MODELCASCADE — Intelligent Model Router

Tier	Coverage	Cost / 1K req	Status
LOCAL Ollama · llama.cpp · vLLM · R2	74% of all requests handled locally	$0	PROVEN
FAST claude-haiku-4-5 · Groq · OpenAI mini	+18% of requests escalated from local	$0.001	LIVE
CAPABLE claude-sonnet-4-6 · GPT-4o · Gemini	+8% hard tasks only · full coverage	$0.005	LIVE
MAX 20X Flat rate · Undifferentiated · April 2026	100% at uniform cost — the arbitrage they closed	$200/mo	DISPLACED

! Coming from a flat-rate plan?

The $200 credit
expires April 17.

If your AI tools just got more expensive because flat-rate access ended, you're not alone. 74% of those requests didn't need a cloud model. ModelCascade routes them locally at $0 — your keys, your hardware, your data. The rest cascade to the cheapest model that works.

# mc.yaml — migrate from flat-rate in 60 seconds

providers:
  local:
    type: ollama
    model: llama3.2:3b        # free, runs on your machine
    cost_per_1k: 0.0

  fallback:
    type: anthropic
    api_key: ${ANTHROPIC_API_KEY}  # your key, your billing
    model: claude-haiku-4-5-20251001
    cost_per_1k: 0.001

routing:
  cost_ceiling: 0.01
  cascade_on_failure: true
  calibration: preset_v1

02 Installation

Download.
Configure.
Route.

Install on Windows (.exe) or Linux (.deb). Bring your own API keys — Anthropic, OpenAI, Groq, Ollama. The router classifies each request, routes to cheapest capable model, escalates on quality failure or timeout. Desktop app runs locally. No cloud dependency. Also available as a Python package for framework integration.

bash · ~/your-agent

── Windows ──

>modelcascade-setup.exe

✓ Installed to C:\Program Files\ModelCascade

── Linux / Debian ──

$sudo dpkg -i modelcascade_0.1.0_amd64.deb

✓ modelcascade installed

── Python (framework integration) ──

$pip install modelcascade

Successfully installed modelcascade-0.1.0

$modelcascade route "explain quantum computing"

→ routed LOCAL · $0.000 · 47ms

03 Configuration

YAML.
BYOK.
Done.

One file. No signup. No dashboard. Your API keys stay local.

cost_ceiling — hard stop per request, no runaway spend

cascade_on_failure — escalates on quality failure or timeout

preset_v1 — production-calibrated routing tables from 21/21 A/B tests

Mix any providers — local Ollama + Anthropic + Groq in one cascade

# modelcascade.yaml

providers:
  local:
    type: ollama
    model: llama3.2:3b
    cost_per_1k: 0.0

  fast:
    type: anthropic
    api_key: ${ANTHROPIC_API_KEY}
    model: claude-haiku-4-5-20251001
    cost_per_1k: 0.001

  capable:
    type: anthropic
    api_key: ${ANTHROPIC_API_KEY}
    model: claude-sonnet-4-6
    cost_per_1k: 0.005

routing:
  cost_ceiling: 0.01        # per request, hard stop
  cascade_on_failure: true
  calibration: preset_v1  # production-calibrated
  escalation: quality_or_timeout

# Works with: LangChain · CrewAI · Claude Code · Custom

04 Design principles

01 · CLASSIFICATION

Classify first,
spend second

Every request gets a difficulty score before a provider is chosen. The score drives the routing. The routing drives the cost. Not the other way around.

02 · ESCALATION

Fail cheap,
succeed capable

When a lower tier fails — quality threshold not met, timeout, rate limit — the executor escalates automatically to the next capable model. No manual intervention.

03 · OWNERSHIP

Your keys.
Your data.

modelcascade never routes through our servers. Your API keys stay on your machine. The calibration tables are yours to inspect and override. No telemetry.

05 Pricing

Every unrouted API call is money you already lost.

74% of requests can route locally at $0. The rest cascade to the cheapest model that meets your quality bar.

Average team API spend: $200/mo → With ModelCascade: from $0

Open Source

Save ~$148/mo

Full routing engine. No artificial limits. MIT forever.

Unlimited requests
Full cascade routing engine
BYOK — all providers
YAML configuration
CLI interface
Community support (GitHub)

⬇ Download Now

Most chosen by developers

Pro

Save ~$171/mo

$29/mo

$171 saved / $29 spent = 5.9x return

Managed routing + analytics. Overnight batch routes 95% locally.

50,000 managed requests/mo
Web dashboard + cost tracking
Pre-configured providers
Desktop app (Windows + Linux)
Email support (48h)
Overage: $0.0003/request

Start Routing

Scale

Save ~$450/mo

$99/mo

One workspace. Your whole team. No per-seat math.

Shared dashboard, cost tags, overnight batch optimization. Morning report your manager can forward to finance.

500,000 managed requests/mo
Unlimited users per workspace
Per-project cost allocation tags
A/B model testing
Priority support (4h)
Overage: $0.0002/request

Start Routing

Dedicated

Custom

Your models, your infrastructure, our routing — deployed inside your VPC with an SLA attached.

Unlimited managed requests
SSO + audit trail
On-premise / VPC deployment
Custom model hosting
Dedicated Slack + SLA
Volume discounts

Talk to Us

What happens if I hit my request limit?

Nothing breaks. Your requests keep routing. We notify you at 80% and 100%. Overage is billed at the same per-request rate — no surprise multipliers. You can also set a hard cap: "never exceed $X/mo" and we stop routing to paid models when you hit it.

What happens to my requests at 3 AM?

They get cheaper. Overnight routing shifts batch requests to the lowest-cost models — typically 95% local routing vs 74% during peak hours. Your morning report shows exactly what ran, which models were used, and what it cost. You review, not relay.

Is this just another API proxy?

No. ModelCascade is open source (MIT). The routing engine runs in your environment. We never see your prompts and never proxy your API calls. Paid tiers add managed keys, analytics, and support — not routing capability. You can leave anytime and keep the same pip package running.

My team is 3 people. Do I need the Scale tier?

Probably not yet. Start with Pro. When you want a shared dashboard or need cost-per-project tags for your manager's quarterly review, upgrade. Scale exists for organizational accountability, not technical capability.

Cancel anytime · Export your config · Self-host forever · Your keys never leave your machine
Open Source tier is permanent. Paid tiers add operational tooling — the MIT core never changes.
Prices locked for 12 months from your start date.

WTP = (Cost_Without − Cost_With) × Risk × Trust

MODEL CASCADE ROUTER