Uber Achieves 52% Reduction in Per-Session AI Agent Costs Through Cost Structure Analysis and Optimization

Uber achieved a 34% reduction in cost per 1,000 requests and a 52% reduction in cost per session by fixing the model for comparison, despite a 7x increase in weekly active users and a 9.4x increase in agent requests from February to August 2026
Using the formula Total Cost = Number of Users × Number of Sessions × Number of Turns × Number of Requests × Number of Tokens × Token Unit Price, reducing waste factors (turns, requests, tokens) by 20% each results in a 49% reduction in total cost
Separated orchestrator and sub-agent models, separated prompt cache TTLs, and optimized MCP tool schemas to eliminate context retransmission and wasted tokens
Built a system for real-time cost tracking and budget threshold alerts based on cost allocation tags, model call logging, and OpenTelemetry in Amazon Bedrock

Cost Optimization Strategies

Downgrading sub-agent models: Claude Code separates the Opus/Sonnet 5.5 orchestrator from the Haiku 5.5 sub-agent, while Codex separates GPT-6 Sol from Luna
Context management: Automatic compaction at 400,000 tokens, separation of 1-hour/5-minute cache TTLs, and enabling tool search to prevent 50,000 to 70,000 token waste when there are more than 100 tools
Strengthening grounding: Confirmed the importance of recording core context through cases where context graph queries yielded correct answers in 38 seconds, whereas the absence of a graph took 20 minutes and resulted in errors

Cost tracking: Cost allocation tags for application inference profiles, model call logging, CloudWatch real-time metrics, and four-source collection of OTel session data
Alert system: Slack notifications for 50/80/100% budget thresholds via integration of AWS Budgets and Amazon SNS, and surge detection using Cost Anomaly Detection
Caveats: 1-hour cache TTL doubles the input unit price, batch inference does not support tool calls, and Claude 5.5 models support only the Standard tier