Skip to content
IQ Routing

The Cognitive Runtime for AI Agents.

Route each inference call to the cheapest model that holds quality. Cut agent loop costs 40 to 60 percent, live in thirty seconds.

40 to 60 percent spend cut
approximately 50 percent cache hits
38 milliseconds median cache-hit latency

How we measure savings

Drop in front of the OpenAI or Anthropic SDK you already use

OpenAI·Anthropic·Google

Any OpenAI- or Anthropic-shaped endpoint drops in behind one base URL. Groq, Mistral, xAI, and self-hosted models are landing on the same shape.

The problem

Every step of every agent loop hits your top model. The bill keeps climbing and nobody can say why.

Pin one frontier model on planning, retrieval, tool calls, and cleanup alike and you pay reasoning rates for work a cheap model would nail. Then the invoice lands as one line, shipping outruns anything finance can categorise, the cost report is a guess, and the audit trail is a Slack thread.

Which team spent $14k last week?
team_id=engineering · 9,212 calls · 89% on opus
Was that PII in the prompt?
2 spans redacted when enabled · ssn, email · masked pre-provider
How do we book this to a project?
tagged at gateway · ERP export on Enterprise

Before and after on a single agent loop

The same five-step LangChain loop, routed two ways.

This is the shadow-mode report we hand you. The Before trace pins every step on the same top-tier reasoning model. The After trace lets the resolver pick the cheapest variant that holds the quality bar on each step. The total drops 49 percent, and nothing in the agent code changes.

How we measure savings

Before · vanilla LangChain

one provider, one budget

Every step on claude-opus-4-8 at thinking=high

  • Planning
    claude-opus-4-8 · thinking=high
    $0.42
  • Retrieval
    claude-opus-4-8 · thinking=high
    $0.38
  • Tool call
    claude-opus-4-8 · thinking=high
    $0.28
  • Synthesis
    claude-opus-4-8 · thinking=high
    $0.41
  • Verification
    claude-opus-4-8 · thinking=high
    $0.35
Loop total$1.84

After · IQ Routing

per-step resolver, session envelope

The resolver picks the cheapest variant per step

  • Planning
    cap:reason-heavy · claude-opus-4-8 high
    $0.42
  • Retrieval
    cap:long-context-128k · gpt-4o
    $0.06
  • Tool call
    cap:tool-call-strict · gpt-4o no thinking
    $0.04
  • Synthesis
    cap:reason-heavy · claude-opus-4-8 high
    $0.41
  • Verification
    cap:cheap-fast · gpt-4o-mini
    $0.01
Loop total
$0.94
49 percent less

The fix

One gateway. Six superpowers.

Each one shows up on the dashboard the moment your first request lands.

State

Works inside the apps that lock models.

Some tools only allow one model family. Point Claude Code, Cursor, or your own app at IQ and it stays inside the family they require, routing to the right tier for each step and never leaking to another vendor. It holds the conversation too, so switching models mid-thread stays safe.

Read the docs →

Cache

Repeats resolve before they bill.

When a request matches one IQ has already handled, the answer comes straight back and you are not charged for it again. IQ catches exact repeats and the ones that just mean the same thing, and every team's cache stays walled off from the rest.

Read the docs →

Route

The router is the hard part.

Routing well is the whole product, and it is our core IP. A purpose-built classifier reads each request for true difficulty, weighs it against live cost and latency across providers, and picks the model that clears your quality bar for the least spend, with instant fallback if one slips. It is tuned on a graded prompt set, so it holds quality where a naive cheapest-model router quietly drops it.

Read the docs →

Book

Tagged to your GL.

Per-team budgets, per-key limits, and a four-way input, output, cached, and thinking cost split, written to the ledger and enforced before the provider call.

Read the docs →

Session

A session is one accountable unit.

Every multi-step agent loop tracks per-turn cost, latency, tokens, and step class, and a governor caps how much thinking the session can spend.

Read the docs →

Step

Each step gets the right model.

An agent runs many steps, and most of them don't need your most expensive model. IQ sizes up each one and routes it to the cheapest model that can still do it well, so you get frontier quality where it counts and pay a fraction everywhere else.

Read the docs →

What the dashboard looks like

Spend, budgets, and per-team usage on one page.

Pricing

Priced so it pays for itself.

Start free in thirty seconds. Scale into Team and Enterprise when the bill makes the case.

Free

$0Self-serve

One URL for every provider. Bring your own provider keys (BYOK).

Usage limit240 requests/min
  • OpenAI, Anthropic, and Google, with one URL for all three
  • Bring your own keys (BYOK)
  • Semantic cache
  • Email support
Start free

Team

Most popular
$70per month

Everything in Free plus alerting, org-scoped access, and ERP cost-event export (preview).

Usage limit600 requests/min
  • Everything in Free
  • Per-team budgets
  • Audit log
  • Per-team alerting
  • Org-scoped access controls
  • ERP cost-event export (preview)
  • Priority email support
Upgrade to Team

Enterprise

CustomAnnual

On-prem deployment, SOC2 evidence pack on request, dedicated support.

Usage limit6,000+ requests/min
  • Everything in Team
  • On-prem or VPC deployment
  • SOC2 evidence pack on request
  • ERP integrations (preview)
  • Dedicated support engineer
Contact sales

Three lines today. A clean bill tomorrow.

Route real traffic in under five minutes.