Model provider

A model provider is the service that runs inference for you. Usually that's the lab that trained the model (Anthropic, OpenAI, Google), sometimes a cloud that resells it (Amazon Bedrock, Google Vertex, Microsoft Azure), and sometimes your own hardware running an open-weight model through Ollama or llama.cpp. The harness sends each model provider request to the provider and gets tokens back.

The distinction matters because the provider, not the harness, sets the limits you bump into: price per token, rate limits, the context window size, which features exist (prompt caching, extended thinking, batch discounts) and how long the prefix cache lives. When the same tool behaves differently on two accounts, check whether they're pointed at different providers or tiers.

Most coding harnesses let you swap the provider. Claude Code can route to Bedrock or Vertex for teams that already pay a cloud. Codex and Cursor can be pointed at other endpoints or your own keys. Local providers give you privacy and no per-token bill, at the cost of weaker models. Pick on capability first; a cheap provider that can't finish the task isn't cheap.