We tracked every dollar spent on 4 production agentic AI systems over 6 months (October 2025 - April 2026). This article publishes the full cost breakdown: API spend per model, infrastructure costs, cost-per-execution metrics, latency data, and the optimizations that cut costs by 40-70% without reducing quality.

Why We Published This Data

When we started building agentic AI systems for clients, the most common question was: "What will this cost to run every month?" We couldn't find a single source with real production numbers. Every article was either theoretical or based on toy benchmarks. So we tracked everything.

The Four Systems We Tracked

System A - Support Triage: Single agent with 3 tools, built on LangGraph. 2.4 average steps per run, 12,000 runs/month. Uses Claude Haiku.

System B - Document Processor: Sequential chain with 5 tools, built on LangGraph. 4.8 average steps per run, 8,500 runs/month. Uses GPT-4o.

System C - Sales Research Crew: Multi-agent system (3 agents), built on CrewAI. 8.2 average steps per run, 3,200 runs/month. Uses Claude Sonnet + Haiku.

System D - Code Review Agent: Single agent with 7 tools, custom framework. 6.1 average steps per run, 5,800 runs/month. Uses GPT-4o.

Monthly API Cost Breakdown

LLM API calls account for 60-80% of total operating cost across all four systems.

System A (Support Triage): $0.018/run, $216/month (Claude Haiku, 3,200 tokens/run)
System B (Document Processor): $0.14/run, $1,190/month (GPT-4o, 12,400 tokens/run)
System C (Sales Research Crew): $0.31/run, $992/month (Sonnet + Haiku, 28,600 tokens/run)
System D (Code Review Agent): $0.22/run, $1,276/month (GPT-4o, 18,900 tokens/run)

Key insight: System A processes 3.7x more runs than System C but costs 78% less per month. Model choice x step count is the cost formula.

Infrastructure Costs

Compute (ECS/EKS): $180-$420/month
Cache (Redis/ElastiCache): $55-$110/month
Database (PostgreSQL/RDS): $60-$180/month
Observability (LangSmith/Datadog): $120-$350/month
Queue (SQS): $5-$15/month
Total infrastructure per system: $420-$1,075/month

Total Monthly Cost (All-In)

System A (Support Triage): $636/month total, $0.053/run all-in
System B (Document Processor): $1,770/month total, $0.208/run all-in
System C (Sales Research Crew): $1,642/month total, $0.513/run all-in
System D (Code Review Agent): $1,996/month total, $0.344/run all-in

The range: $636 to $1,996/month for production agentic AI systems handling 3,200-12,000 runs/month. That's $0.05-$0.51 per agent execution all-in.

Latency Data

System A: P50 1.8s, P95 3.2s, P99 5.1s
System B: P50 8.4s, P95 14.2s, P99 22.8s
System C: P50 18.6s, P95 32.4s, P99 48.1s
System D: P50 12.1s, P95 21.8s, P99 35.2s

Latency scales roughly linearly with agent steps. For user-facing agents, keep it under 4 steps or use streaming.

Three Optimizations That Cut Costs 40-70%

1. Model Routing: Cheap Models for Simple Steps (35-55% savings)

Route simple classification steps to Haiku/GPT-4o-mini, reserve expensive models for complex reasoning. System B went from $1,190 to $620/month in API costs - a 48% reduction with no measurable quality loss.

2. Tool Result Caching (15-35% savings)

Redis cache with 1-hour TTL for tool results. System A saw 35% cache hit rate on knowledge base lookups, reducing average steps from 2.4 to 1.8 and cutting API costs by 25%.

3. Step Limits + Early Exit (10-20% savings)

Max_steps hard limit plus early exit on high confidence. System C's P99 latency dropped from 68s to 48s, monthly API cost dropped by $180 (18%).

Post-Optimization Results

System A: $636 -> $502/month (21% savings)
System B: $1,770 -> $980/month (45% savings)
System C: $1,642 -> $890/month (46% savings)
System D: $1,996 -> $1,180/month (41% savings)

Cost Projection Formula

Monthly Cost = (runs/month x avg_steps x avg_tokens_per_step x model_price_per_token) + infra_base

Example: 10,000 runs/month x 4 steps x 5,000 tokens x $0.000003/token (Claude Haiku) + $500 infra = $1,100/month

FAQ

How much does it cost to run an AI agent in production?

Based on our data, a single agentic AI workflow costs $0.02-$0.35 per execution. A simple 2-tool agent using Claude Haiku costs ~$0.02/run. A complex multi-agent system costs $0.15-$0.35/run. At 10,000 executions/month, expect $200-$3,500/month in API costs alone.

What's the biggest cost driver for agentic AI?

LLM API calls account for 60-80% of total cost. The number of agent steps is the primary multiplier. Model choice is the second biggest factor: GPT-4o costs 5-8x more than Claude Haiku.

How do you reduce agentic AI costs without losing quality?

Three strategies: (1) Route simple tasks to cheaper models. (2) Cache frequent tool call results. (3) Limit max agent steps. These cut costs 40-70% without quality tradeoffs.

Is it cheaper to use CrewAI or LangGraph?

The framework doesn't meaningfully affect cost. Both make the same API calls. Cost difference comes from architecture: number of steps, model selection, and caching strategy.

What infrastructure do you need for production AI agents?

Minimum: 2-node Kubernetes cluster ($150-$400/month), Redis ($50-$100/month), PostgreSQL ($50-$200/month), observability stack ($100-$300/month). Total infra: $350-$1,000/month excluding API costs.

The Real Cost of Running Agentic AI in Production: 6 Months of Data