Back to full roadmap
topiccore
Parallel Tool Calls
Model can call N tools in one turn — latency drops N×.
2 hours1 resources1 prereqs
Old pattern: sequential tool calls (getWeather → wait → getNews → wait). New: parallel in one turn.
Claude: disable_parallel_tool_use: false by default. Model can return [tool_use, tool_use, tool_use].
OpenAI: parallel_tool_calls: true by default.
Benefit: "Give me weather for Istanbul + Ankara + Izmir" → 3 parallel calls, 1s total (sequential would be 3s).
Caveat: parallel tools must be independent. If one's output feeds another, they can't be parallel. The model usually figures this out.