ContrieContrie/Docs
PlaygroundDashboard
OverviewAuthenticationScrapeExtractStreamingModelsMCPFor agentsRate LimitsErrorsExamples

AI Models

Contrie uses a fleet of 17 models across 8 providers. The router selects the cheapest model that can handle the classified complexity, and escalates a tier whenever the quality score lands under 70. This table renders from the same source file the router reads — it cannot go stale.

Model Fleet

ModelTierContextVisionPriority
minimax/minimax-m3:freefree1049Kyes10
nvidia/nemotron-3-super-120b-a12b:freefree262K—9
z-ai/glm-5.2:freefree256K—8
nvidia/nemotron-3-ultra-550b-a55b:freefree1000K—6
minimax/minimax-m2.7:freefree197K—4
google/gemma-4-31b-it:freefree262Kyes2
deepseek/deepseek-v4-pro-0813budget1049K—15
deepseek/deepseek-v4-flashbudget1049K—12
deepseek/deepseek-chat-v3-0324budget164K—10
openai/gpt-5.6-lunastandard1050Kyes10
openai/gpt-5-ministandard400Kyes8
anthropic/claude-haiku-4.5standard200Kyes6
bytedance/ui-tars-1.5-7bvision128Kyes16
google/gemini-2.5-flash-litevision1049Kyes10
google/gemini-3.1-flash-litepremium1049Kyes12
google/gemini-3.7-flashpremium1049Kyes8
anthropic/claude-sonnet-5frontier1000Kyes10

How Model Routing Works

  1. The classifier scores the page — simple, complex, visual, or adversarial — from HTML size, structure, and anti-bot fingerprints
  2. The router picks the cheapest capable model within your tier (keyless demo and free keys use free models only)
  3. Every result is quality-scored 0–100; anything under 70 escalates to the next tier and retries, keeping the best attempt
  4. The free tier walks all free models in priority order before giving up

Vision Cascade

Requests that include a screenshot route through vision-capable models, escalating by tier until quality clears the bar:

ui-tars-1.5-7b (vision) → gemini-3.1-flash-lite (premium) → claude-sonnet-5 (frontier)

The full control flow, with a diagram, lives on the architecture page.