One key.
Every model.
Point your existing SDK at one endpoint and reach Claude, GPT, Gemini, Grok, Kimi and DeepSeek. Same OpenAI and Anthropic wire formats, same streaming — at a fraction of list price.
- 42
- models
- 6
- providers
- 91%
- off list price
- 1
- API key
The whole pitch, as arithmetic
The same model. A fraction of the bill.
Every row is pulled live from the running price catalogue — not a screenshot. Output price per million tokens, which is what actually dominates an inference bill.
| Model | Official | Here | You save |
|---|---|---|---|
 Claude Opus 5Claude | $25.00 | $2.86$20 credits | 89% |
| Claude Sonnet 5Claude | $10.00 | $1.14$8 credits | 89% |
| GPT-5.2OpenAI | $10.00 | $1.29$9 credits | 87% |
| GPT-4oOpenAI | $10.00 | $1.29$9 credits | 87% |
 Gemini 3 ProGemini | $12.00 | $1.20$8.4 credits | 90% |
 DeepSeek ReasonerDeepSeek | $2.19 | $0.19$1.31 credits | 91% |
 Kimi K2Kimi | $2.50 | $0.21$1.5 credits | 91% |
 Grok 4Grok | $15.00 | $1.29$9 credits | 91% |
If you spend this on official APIs each month
$200
the same traffic here costs
$22
At the 89% average discount across the models above. Your real figure depends on the tier you pick and your cache-hit rate.
Billing runs in credits — one credit is one unit of what you top up. The dollar figures are those credits converted back at the reference rate, so both columns compare like for like.
Two lines to migrate
Nothing to rewrite.
The gateway speaks both the OpenAI and Anthropic wire formats and translates between them — requests, responses and SSE streams. An OpenAI SDK can call a Claude model; Claude Code can call GPT. Change the base URL and the key, keep everything else.
- Streaming passes straight through, token by token
- Tool calls and vision payloads translated both ways
- Automatic failover to the next upstream on 429 or 5xx
from openai import OpenAI
client = OpenAI(
api_key="OMNI_KEY",
base_url="https://api.omnikeyai.live/v1",
)
client.chat.completions.create(
model="claude-opus-5", # or gpt-5.2, gemini-3-pro, grok-4…
messages=[{"role": "user", "content": "hi"}],
)Built like infrastructure, not a reseller script.
Priority routing with failover
Channels are ranked by priority, then weighted-random inside a band. A 429 or 5xx rolls to the next one mid-request.
Metered on real usage
Token counts come from the upstream’s own usage payload. Cache reads and writes are priced separately, never guessed.
Limits that actually hold
Per-key concurrency slots and RPM windows in Redis, plus a per-day ceiling and a model allow-list on the free tier.
Keys you can scope
Pin a key to a pricing tier, a quota, an expiry, a model list or an IP range. Rotate the secret without losing the key.
Start at zero
Three ways to run without paying anything.
$100 on sign-up
Credited the moment the account exists. Spend it on any model, any tier — it never expires.
Create an account$10 per referral
Paid straight into your balance when someone signs up with your code, plus 5% of everything they ever top up.
Get your linkCredits for social tasks
Follow, share, subscribe or write a review. Small tasks pay out on submit; larger ones after a quick review.
See the tasksYour next request could be running in a minute.
No card, no sales call. Create a key and point your SDK at it.