Models

Every request is routed through the gateway, which picks a model from the free pool or the premium catalog — and falls back across candidates when one fails.

Auto (default)

With no model selected, the gateway resolves the plan's default model — every tier defaults to nvidia/nemotron-3-ultra-550b-a55b:free. The router then builds a candidate chain: requested model → openrouter/free load balancer → the rest of the free pool, skipping models whose circuit breaker is open.

Override

Use the model dropdown in the Brody AI window to pin a specific model — premium models are available on Pro and above. An empty override ("") means auto. Your chosen model and key are forwarded to the gateway so the selection is honored server-side.

Tip: if a model keeps failing, the router falls through the pool automatically — a pinned model never blocks your session.

Tool-call pinning

Native function calling requires a tool-capable model. When a request includes tools and the resolved model is a prose-only one (free load balancers, Nemotron, reasoning models), the router silently swaps to cohere/north-mini-code:free — the live-verified tool-caller — so the 76-tool agent always works, even on the free tier.

Free pool

Curated from the OpenRouter catalog — $0 at the provider. Overridable at gateway boot viaOPENROUTER_FREE_MODEL_POOL_CSV.

ModelNote
openai/gpt-4o-miniCheapest paid-tier quality/price leader; leads the fallback pool.
meta-llama/llama-3.3-70b-instructStrong reasoning at low cost.
openrouter/freeOfficial OpenRouter auto free-load balancer.
nvidia/nemotron-3-ultra-550b-a55b:freeDefault plan model for every tier (550B).
cohere/north-mini-code:freeOptimized for C# / Unity code generation — emits real tool calls.
google/gemma-4-31b-it:freeVision + tools capable.
nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:freeReasoning engine.
nvidia/nemotron-nano-12b-v2-vl:freeVision and multimodal.
deepseek/deepseek-r1:freeDeepSeek R1 reasoning.
qwen/qwen-2.5-coder-32b-instruct:freeCoder-specialized.
mistralai/pixtral-12b:freeMultimodal / vision.
google/gemini-2.0-flash-exp:freeGemini 2.0 Flash experimental.
meta-llama/llama-3.1-405b-instruct:freeLargest Llama on the free pool.
inclusionai/ling-3.0-flash:freeFast flash model.
poolside/laguna-s-2.1:freeCode-oriented poolside model.

Premium models

Available on paid plans — provider cost per 1M tokens is on the pricing page.

ModelNote
anthropic/claude-3.5-sonnetFrontier coding model, direct Anthropic fallback supported.
anthropic/claude-3-5-haikuFast, cheap Claude.
openai/gpt-4oFrontier OpenAI model.
openai/gpt-4o-miniCheap and fast.
google/gemini-3.5-flashPaid safety net for tool calls.
google/gemini-3.6-flashLatest Gemini flash.
google/gemini-2.0-flash-001Gemini 2.0 Flash.
google/gemini-proGemini Pro.
deepseek/deepseek-chatCheap chat model.
deepseek/deepseek-r1DeepSeek R1 reasoning (paid).

Vision

Screenshot analysis (the analyze_screenshot tool and the /api/v1/vision/analyze endpoint) uses a dedicated vision pool, tried in order:

  1. google/gemma-4-31b-it:free — vision + tools, free
  2. cohere/north-mini-code:free — proven tool-caller
  3. nvidia/nemotron-nano-12b-v2-vl:free
  4. google/gemini-3.5-flash
  5. openai/gpt-4o-mini
  6. mistralai/pixtral-12b:free

Bring your own key (BYOK)

OpenAI

GPT models with your key.

Claude (Anthropic)

Claude 3.5 Sonnet / 3.5 Haiku with your key.

Gemini (Google)

Gemini models with your key.

DeepSeek

DeepSeek chat / R1 models.

Ollama

Local models on http://localhost:11434.

OpenRouter

One key, hundreds of models.

Groq

Fast inference provider.

Custom

Any OpenAI-compatible endpoint.

Keys are stored in editor preferences and forwarded per request, so your selected provider and model are honored end-to-end. With BYOK you pay the provider directly at their published rates.