router.Nyuro.ai

Models & aliases

Everything you can pass as `model` — concrete aliases, industry tags, strategy hints, and auto.

The model field is the steering wheel. You can name a model directly, or hand the choice to the router with a tag, a strategy, or auto. Every form resolves through the same routing pipeline.

What you can pass as model

Concrete alias

A specific model from the curated set below: gpt-4o, claude-3-5-sonnet, gemini-flash-latest, Qwen/Qwen3-32B-AWQ, … Forwarded to the backend that serves it.

Industry tag

industry:legal, industry:healthcare, industry:code, … — the router picks the curated best model for that domain and falls through a priority list if your top pick is offline.

Strategy hint

strategy:cost, strategy:quality, strategy:latency, strategy:local — route by intent rather than name. See Routing.

Auto

auto — a balanced default. The classifier inspects the prompt and routes to the best model for the task, so you can ship once and let us tune.

Industry tags

TagTuned for
industry:legalContracts, citations, careful nuance
industry:healthcareClinical Q&A, polite refusals
industry:financeQuantitative reasoning, numeric reliability
industry:codeCode generation, debugging, refactors
industry:creativeLong-form writing, brand voice
industry:supportFast, polite, tool-aware
industry:salesLead enrichment, outreach drafts
industry:dataSQL, dataframes, chart reasoning
industry:educationStep-by-step tutoring
industry:generalBalanced default

The curated set

Nyuro runs a deliberately small, curated roster rather than a long tail: the frontier models from OpenAI, Anthropic and Google, plus open-weight models we host ourselves on RunPod. Every model below is routable today.

Frontier models

AliasProviderTierContext$/Mtok in$/Mtok out
gpt-4oOpenAIquality128K5.0015.00
gpt-4o-miniOpenAIcost128K0.150.60
claude-3-5-sonnetAnthropicquality200K3.0015.00
claude-3-5-haikuAnthropiccost200K0.804.00
gemini-pro-latestGooglebalanced1M1.2510.00
gemini-flash-latestGooglecost1M0.302.50

All six support tool calling. gpt-4o and both Gemini models also support vision; claude-3-5-sonnet is the one carrying the code capability.

Open-weight models we host

Served from our own RunPod deployment, so there is no third-party provider in the path and no per-token provider meter:

AliasFamilyTierUse
Qwen/Qwen3-32B-AWQQwen 3balancedGeneral chat, quantized for throughput
Qwen/Qwen3-30B-A3B-Instruct-2507Qwen 3 (MoE)balancedGeneral chat, instruction-tuned
ibm-granite/granite-4.0-h-smallIBM Granite 4balancedGeneral chat
Qwen/Qwen3-Embedding-0.6BQwen 3costEmbeddings/v1/embeddings and collections

Qwen/Qwen3-Embedding-0.6B is the embedding model, not a chat model — pass it to /v1/embeddings, not /v1/chat/completions. It produces 1024-dimension vectors and is the model backing managed collections.

Discovering models programmatically

Two endpoints, for two different questions:

# What can I route to right now? (needs the models.read scope)
curl https://api.routing.nyuro.ai/v1/models \
  -H "Authorization: Bearer nyu_live_…"

# Full catalog with pricing, capabilities and latency
curl https://api.routing.nyuro.ai/api/v1/catalog/models

The catalog is larger than the routable set

/api/v1/catalog/models returns thousands of entries enriched from the upstream registry, but only the curated models above carry "routable": true. Filter on that flag — naming a non-routable catalog entry will not route. /v1/models returns the routable set directly.

Open-weight & self-hosted models

Concrete aliases can also cover models you host yourself. Configure an OLLAMA_BASE_URL or VLLM_BASE_URL on the gateway and those models join the catalog alongside the curated set — including under strategy:local. See BYOK & local models.

On this page