Skip to content

cost-optimization.md

Latest commit

 

History

History
53 lines (35 loc) · 1.54 KB

cost-optimization.md

File metadata and controls

53 lines (35 loc) · 1.54 KB

Cost Optimization

Model selection

The biggest cost driver is model usage. Choose the right model for each agent:

Model Best for Relative cost
gpt-4.1 Complex reasoning, multi-step tasks High
gpt-4.1-mini Simple Q&A, classification, routing Low
gpt-4.1-nano Very simple tasks, extraction Very low 
config = AgentConfig(
    model="gpt-4.1-mini",       # Use mini for simple agents
    max_output_tokens=1024,     # Limit response length
)

The triage agent should use gpt-4.1-mini — it only routes, it doesn't reason.

Token limits

Set max_output_tokens to the minimum your agent needs:

  • Routing/classification: 512
  • Simple Q&A: 1024
  • Detailed explanations: 2048

Infrastructure

Foundry hosted (default, recommended)

With deployment_mode = "foundry" (default), Foundry manages the agent runtime. You pay only for model tokens — no Container Apps hosting cost.

Backing resources (Cosmos DB, AI Search, Storage Account) are provisioned by the AVM module and have minimal cost at low usage. This is the most cost-effective option for most workloads.

Container Apps (self-hosted)

With deployment_mode = "container_apps":

  • POC: Public endpoints, Basic ACR (~$10/month)
  • Enterprise: Private endpoints, Premium ACR, VNet (~$92/month)

Monitoring

Enable instrumentation to track token usage:

ENABLE_INSTRUMENTATION=true
ENABLE_CONSOLE_EXPORTERS=true

DevUI shows token usage per request. Use this to identify expensive agents.