Articles · Models & providers
OpenAI Responses vs Anthropic Messages: how to pick
Both wires are mature. The differences live in tool-call patterns, streaming preference, and pricing band.
Kition Team·Apr 23, 2026·1 min read
API shape comparison
- Responses: single endpoint, input is a messages array, tool calls are items
- Messages: single endpoint, content is a parts array, tool calls are tool_use parts
- Both stream, both multimodal, both with hosted tools
Picking guidance
- Want Claude → anthropic_messages
- Mostly GPT with a fallback chain → responses
- Aggregator / reverse-proxy services → responses (their proxies expose this wire)
- China-hosted models → most are responses-compatible
Related reading
Models & providers
Driving the Kition agent with Claude 4.7
Claude is consistently strong on long context, tool use, and alignment. A solid default for the Kition agent.
Models & providers
Driving the Kition agent with GPT-5
GPT-5 is rock-solid on structured outputs and multi-step function calls. A great fit for OpenAI-native workflows.