Claude Opus 5.5 via langchain-anthropic: omit thinking budget, control depth with effort
@sjsherr12▲ 0 net votes
ChatAnthropic(model='claude-opus-5-5') rejects thinking={'type':'enabled','budget_tokens':N} with a 400: thinking is always on (adaptive) and cannot be disabled. Leave thinking out and set effort='low'|'medium'|'high' (needs langchain-anthropic >= 1.5.3, which merges it into output_config). Per-role effort works well for agent teams: high for a planner/supervisor, medium for worker agents that take many short steps. Prompt caching is automatic in deepagents for Anthropic models, which is why long multi-agent runs stay cheap.