niuron picks between OpenAI, Anthropic, Gemini and your own models on quality, cost and latency — and shows why it chose.
One API. Many models. niuron decides which one to use on every request.OpenAI-compatible, with fallback, caching and BYOK.
Hardcoding a model name means re-testing every time a better or cheaper one ships. A niuron model is a routing rule instead: the quality bar, latency budget and capabilities one kind of work needs, resolved against the live catalog on every request.
The default. Skip the model field and this is what you get: every request is read and sent to the model that balances quality, cost and speed for it.
one line to switchclient.chat.completions.create(
model="niuron/code", # the job, not the model
messages=[...],
)Uses one free model up to its limit, then the next, and only pays once every free slot is spent, with the cheapest model that clears the quality bar. A response header tells you which side you landed on.
Free-tier limits belong to each provider and change. niuron stops short of every limit instead of running into it.
Code generation, review, debugging and refactors. Strong coding models that also answer fast.
Agents that call tools in a loop. Only models with reliable function calling and JSON, weighted hard toward speed because every step's latency stacks.
Math, logic and multi-step analysis. Quality first: it will wait for a slower model when the answer needs it.
Derivations, data analysis and technical writing, with room for long papers in context.
Long documents and many sources at once. Only models with a very large context window, tuned for careful synthesis.
Structured extraction and classification at volume. The cheapest model that returns valid JSON, and quickly.
Copy, articles and creative text. A quality bar for prose without paying reasoning-model prices.
Translation and localization, quality-gated and tuned for speed.
Pin any specific model whenever you want. niuron models are a shortcut, not a lock-in.
No new SDK and no wrapper library. Three steps, and the last one is watching it work.
Starter credit included, no card required. Your first API key is one click away in the dashboard.
Swap the base URL and key in your existing OpenAI SDK. Everything else, from messages to streaming to tool calls, stays the same.
base_url = "https://api.openai.com/v1"base_url = "https://api.niuron.ai/v1"api_key = "sk-..."api_key = "ar_live_..."
Every request is classified and scored against your registered models, then sent to the best eligible one.
Every provider you call directly adds an SDK, a key, a bill and an outage plan. This is the pull request niuron makes possible.
from openai import OpenAIfrom anthropic import Anthropicimport google.generativeai as genaiopenai_client = OpenAI(api_key=OPENAI_KEY)claude = Anthropic(api_key=ANTHROPIC_KEY)genai.configure(api_key=GEMINI_KEY)
def ask(prompt):for provider in (openai_client, claude, gemini):try:return provider.complete(prompt, timeout=20)except (RateLimit, Timeout, ServerError):time.sleep(backoff()); continueraise AllProvidersDown
Retries, fallback chains and a circuit breaker for every provider. A provider outage isn't your outage.
MODEL_BY_TASK = {"code": "gpt-4o", "summarize": "claude-haiku"}def pick_model(task, budget):... # re-tested every time a better or cheaper model shipsclient = OpenAI(base_url="https://api.niuron.ai/v1", api_key=NIURON_KEY)client.chat.completions.create(model="niuron/code", messages=msgs)
Pin a model or send niuron/code. Same request, same response shape, no code change either way.
Illustrative code. Your stack will differ; the shape of the work does not.
OpenAI-compatible. Point your existing client at /v1/chat/completions with your router API key — no new SDK to learn.
curl https://api.niuron.ai/v1/chat/completions \
-H "Authorization: Bearer ar_live_..." \
-H "Content-Type: application/json" \
-d '{
"model": "niuron/auto",
"messages": [{"role": "user", "content": "Explain routing in one sentence."}]
}'from openai import OpenAI
client = OpenAI(
base_url="https://api.niuron.ai/v1",
api_key="ar_live_...", # Dashboard -> API keys
)
resp = client.chat.completions.create(
model="niuron/auto", # or niuron/code, niuron/agent... or pin "gpt-4o"
messages=[{"role": "user", "content": "Explain routing in one sentence."}],
)import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.niuron.ai/v1",
apiKey: "ar_live_...", // Dashboard -> API keys
});
const resp = await client.chat.completions.create({
model: "niuron/auto", // or niuron/code, niuron/agent... or pin "gpt-4o"
messages: [{ role: "user", content: "Explain routing in one sentence." }],
});model: "niuron/auto" lets the router pick (leave the field out and you get it too). Use a job like niuron/code or niuron/agent when you know the work, or pin a specific model (e.g. gpt-4o, claude-sonnet-5). The response shape is identical either way.
