Build a LangChain Agent That Costs ₹0.02 Per Run

Why agent cost matters

A single LangChain agent run can involve 3–8 LLM calls depending on tool use and self-correction loops. At GPT-4o prices, that adds up to ₹6–20 per run — fine for prototypes, painful at 10,000 runs per day.

The key insight is that most agent steps do not need a frontier model. Routing, tool selection, and parameter extraction are well within the capability of sub-₹1-per-million-token models.

Setup: connect LangChain to AICredits

pip install langchain langchain-openai

from langchain_openai import ChatOpenAI
 
# Use any model via AICredits — change only base_url and api_key
llm = ChatOpenAI(
    model="anthropic/claude-3-5-haiku-20241022",
    base_url="https://api.aicredits.in/v1",
    api_key="sk-your-aicredits-key",
    temperature=0,
)

The rest of your LangChain code is unchanged. Tools, chains, agents, memory — all work as normal.

Choosing the right model for each agent step

Use Claude 3.5 Haiku or GPT-4o Mini for tool selection and parameter extraction — these steps are structured and predictable. Use Claude 3.5 Sonnet or GPT-4o only for the final synthesis step where response quality visibly matters to your user.

from langchain.agents import create_tool_calling_agent, AgentExecutor
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.tools import tool
 
# Cheap model for tool selection steps
cheap_llm = ChatOpenAI(
    model="openai/gpt-4o-mini",
    base_url="https://api.aicredits.in/v1",
    api_key="sk-your-aicredits-key",
)
 
# Better model for final answer synthesis
smart_llm = ChatOpenAI(
    model="anthropic/claude-3-5-haiku-20241022",
    base_url="https://api.aicredits.in/v1",
    api_key="sk-your-aicredits-key",
)
 
@tool
def get_weather(city: str) -> str:
    """Get current weather for a city."""
    return f"Weather in {city}: 28°C, partly cloudy"
 
@tool
def calculate(expression: str) -> str:
    """Evaluate a mathematical expression."""
    return str(eval(expression))  # simplified example
 
prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful assistant. Use tools when needed."),
    ("human", "{input}"),
    ("placeholder", "{agent_scratchpad}"),
])
 
agent = create_tool_calling_agent(cheap_llm, [get_weather, calculate], prompt)
executor = AgentExecutor(agent=agent, tools=[get_weather, calculate], verbose=True)
 
result = executor.invoke({"input": "What is the weather in Mumbai? Also, what is 15% of 2400?"})
print(result["output"])

Hitting ₹0.02 per run with the right model

For many agentic tasks — web search summarisation, document Q&A, simple data extraction — GPT-4o Mini handles all steps adequately. At $0.15 per million input tokens (₹14.4 in INR), a 5-step agent with 300 tokens per call costs roughly:

5 steps × 300 tokens × ₹14.4 / 1,000,000 = ₹0.02 per agent run

The pattern: profile your agent on 50 sample inputs, measure task success rate with GPT-4o Mini, and only escalate individual failing steps to a more capable model. Most tasks land between ₹0.02 and ₹0.20 per run.

Cost tracking per agent run

AICredits logs every LLM call with model, token counts, and INR cost. Export a CSV from the Usage tab in the dashboard to see per-request breakdowns.

Tag your requests by passing a unique identifier in the user field of the API call. This lets you filter the usage CSV by agent type and attribute cost to specific product features.