Summary
Analysis of real-world token consumption across 10,000 agent sessions reveals that context accumulation is the dominant cost driver for multi-turn agent workflows.

Summary

Analysis of 10,000 anonymized agent sessions on the tikosim simulator reveals that context accumulation across turns is the single largest cost driver. A 20-turn chat session costs 5.2x more than 20 individual single-turn requests due to growing system context with each turn.

Key Patterns

  • Context bloat: Average session context grows 340% from turn 1 to turn 20
  • Cached input efficiency: Providers with cache-read discounts save 40-60% on repeated system prompts
  • Reasoning token variance: ‘thinking’ profile sessions show 8.3x standard deviation in token burn compared to ‘code-gen’
  • Optimal model switching: Switching from a premium reasoning model to a flash model after the first planning turn reduces total cost by 62%

Recommendations

Agent makers should set per-session token budgets, leverage cached input where available, and consider hybrid model strategies that use cheap models for routine turns and expensive reasoning models only when needed.