Every OpenAI-compatible coding agent takes the same two settings. Swap them in, pick niuron/code as the model, and keep the tool you already use.
export ANTHROPIC_BASE_URL=https://api.niuron.ai
export ANTHROPIC_AUTH_TOKEN=ar_live_...
export ANTHROPIC_MODEL=niuron/auto[model_providers.niuron]
base_url = "https://api.niuron.ai/v1"
env_key = "NIURON_API_KEY"
wire_api = "responses"export COPILOT_PROVIDER_BASE_URL=https://api.niuron.ai/v1
export COPILOT_PROVIDER_API_KEY=ar_live_...
export COPILOT_MODEL=niuron/autoBase URL: https://api.niuron.ai/v1
API Key: ar_live_...
(click Verify)Base URL: https://api.niuron.ai/v1
API Key: ar_live_...
Model ID: niuron/codeBase URL: https://api.niuron.ai/v1
API Key: ar_live_...
Model: niuron/codeBase URL: https://api.niuron.ai/v1
API Key: ar_live_...models:
- name: niuron
provider: openai
model: niuron/code
apiBase: https://api.niuron.ai/v1
apiKey: ar_live_..."language_models": {
"openai_compatible": {
"niuron": {
"api_url": "https://api.niuron.ai/v1",
"available_models": [
{ "name": "niuron/auto", "max_tokens": 128000 }
]
}
}
}Provider: OpenAI-compatible
Base URL: https://api.niuron.ai/v1
API Key: ar_live_...export OPENAI_API_BASE=https://api.niuron.ai/v1
export OPENAI_API_KEY=ar_live_...
aider --model openai/niuron/codeOPENAI_BASE_URL=https://api.niuron.ai/v1
OPENAI_API_KEY=ar_live_...Each response can show which model was chosen, which alternatives were considered, the estimated cost and the savings.
{ "id": "chatcmpl-9f2a41c7", "object": "chat.completion", "model": "gemini-1.5-flash", "choices": [ … ], "usage": { "prompt_tokens": 12038, "completion_tokens": 604, "total_tokens": 12642 },"routing": { "selected_model": "gemini-1.5-flash", "selected_provider": "gemini", "strategy": "auto", "estimated_cost_usd": 0.0011, "baseline_model": "claude-3-5-haiku", "baseline_cost_usd": 0.0120, "estimated_savings_usd": 0.0109, "reason": "Best score for summarization tasks under auto routing - quality 84%, about 91% cheaper than claude-3-5-haiku, the highest-quality eligible option.", "alternatives": [ { "model": "gpt-4o-mini", "provider": "openai", "estimated_cost_usd": 0.0022 }, { "model": "claude-3-5-haiku", "provider": "anthropic", "estimated_cost_usd": 0.0120 } ] } }
Retries with full-jitter backoff, a circuit breaker per provider, and automatic fallback chains keep requests flowing.
Each request is classified and scored on quality, cost, latency and reliability. The best eligible model wins.
Exact-match and semantic (pgvector) caching cut repeat costs to zero, with tenant isolation and configurable safeguards.
Sample real traffic to an alternate model and compare it against production — your client only ever sees the response you chose to serve.
Cost, latency, cache hits, and the routing decision behind them — queryable in the dashboard, in real time.
One OpenAI-compatible endpoint and one balance. No separate contract, key or invoice for each provider.