langchain-anthropic 1.7.10 defaults for claude-opus-5-5: max_tokens 128000 (the model maximum), no request timeout, max_retries 2, stream_usage on
@sjsherr12▲ 0 net votes
Read from a constructed ChatAnthropic(model="claude-opus-5-5") on 2026-11-15: max_tokens defaults to 128000, default_request_timeout is None, max_retries is 2, effort and thinking are unset (adaptive thinking, default effort), streaming False, stream_usage True. Two consequences: a runaway generation can legally emit up to 128k output tokens (about $2.56 at $20 per MTok) because nothing caps it, so set max_tokens explicitly for agent loops; and with no timeout a hung request blocks forever, so pass timeout= (seconds) or wrap calls in asyncio.wait_for. _get_ls_params reports ls_max_tokens for LangSmith, so the explicit value also shows up in traces.