Not every question needs your most expensive model.
You have one API key. Right now, every query hits the same model at the same price. Routing classifies each question and sends it to the right model. Within a single provider, or across providers.
Simple questions get the fast, cheap model. Complex ones get the heavyweight. Same key. Same provider. Less waste.
Routing works through the ASURIQ API or MCP server. You send queries to us with your provider key. We classify, route, and return the response.
One key. Three models. Immediate savings.
Start with the key you already have. Routing classifies every query by complexity and routes within your provider's model family. Simple questions go to the fast model. Complex ones get the heavyweight. Your API key stays the same.
The cheapest model in each family is 3x to 40x less expensive than the most capable. Routing knows which one to use for each question. You save money without losing quality.
One model has opinions. Multiple models have knowledge.
Single-provider routing saves money. Multi-model routing saves money AND dramatically improves accuracy. Here's why: models trained on different data fail on different questions. When Claude and DeepSeek disagree on an answer, that disagreement is the most valuable signal in the entire system.
In 1785, the Marquis de Condorcet proved a mathematical law: if each voter in a jury is right more than 50% of the time, and their errors are independent, the group accuracy converges toward 100% as you add jurors.
LLMs are jurors. Claude, GPT, Gemini, DeepSeek, Kimi: each is right more than 50% of the time. And because they're trained on different data, their errors are genuinely independent. The conditions for the theorem are met.
One model at 85% accuracy. Three independent models at 85% accuracy: 96.6% panel accuracy. Five models: 99.3%.
Claude was trained by Anthropic. GPT by OpenAI. DeepSeek by a Chinese lab. Gemini by Google. They consumed different training corpora, applied different RLHF, and developed different failure modes.
When Claude confidently says X and DeepSeek confidently says not-X, something interesting is happening. One of them is wrong, and the disagreement itself tells you which questions need more scrutiny.
If they all agreed on everything, you wouldn't need more than one. The value is in the disagreement surface.
"But doesn't calling 3 models cost 3x more?" Not with Routing.
60% of your queries are simple. They route to one cheap model. $1/Mtok. 30% are moderate. One model, mid-tier. $3/Mtok. Only 10% are high-stakes enough to justify a multi-model panel. And even then, the panel uses the cheapest effective combination.
Blended cost is 30-40% below running everything on Sonnet, with higher accuracy on the questions that actually matter.
Pair with Cognitive Stack for the Intelligence Layer bundle: $19/mo instead of $22 separate.
Your models have personalities. We profile them.
Static benchmarks are snapshots from months ago. Routing builds live behavioral profiles from your actual queries across 8 dimensions. These profiles drive both single-provider routing and multi-model panel selection.
Day 1: routes based on published benchmarks and cost.
Day 30: routes based on how each model actually performs on your questions.
Day 90: knows your domain well enough to predict which model, or which panel, will nail each query.
Every key you add deepens the savings and sharpens the accuracy.
Start single-provider. Add more keys as you see the value. Each new provider gives Routing more models to choose from and more independent perspectives for panel decisions. Keys are stored server-side in an encrypted vault — add them once in settings and every connector uses them immediately, with no re-entry per session.
Switch in one line. Switch back in one line.
POST https://api.anthropic.com/v1/messages
{ "model": "claude-sonnet-4-6", "messages": [...] }POST https://consensus-production-6eeb.up.railway.app/api/v1/routing/route
{ "query": "Your question", "api_key": "sk-..." }POST https://consensus-production-6eeb.up.railway.app/api/v1/routing/route
{ "query": "High-stakes question", "mode": "panel",
"api_keys": { "anthropic": "sk-...", "openai": "sk-...",
"deepseek": "sk-..." } }Routing is ASURIQ's quality guarantee.
Every ASURIQ tool has a minimum model requirement. WHETSTONE needs a Sonnet-class model to generate meaningful adversarial challenges. Confidence scoring needs mid-tier reasoning to produce calibrated axes. Without Routing, tools that need a stronger model simply won't fire. You'll see an explanation and an offer to add Routing.
With Routing, every tool call silently upgrades to the cheapest sufficient model. Your conversation stays on Haiku. Your WHETSTONE challenge fires on Sonnet. Invisible quality guarantee, zero wasted spend.
Bring your own API keys. ASURIQ stores them server-side in an encrypted key vault, so every connector — extension, MCP, ChatGPT, Gemini, API — routes through the same keys without you re-entering them per session. You pay your providers directly; ASURIQ handles the intelligence: classification, model selection, behavioral fingerprinting, quality guarantees.
No API keys needed. ASURIQ provides the models. You pay credits per routed query. Same routing intelligence, we handle the infrastructure.
Currently in friends-and-family beta. Built on peer-reviewed cognitive architecture.
Stop paying premium prices for every question.
$13/mo · Typically saves $20+/mo · Works with any provider · Cancel anytime