Standardized reasoning_effort

What it does

reasoning_effort is a new standard parameter on the LangChain chat model interface (langchain-core 1.4.7+). Pass reasoning_effort=<int> to any supported chat model and it controls how much compute the model spends on its internal reasoning chain before answering. Lower values mean faster, cheaper responses; higher values mean longer thought chains and better quality on complex tasks. When unset, the model uses its built-in default.

Why it matters

Before this change every provider had a proprietary knob: xAI Grok used its own parameter, Anthropic had model-specific knobs, OpenAI used o_reasoning_effort, and Fireworks had fireworks_reasoning. Multi-model apps had to write a different code path per vendor. With reasoning_effort standardized, you write one code path and swap the model string to get consistent control over reasoning depth across Grok, Claude, GPT-4o, and Fireworks.

Example

$ grok_fast = init_chat_model("xai:grok-3", temperature=0)
print(grok_fast.invoke("What is 17 × 23? Answer in one sentence.", reasoning_effort=500).content)
claude_deep = init_chat_model("anthropic:claude-sonnet-4-20250514", temperature=0)
print(claude_deep.invoke("What is 17 × 23? Answer in one sentence.", reasoning_effort=4000).content)

=== Fast (effort=500) ===
17 × 23 = 391.

=== Deep (effort=4000) ===
17 × 23 = 391. The calculation follows from the distributive property:
17 × 23 = 17 × (20 + 3) = (17 × 20) + (17 × 3) = 340 + 51 = 391.

Common flags

reasoning_effort=<int>

Standard chat-model param controlling internal reasoning compute

init_chat_model("provider:model")

Universal entry point to initialize any supported chat model

BaseChatModel.invoke()

Standard streaming-agnostic invocation accepting reasoning_effort