LLM API, Prompt Caching & Cloud FinOps
Empirical cost benchmarks, prompt cache amortization curves, and multi-turn agent token compounding data across Anthropic, Google, and AWS.
💰 Companion Utility: Agent Loop Cost Calculator
Model multi-turn token compounding and compare real monthly cloud bills across Gemini Flash, Claude Sonnet, and OpenAI.
- Token Economics
Why 15-Step Agent Loops Burn 400K Input Tokens (And 4 Config Changes That Cut Bills by 78%)
Compound context accumulation burns thousands of dollars on long turns. How subagents and tool pruning tame agent loop bills.
- Cache Benchmarks
Prompt Caching Break-Even Benchmarks: How Many Turns Before Cache Writes Pay Off?
Anthropic and Gemini charge a 25% write premium on cached prefixes. We benchmarked the exact turn count where caching beats on-demand.
- AWS Bedrock
AWS Bedrock Batch Inference vs. On-Demand: Real Latency and 50% Discount Cutoff Data
When does 24-hour SLA batch inference actually return? We stress-tested 10,000 synthetic queries to measure turnaround times and cost cutoffs.
- Model Routing
Multi-Model Routing Config: Pairing Cheap Flash Models for Scraping with Sonnet for Code Edits
Stop running 200k-token web scrapes through $15/M-token frontier models. Production routing blueprints using Gemini Flash + Claude Sonnet.
- Schema Audit
We Measured the Exact Input Token Cost of 10 Popular GitHub MCP Servers
Eager tool injection steals up to 22% of your agent's context window before you type a word. Token counts and schema pruning for 10 servers.