Build against one endpoint
Helyx AI speaks the OpenAI Chat Completions protocol. Point any existing OpenAI client at https://helyxai.space/v1, swap the key, change the model string — done. There is no SDK of ours to install.
Quickstart
Create an account, grab a key from the dashboard, and make a call. Every claimable model's daily grant is taken automatically the moment you sign in.
Authentication
Every request carries a bearer token. Keys are created and revoked on the dashboard. We store only a SHA-256 hash, so a lost key cannot be recovered — revoke it and create a new one.
A missing or malformed header returns 401. Keys are account-wide: any key can call any active model.
Chat completions
Request body
Response
Streaming
Set "stream": true to receive server-sent events. Chunks arrive as data: lines and the stream ends with data: [DONE]. Usage totals arrive in the final chunk, so billing stays accurate on streamed calls.
Token accounting
You are charged for what you actually send and receive.
- Gateway preamble is not billed. Upstream providers prepend a fixed system preamble — roughly 7,100 tokens on Claude Opus 5. We subtract it before anything is logged or charged. A one-word prompt therefore records 0 input tokens, not 7,000.
- Only the remainder counts as input. If the provider reports 15,037 prompt tokens, your logged input is 15,037 − 7,100 = 7,937.
- Output is counted as reported. No adjustment.
- Free tokens are generic. A grant of 1M covers 1M combined input + output tokens.
Free-token waterfall
There are no plan tiers and no per-model daily caps. Every request draws from three sources in order:
The daily claim requires a signed-in session: visiting the dashboard claims it on every eligible model. The API never self-grants tokens, so a key alone cannot mint free usage. Unused daily tokens do not roll over — each day starts fresh and yesterday's remainder is gone. The claim per model is listed in the table below.
When a single request spans two sources, cost is split proportionally: if 30% of the tokens fall past your free tokens, you pay 30% of that request's price.
Retries & overload
Every model call is attempted up to three times against the upstream provider before we give up. Transient failures — timeouts, 5xx, rate limits, dropped connections — are retried transparently with a short backoff, so most blips never reach your code.
If all three attempts fail, the API returns 503:
Failed requests are never billed and never consume free tokens. If 503s persist, message us on WhatsApp or Telegram @kairalmas.
Data logging
We are currently retaining request and response content. While the platform is being tuned across all models, both the prompts you send and the completions we return are written to internal log files.
Specifically, for each API call we store the request payload, the flattened prompt text, the model's full output, the model slug, token counts, HTTP status, retry count, and latency. Nothing is shared with third parties beyond the upstream model provider that serves your request.
This capture is temporary and exists so we can compare model behaviour, reproduce failures, and verify billing. It will be switched off once that work is finished. If you are sending data you would rather not have retained, avoid it during this period or contact us to have your account excluded.
Logs are stored outside the web root and are not reachable over HTTP.
Error codes
Model catalog
5 models are active right now. Rates are per 1M tokens; a struck-through price means a discount is live and is applied automatically at request time.
| Slug | Model | Context | In / 1M | Out / 1M | Daily claim |
|---|---|---|---|---|---|
| DeepSeek-V4-Flash | DeepSeek-V4 Flash DeepSeek | 1M | $0.140 | $0.280 | — |
| DeepSeek-V4-Pro | DeepSeek-V4 Pro 50% off DeepSeek | 1M | $0.800 $0.400 | $1.50 $0.750 | — |
| gemini-3.1-flash-lite | Gemini 3.1 Flash Lite Google | 1M | $0.250 | $1.50 | — |
| gemma-4-31B-it | Gemma4 31B Google | 262K | $0.200 | $0.500 | — |
| MiniMax-M3 | MiniMax M3 MiniMax | 1M | $0.300 | $1.20 | — |