Model catalog

Routic offers a curated catalog of AI models through a single API endpoint. All models are accessible via the same Base URL; you select which model to use with the model parameter in your request.

Base URL

https://api.routic.ai/v1

Authentication

All requests require an API Key (sk-*) in the Authorization header:

Authorization: Bearer sk-xxxxxxxx

How to select a model

Use the model parameter in the request body. Two naming styles are supported:

Style 1: Canonical model name (recommended)

Use the industry-standard model identifier. This is the primary way to call a specific model:

{
  "model": "deepseek-r1",
  "messages": [{ "role": "user", "content": "Hello" }]
}

Style 2: Smart routing name

Use a Routic-managed routing identifier. Routic automatically picks the best available model for that capability:

{
  "model": "auto/reasoning",
  "messages": [{ "role": "user", "content": "Hello" }]
}

Both styles work on the same endpoint (/v1/chat/completions). Canonical names give you precise control; smart routing names let Routic optimize routing behind the scenes.

Available models

Reasoning models

Reasoning models use extended chain-of-thought processing before producing a response. They excel at complex problem-solving, mathematical reasoning, and logic-heavy tasks.

General chat models

General-purpose conversation models optimized for speed and cost. Suitable for everyday Q&A, summarization, and content generation.

Code models

Models optimized for code generation, programming assistance, and software engineering tasks.

Distill models (cost-effective)

Distilled versions of top models, offering good performance at lower cost. Ideal for high-volume, latency-sensitive workloads. These models are trained on reasoning outputs but do not support the thinking API parameter.

Partner models

Models hosted through partner infrastructure with specific routing configurations.

Availability note: deepseek-r1-0528 and deepseek-v3-0324 are listed as salable SKUs but may not have a live upstream deployment at all times. If you need guaranteed availability, use deepseek-r1 or deepseek-v3 instead, or enable smart routing for automatic failover.

Cache pricing: The "Cache (input)" column shows the per-model cache hit rate. A dash (—) means cache hit pricing is not yet published for that model; check the console for the latest rate. See Context caching for details.

Per-model parameter reference

Thinking Mode (reasoning)

For details, see Thinking Mode.

Context caching

Context caching is enabled by default on all models. Cache hit tokens are billed at a lower rate (approximately 10x cheaper than standard input). The "Cache (input)" column in the model tables above shows the per-model cache hit rate.

See Context caching for details and best practices.

API endpoints

All models share the same API endpoints:

Capability matrix

Note: "Reasoning" refers to extended chain-of-thought processing (Thinking Mode). Models marked "Yes" support the thinking parameter. "Function calling" refers to the tools / tool_choice parameter support.

Rate limits

Each API Key has default RPM (requests per minute) and TPM (tokens per minute) limits. Default values:

To request higher limits, contact support.

Model selection best practices

1. Use canonical names for predictable behavior and reproducible results.
2. Use reasoning models (deepseek-r1, qwq-32b) for math, logic, and complex analysis.
3. Use general chat models (deepseek-v3.*) for Q&A, summarization, and content generation.
4. Use code models (qwen3-coder-plus) for code generation, refactoring, and programming assistance.
5. Use distill models for high-volume, cost-sensitive workloads where latency matters.
6. Use minimax-m2.5 when you need ultra-long context (up to 1M tokens).
7. Enable context caching for multi-turn conversations with repeated context (default: on).
8. Use smart routing for automatic failover when model availability changes.

Coming soon

The following capabilities are planned but not yet available:

• Vision models (image understanding)
• Video generation
• Embedding / vector models
• FIM (Fill-In-the-Middle) completion for code

See also

• Quickstart
• Chat completions endpoint
• Authentication & keys
• Thinking Mode
• Context caching
• Smart routing