Why OpenAI, Anthropic, Gemini, and Ollama Prompt Formats Differ

Published March 16, 2026 · Updated March 17, 2026 · 3 min read

OpenAI, Anthropic, Gemini, and Ollama all use conversational message formats for their APIs, but the structures are different enough that a prompt formatted for one provider won't work directly with another. The differences are intentional — each provider made independent decisions about how to structure multi-turn conversations, system instructions, and role names.

OpenAI format

OpenAI's chat completions API uses a flat messages array where system, user, and assistant messages are all peers:

{
  "model": "gpt-4o",
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is JSON?"},
    {"role": "assistant", "content": "JSON is a text format..."},
    {"role": "user", "content": "Give me an example."}
  ]
}

Anthropic format

Anthropic separates the system prompt from the conversation history — the system prompt is a top-level parameter, not a message:

{
  "model": "claude-3-5-sonnet-20241022",
  "system": "You are a helpful assistant.",
  "messages": [
    {"role": "user", "content": "What is JSON?"},
    {"role": "assistant", "content": "JSON is a text format..."},
    {"role": "user", "content": "Give me an example."}
  ]
}

Gemini format

Gemini uses a contents array instead of messages, wraps content in a parts array, uses model instead of assistant for the AI role, and separates the system instruction as a dedicated top-level field:

{
  "system_instruction": {"parts": [{"text": "You are a helpful assistant."}]},
  "contents": [
    {"role": "user", "parts": [{"text": "What is JSON?"}]},
    {"role": "model", "parts": [{"text": "JSON is a text format..."}]},
    {"role": "user", "parts": [{"text": "Give me an example."}]}
  ]
}

Ollama format

Ollama's API is designed to be OpenAI-compatible — the structure is similar to OpenAI's, with messages array and the same role names. The main difference is the model field references a local model name rather than an OpenAI model identifier:

{"model": "llama3.2", "messages": [...]}

System prompt support varies by model. Llama 3 handles system messages in the messages array well. Some older models work better with the system prompt embedded in the first user message using the model's instruction format.

Why these differences exist: each provider built their API independently, making different design choices. OpenAI was first; others built compatible-but-different formats. There's no single standard for conversational AI APIs, though OpenAI's format has become an informal reference point that Ollama explicitly targets for compatibility.