Harvard paper: 94.8% of coding agent tokens are cached re-reads, 25x energy gap
rohanpaul_ai · x · 2026-11-18
A new Harvard paper logged 10 weeks of real Claude Code use by one engineer—about 1 billion tokens across 42 sessions—and found 94.8% of the text an agent processes is re-reading cached earlier context, not new work.
Key findings:
- The authors built a cache-aware energy estimator that charges cached tokens for attention reads but not full reprocessing; conventional estimators like EcoLogits price single request-response calls and miss agents' repeated context reads and subagent handoffs.
- About 260 cached tokens were read per generated token, and context grew 5.3x within a session.
- The cache-aware estimate came to 250 kWh over 10 weeks versus 10 kWh from the output-based estimator—a roughly 25x gap.
Practical advice: when tracking agent cost or emissions, count cached reads and subagent calls, and start new tasks in fresh sessions where possible.
Related event: Harvard Paper: 94.8% of Coding Agent Tokens Are Cache Re-Reads(2 posts)→
More from Infra
- Bloomberg: New US Chip Rule Freezes $480M Nvidia GPU Loan for ByteDance in Malaysia — rohanpaul_ai · 2026-11-18
- AI raises the cost of weak cloud architecture, says resilience expert — DavidLinthicum · 2026-11-18
- French team drops vLLM/PyTorch for Rust+CUDA and FP4/FP8, huge single-GPU inference gains — IgorCarron · 2026-11-18
- LlamAmpere v0.7: Qwen 27B at Full Context on 12GB Cards, 70 TPS — Brief-Tap-6616 · 2026-11-18
- Qwen4-Max lands on DeepInfra with 1M context at $1.413/M input tokens — gharik · 2026-11-18
- TensorScale Visualizes Every Inference Stage and Benchmarks Against a Leading Hosting Provider — TerryTangYuan · 2026-11-18