Concepts
Models, providers and cost
Providers and models
A provider is a company or server that runs AI models: Anthropic, OpenAI, OpenRouter, DeepSeek, Google Gemini, xAI, Groq, Mistral, any server that speaks the OpenAI API, or Ollama on your own computer. A model is one of the models that provider offers.
Every agent has its own model, so you can give the Orchestrator a strong model for planning and a cheaper or local one to an agent that does simple, repetitive work. Each agent also has an Effort setting: more effort thinks longer, which helps on hard work and costs more.
When you add a provider's key in Settings > Models, the dashboard tests it, keeps it in data/, and gives it to every agent's profile. Hermes reads the key at every turn, so a replaced key works from the next turn on, with no new conversation. See Add a model provider.
The model catalog
Every model picker lists each provider's models under the provider's name. Each option shows the model's context window (how much text it can take in at once), its price in and out per million tokens, and anything it cannot do, such as "no tools". Agents need a model that can call tools to use skills, apps and files.
Refresh catalog in Settings > Models fetches the lists again; otherwise they are kept for 10 minutes.
Fallback
A fallback model takes over when an agent's model fails: the turn tries again as many times as you set, then once on the fallback. A fallback from another provider also covers that provider being down. See Choose each agent's model.
What a turn costs
A turn's cost depends on the tokens it reads and writes. Every turn reads more than your message:
- the agent's identity, your brief and its memory;
- the list of its skills, and its loaded tools;
- the conversation so far.
So a long conversation costs more per message than a new one, and an agent with many tools loaded costs more per turn than one with few. Skills and tools > Built-in tools shows how many tokens each agent's loaded tools add, and moving rarely used tools to Tools found when needed saves them on every turn.
Settings > Usage adds it all up. Costs there are estimates from Hermes (or, for a model Hermes has no price for, from the catalog); turns on a local model are $0.00. Your provider's bill is the true figure, and it can be higher: a failed turn counts as $0.00 in Usage. See Track usage and cost.
Local models
Ollama on your own computer is a provider like the others, with no key and no cost. Local models are slower and need a 64,000-token context window to work as agents. See Run agents on local models with Ollama.