Nicrron gateway
One API for every model.
Route, cache and observe your AI traffic through one OpenAI-compatible endpoint. Pay per token, no subscription.
$10 to start · deposits matched up to $500 every 90 days · chat prompts never stored
import OpenAI from 'openai';
const client = new OpenAI({
apiKey: process.env.NICRRON_API_KEY,
baseURL: 'https://api.nicrron.ai/v1',
});
const response = await client.chat.completions.create({
model: 'anthropic/claude-sonnet-5',
messages: [{ role: 'user', content: 'Hello!' }],
});01 — Routing modes
Name the job, not the model.
Four virtual slugs pick the right model per request. Pin a specific model whenever you need to, or hand the gateway an ordered fallback list.
- nicrron/auto weighs prompt length, complexity and live provider health.
- Fallback arrays cascade on 429 or 5xx before the first streamed byte.
- Every request is one OpenAI-compatible call; nothing new to learn.
nicrron/autoRight-sizes the model per request from prompt length, complexity and live provider health.
nicrron/reasoningDeepSeek V4 Pro, GPT-6 Sol and Claude Opus 5.5 for multi-step logic.
nicrron/codingClaude Opus 5.5, Claude Sonnet 5.5 and Kimi K2.7 Code for software work.
nicrron/fastestClaude Haiku 5.5, GPT-6 Luna and DeepSeek Flash for the lowest latency.
02 — Failover and caching
Uptime you don't have to engineer.
Provider health is scored continuously with an EWMA latency model and circuit breakers. Deterministic requests are served straight from Redis, in milliseconds, for free.
- If a provider rate-limits or fails, the request moves to the next one instantly.
- Exact-match cache: 12 ms and $0.00 on a hit, scoped to your account, keyed by the full request.
- Send temperature 0 or x-nicrron-cache: true to opt in per request.
Latency, same request
cached requests are not billed
03 — Zero retention
We meter tokens, never text.
Chat, embedding and audio payloads stream through memory and are never written to disk or a database. Only the metadata needed to bill and audit is kept.
- Stored: latency, status, token counts and a hash of the request.
- Never stored: prompts, completions, embedding vectors, audio.
- Generated video and images stay 30 days for download, then are deleted.
Kept
- latency_ms
- status_code
- prompt_tokens · completion_tokens
- request_hash
Never kept
- messages[]
- choices[].message
- embedding vectors
- audio input
04 — Per-token billing
One balance, every cent visible.
Every request lands in your logs with its exact cost. Upstream prompt-cache discounts pass through automatically, and teams share one wallet with roles and alerts.
- Anthropic cache reads are 90% off, OpenAI cache reads 50% off, recorded in usage.
- Filter logs by date, provider, status and cost; export to CSV or JSON.
- Owner, Admin, Developer and Viewer roles over a shared credit pool.
POST /hooks/nicrron HTTP/1.1
X-Nicrron-Signature: sha256=9b4a1f…
X-Nicrron-Event: alert.low_balance
{
"event": "alert.low_balance",
"currentBalance": 4.50,
"threshold": 10.00
}05 — Also in the gateway
More than chat completions.
- Model Arena
- Benchmark two or three models on one prompt: TTFT, tokens per second and cost, with a blind mode.
- Teams and shared wallets
- One credit pool per organization with Owner, Admin, Developer and Viewer roles.
- Webhook alerts
- HMAC-SHA256 signed payloads to your endpoint when balance runs low or spend spikes.
- Video and images
- Seedance 2.0, Veo 3.1 Fast, Kling 2.5, Hailuo 02 and Nano Banana as async jobs, cost charged upfront and refunded on failure.
- Audio
- OpenAI TTS and Whisper transcription through the standard SDK, billed on true audio duration.
- Embeddings
- OpenAI, Mistral and Voyage embeddings behind /v1/embeddings with automatic failover.
Providers
Every major lab, one key.
Frontier chat, reasoning, coding, embeddings, audio, image and video models behind a single endpoint.
Questions
Frequently asked.
How does the $500 deposit-match offer work?
We match your deposits 1:1 in bonus credits. Sign up, add $100, and we add another $100 free. Every additional $100 you deposit is matched with $100 more — up to $500 in bonus credits per rolling 90-day period. Each match is applied after a brief account review, typically within 24 hours of your deposit (occasionally longer during high volume); once your account is verified, the credits always arrive.
Do you store or log my prompt contents or embedding vectors?
No. Nicrron enforces a strict Zero Data Retention (ZDR) architecture for chat, embeddings, and audio: prompts, completions, and embedding vectors are streamed in-memory only and never persisted to disks or databases. We only record metadata (latency, status codes, token counts, and cryptographic request hashes) for auditability and settlement. The one exception is asynchronous video/image generation, where the output is retained for 30 days so you can download it, then permanently deleted.
What is the Model Arena and how does blind benchmarking work?
The Model Arena allows developers to benchmark multiple models concurrently on identical prompts to evaluate Time-to-First-Token (TTFT), generation speed (TPS), and real-time cost differences. In Blind Battle mode, model identities are masked (Model Alpha vs Beta) for unbiased qualitative evaluation.
How do Organization Teams & Shared Wallets work?
Teams allow multiple engineers to route LLM queries against a shared universal credit pool. Organization owners can assign fine-grained RBAC roles (Owner, Admin, Developer, Viewer) and set alert thresholds.
How do Spend & Reliability Webhook Alerts work?
You can configure automated webhook triggers that dispatch cryptographically signed HMAC-SHA256 JSON payloads to any HTTPS endpoint you control when credit balances drop low or daily spend velocities spike — relay them into Slack, Discord, or your incident tooling from there.
How does Smart Routing (nicrron/auto, reasoning, coding) work?
Nicrron dynamically analyzes prompt complexity, token volume, and live provider health. Use 'nicrron/reasoning' to route to DeepSeek V4 Pro and GPT-6 Sol, 'nicrron/coding' to route to Claude Opus 5.5 and Kimi K2.7 Code, or 'nicrron/auto' for general smart routing.
Do you support OpenAI-compatible Embeddings?
Yes. Use your standard OpenAI SDK or cURL pointing to /v1/embeddings to generate embeddings across OpenAI (text-embedding-3-small/large), Mistral (mistral-embed, codestral-embed), and Voyage (voyage-4 family) with automatic provider failover.
Can I generate videos, images, and audio through Nicrron?
Yes. POST /v1/videos/generations runs Seedance 2.5 and 2.0 (native audio), Seedance 1.0 Pro, Veo 3.1 Fast, Kling 2.5 and Hailuo 02 as async jobs with the exact cost charged upfront and auto-refund on failure; /v1/images/generations serves Nano Banana from $0.039/image; /v1/audio/speech and /v1/audio/transcriptions provide OpenAI TTS and Whisper with the standard OpenAI SDK. Finished media is re-downloadable for 30 days.
How do Prompt Caching discounts work?
Prompt caching discounts are passed through automatically. Cache reads with Anthropic receive a 90% discount, and OpenAI cache reads receive a 50% discount, recorded directly in your usage metrics.
Can I export my telemetry logs for accounting or analytics?
Yes. Telemetry logs can be filtered by date range, provider, status, and cost, and exported with one click to RFC 4180 CSV or structured JSON.
Developers
Point your SDK at api.nicrron.ai.
Two lines change: the base URL and the key. Every model, one balance.
Builders
Describe an app and watch it get built.
Ninoo runs on the same account and balance, in the browser or in your terminal.