; burst overflow rides the gateway.

Data residency

Pin a key to EU or US processing. Zero-retention agreements with every upstream; request archive is opt-in per key.

Popular right now

lum-route:smartLumenal Router$0.09 in · $0.36 out / 1M
lumen-core-70bLumenal · proprietary$0.35 in · $1.40 out / 1M
llama-4.2-90bMeta · open weights$0.12 in · $0.48 out / 1M
deepseek-r2DeepSeek · open weights$0.27 in · $1.10 out / 1M
gpt-5OpenAI$1.25 in · $10.00 out / 1M
claude-sonnet-4.5Anthropic$3.00 in · $15.00 out / 1M

Full catalogue — 38 models →

Questions we actually get asked

Is it really drop-in compatible with the OpenAI API?

Base URL and bearer token are the only changes. /v1/chat/completions, /v1/embeddings, /v1/models, SSE streaming and tool-calling follow OpenAI semantics; vendor extras pass through in extra_body.

Do you train on our prompts?

No. Prompts go to the selected upstream under zero-retention terms and are logged as metadata only (24 h) unless you enable request archiving on the key.

What happens when an upstream degrades?

Your key keeps working: the router moves to the next entry in your failover chain and stamps the response with x-lumenal-fallback. Degradations appear on the status page within 60 seconds.

Can we self-host part of the traffic?

Yes — open-weight catalogue entries ship with checksummed weight mirrors. Run locally, set "burst": "gateway" and only overflow is billed.