What the marketplace installs and rates.

Aion Labs only: Aion-3.0-Mini for everyday requests, Aion-RP 1.0 (8B) for moderate work, Aion-3.0 for hard ones - with complexity rules and a cost-aware objective.

Amazon only: Nova Micro 1.0 for everyday requests, Nova 2 Lite for moderate work, Nova Premier 1.0 for hard ones - with complexity rules and a cost-aware objective.

Anthropic only: Claude 3 Haiku for everyday requests, Claude Opus 4.7 for moderate work, Claude Opus 4.1 for hard ones - with complexity rules and a cost-aware objective.

ByteDance Seed only: Seed 1.6 Flash for everyday requests, Seed 1.6 for moderate work, Seed-2.0-Code for hard ones - with complexity rules and a cost-aware objective.

Cohere only: Command R7B (12-2024) for everyday requests, Command A for moderate work, Command R+ (08-2024) for hard ones - with complexity rules and a cost-aware objective.

DeepSeek only: DeepSeek V4 Flash 0423 for everyday requests, DeepSeek V3.1 for moderate work, DeepSeek V4.1 Flash for hard ones - with complexity rules and a cost-aware objective.

Google only: Lyria 3 Pro Preview for everyday requests, Nano Banana 2 (Gemini 3.1 Flash Image) for moderate work, Nano Banana Pro (Gemini 3 Pro Image) for hard ones - with complexity rules and a cost-

Meta only: Muse Spark 1.3 Contributor for everyday requests, Muse Spark 1.1 for moderate work, Muse Spark 1.2 for hard ones - with complexity rules and a cost-aware objective.

Meta only: Llama 3.1 8B Instruct for everyday requests, Llama 3.3 70B Instruct for moderate work, Llama 3.1 70B Instruct for hard ones - with complexity rules and a cost-aware objective.

MiniMax only: MiniMax M2 for everyday requests, MiniMax M2.1 for moderate work, MiniMax M3 for hard ones - with complexity rules and a cost-aware objective.

Mistral AI only: Mistral Nemo for everyday requests, Mistral Small 4 for moderate work, Mistral Large 2407 for hard ones - with complexity rules and a cost-aware objective.

Moonshot AI only: Kimi K2.5 for everyday requests, Kimi K2.7 Code for moderate work, Kimi Latest for hard ones - with complexity rules and a cost-aware objective.

Nous Research only: Hermes 3 70B Instruct for everyday requests, Hermes 3 405B Instruct for moderate work, Hermes 4 405B for hard ones - with complexity rules and a cost-aware objective.

NVIDIA only: Nemotron 3.5 Content Safety for everyday requests, Nemotron 3 Nano 30B A3B for moderate work, Nemotron 3 Ultra for hard ones - with complexity rules and a cost-aware objective.

OpenAI only: gpt-oss-20b for everyday requests, GPT-5.1 for moderate work, o1-pro for hard ones - with complexity rules and a cost-aware objective.

Perplexity only: Sonar for everyday requests, Sonar Deep Research for moderate work, Sonar Pro Search for hard ones - with complexity rules and a cost-aware objective.

Qwen only: Qwen3.7 Flash for everyday requests, Qwen3.6 Flash for moderate work, Qwen3.8 Max (0902) for hard ones - with complexity rules and a cost-aware objective.

Sakana only: Sakana Namazu for everyday requests, Fugu Ultra for moderate work, Fugu Ultra v2 for hard ones - with complexity rules and a cost-aware objective.

Sao10K only: Llama 3 8B Lunaris for everyday requests, Llama 3.3 Euryale 70B for moderate work, Llama 3.1 Euryale 70B v2.2 for hard ones - with complexity rules and a cost-aware objective.

Tencent only: Hy-MT2-1.8B for everyday requests, Hy3 for moderate work, Hy4 preview for hard ones - with complexity rules and a cost-aware objective.

TheDrummer only: UnslopNemo 12B for everyday requests, Cydonia 24B V4.1 for moderate work, Skyfall 36B V2 for hard ones - with complexity rules and a cost-aware objective.

xAI only: Grok Build 0.1 for everyday requests, Grok 4.3 for moderate work, Grok 4.6 for hard ones - with complexity rules and a cost-aware objective.

Z.ai only: GLM Flash Latest for everyday requests, GLM 4.5V for moderate work, GLM 5.3 for hard ones - with complexity rules and a cost-aware objective.

Every request routed among hosted models that cost nothing per token; quality decides.

Models with extended thinking for multi-step problems, the strongest preferred for hard requests.

Models that read images as well as text, cheapest first, strongest for hard requests.

Hosted open-weights models only - the same stack runs on your own hardware later.

Anything with personal data is pinned to an on-prem model; everything else uses the cloud tiers.

Cheap by default, frontier only for genuinely hard requests, with a monthly budget objective.

Route support traffic to the cheapest capable model; anything with personal data stays on-prem.

Three models, two rules: quality first, private data stays on-prem. Apply in a minute.