How do you choose appropriate models for different agents?

What you need to know

Models differ a lot in price and speed. A small model can cost 10–20 times less per token than a frontier model and answer faster. Using the big model everywhere is simple but wasteful; using the small one everywhere often fails on the hard steps.

Where CrewAI lets you set a model

1from crewai import Agent, Crew, LLM23BIG = LLM(model="openai/gpt-4.1", temperature=0.2)          # example ids4SMALL = LLM(model="openai/gpt-4.1-mini", temperature=0)56classifier = Agent(role="Ticket Classifier", goal="...", backstory="...",7                   llm=SMALL)8investigator = Agent(role="Escalation Investigator", goal="...", backstory="...",9                     llm=BIG, tools=[crm_lookup, order_lookup],10                     function_calling_llm=SMALL)   # cheaper model picks tool args11writer = Agent(role="Reply Writer", goal="...", backstory="...", llm=SMALL)1213crew = Crew(agents=[classifier, investigator, writer], tasks=tasks,14            planning=False)

• llm on an agent — the model for its reasoning.
• function_calling_llm on an agent or crew — the model that formats tool calls, if different.
• manager_llm — the manager in a hierarchical crew.
• planning_llm — the AgentPlanner when planning=True.

Always use the provider prefix (openai/, anthropic/, gemini/) so CrewAI routes to the right provider.

A starting split

Then measure, one change at a time

Change one agent's model, run the golden set, record score and cost, and keep it only if the score holds. Changing several at once hides which change hurt.

A real-life example

An e-commerce company's customer-support escalation crew runs about 3,000 escalations a day. It started with the strong model on all three agents, at about ₹9 per ticket.

The team tried the small model on each agent in turn:

• Classifier: accuracy 97% vs 98%. Switched.
• Writer: tone and accuracy scores the same. Switched.
• Investigator: correct root cause 71% vs 89%. Kept the strong model, but set a small function_calling_llm.

Cost fell to about ₹3.50 per ticket, roughly ₹16,000 saved per day, with no measurable quality loss. Surprise: the writer, which they expected to need the big model, did not; the investigator, which "only looks things up", did.

Follow-up questions to expect

• "Would you use a reasoning model?" — For the investigator or planner, possibly, if the golden set shows a gain. They are slower and cost more, so not for simple steps.
• "Can you mix providers in one crew?" — Yes, each agent can use a different provider. Watch for different tool-calling behaviour and keep one tracing tool across them.
• "How do you avoid surprises when a provider updates a model?" — Pin dated model versions where the provider offers them and rerun the golden set before moving.