OnyRouter: Automatic AI Model Selection Per Request (onysoft/auto API Guide)

With 708+ models in the catalog, "which model should I use?" is a real engineering burden for every request: wire quick tasks to a flagship and your bill balloons; send deep analysis to a budget model and quality visibly drops. OnyRouter takes that decision off your plate — it analyzes the incoming message, classifies the question type and routes the request to the model best suited for the job. Usage is a single line: set the model field to onysoft/auto. Routing itself is completely free; only the selected real model's usage is billed. Few platforms in the world offer this today, and Onysoft is the first Turkish AI gateway to ship it — both inside the panel Chat and directly on the API, ready to power your own applications.

What Is OnyRouter and How Does It Work?

OnyRouter is a two-layer routing engine. The first layer is heuristic and adds zero latency: if the message carries code signals (code blocks, error output, technical commands) it goes straight to the coding category; a very short greeting goes to the quick category. The second layer is a classifier: messages the heuristics can't settle are shown to a light, fast model that answers "what type of question is this?" in a single word. Classification runs on the platform's own infrastructure — you pay neither tokens nor fees for it.

The resolved category maps to a real model through a routing table: quick questions go to the best speed-class model, coding and general chat to a balanced flagship, deep analysis to the strongest reasoning model, bulk mechanical work to the best price/performance model, and image-bearing requests to a vision-capable model. We keep the routing table current as the catalog evolves — when a new model becomes best-in-class we update the route, while your code keeps saying onysoft/auto. The choice is transparent on every response: the model field of the API response shows the actual routed model.

Why Automatic Model Selection?

Production systems tend to fall into one of two traps: sending everything to a single flagship (you pay flagship prices while 80% of traffic is simple work) or sending everything to a single budget model (your users notice the quality drop on critical questions). The right architecture is tiered model selection — but building it by hand means writing request classifiers and tracking the model catalog forever.

OnyRouter gives you that tier as a managed layer. Similar approaches are only now emerging globally — OpenRouter's auto route, or research such as RouteLLM — but Onysoft is the first local Turkish gateway to combine it with TL billing, local support and KVKK compliance. It also works on two surfaces at once: "⚡ OnyRouter" is the default option in the in-panel Chat, and the onysoft/auto virtual model brings the same intelligence to your own SaaS, chatbot or internal tool with a one-line change.

Using It on the API: Python, curl and Streaming

There is no special endpoint — the familiar /v1/chat/completions with the model name onysoft/auto:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.onysoft.com/v1",
    api_key="sk-ony-YOUR_KEY",
)

response = client.chat.completions.create(
    model="onysoft/auto",
    messages=[{"role": "user", "content": "Find and fix the bug in this Python function: ..."}],
)
print(response.model)   # the routed real model (e.g. anthropic/claude-sonnet-5)
print(response.choices[0].message.content)

The same request with curl:

curl https://api.onysoft.com/v1/chat/completions \
  -H "Authorization: Bearer sk-ony-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "onysoft/auto",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Streaming works the same way — pass stream=True and the answer streams live from the selected model; the model field of the chunks tells you which model is answering. onysoft/auto also appears as a virtual model in the /v1/models list, so applications that discover the catalog programmatically will find it too.

Categories: Which Question Goes Where?

OnyRouter works with six categories. The current routing logic in short:

• Quick — short/simple questions, greetings, one-line lookups → the leading speed-class model (currently the Gemini Flash class).
• General — normal conversation, writing, summarization, translation → a balanced flagship (currently the Claude Sonnet class).
• Code — programming, debugging, technical commands → the strongest balanced coding model.
• Deep — multi-step analysis, strategy, math/logic, long-document evaluation → the reasoning flagship (currently the GPT-5.6 class).
• Bulk — high-volume mechanical work such as batch classification/labeling → the best price/performance model (currently the DeepSeek class).
• Vision — image/photo analysis requests → a strong vision-capable model.

The examples in parentheses reflect the routes at the time of writing; we update them as the catalog evolves. For endpoints where you want to pin a specific model, keep using its full name (like anthropic/claude-sonnet-5) — onysoft/auto is an option, not a requirement.

Billing, Transparency and Limits

The billing model is deliberately simple: routing is free, and the request is billed at the selected real model's catalog price. Your usage logs and panel reports show the request under the routed model's name; there is no surprise "router fee" line item. For cost planning, run your expected per-category volume through the cost calculator with the corresponding route's model.

Two limits worth knowing: (1) On keys with a model lock (allowed_models), the request returns 403 if the routed model isn't on the allow-list — locked keys are dedicated to specific models by design, so use explicit model names there. (2) The classifier layer can add a few seconds to time-to-first-token; for clear-cut cases like code or greetings the zero-latency heuristics kick in, so most of the time you won't feel it. A good middle ground: keep latency-critical endpoints on fixed model names and let the rest of your traffic ride onysoft/auto.

To get started: create a free account, generate your sk-ony- key in the panel, and set the model to onysoft/auto on your first request. To try it without writing code, ⚡ OnyRouter is already the default in the panel Chat.