Neuralwatt data: your coding-agent harness decides how often the prompt cache breaks

Neuralwatt says harness design, not just the model, drives prompt-cache misses that can roughly double a request's inference energy.

Inference provider Neuralwatt published an analysis on October 16 arguing that the agent harness, not just the model, decides how much prefill work your sessions trigger. Across a month of traffic (September 8 to October 7, 2026), uncached input ranged from 1.6% to 11.1% by client. Neuralwatt says a mostly uncached request uses roughly twice the energy of a cached one of the same size, closer to 2.5 times on GLM-5.3. Past 128k tokens, one turn that re-reads more than half its prompt costs as much prefill as about 600 normal turns.

The break rate per turn was about 1 in 100 for deepseek-harness against about 3 in 100 for OpenCode and Pi. In Neuralwatt's lab tests, every cache break came from a harness editing content it had already sent. Examples include OpenCode 1.x's plan-mode reminder (fixed in 2.x), omp pruning re-read files (Neuralwatt says setting compaction.supersedeReads: false and compaction.dropUseless: false keeps the cache), dates written into the system prompt, mid-session effort switches and MCP servers connected mid-session.

The practical takeaway matches Claude Code's own caching docs. Those docs list switching models, changing effort, connecting an MCP server and compacting as actions that can invalidate the cache. Make those changes early in a session, not at turn 200.

Caveats: the energy figures come from one provider's cloud, mostly on open-weight models. Vibe Coding Weekly notes that Neuralwatt has a commercial interest as an inference provider. RuntimeWire frames the findings as needing to hold across other deployments.

Sources