Claude Code Just Got 79x Leaner. The Model Never Moved.
The July changelog shipped a 79x cut in transcript size and 7x faster tool rounds — with zero model upgrade. The gains came from the runtime, which is exactly where production lives.
NeuroX AI · July 20, 2026
Anthropic's July Claude Code changelog is worth reading for one reason: none of the biggest wins involve a new model. Session transcript size dropped up to 79x in edit-heavy sessions, and per-tool-call CPU overhead fell far enough to deliver up to 7x faster tool rounds at high tool counts. Same weights. Same benchmarks. A dramatically cheaper, faster runtime.
The mechanism is unglamorous, which is the point. The 79x came from pruning superseded file-history backups instead of hoarding every checkpoint. The 7x came from caching the tool-pool assembly instead of rebuilding it on every single call. No inference breakthrough — just an agent runtime that stopped wasting work it had already done.
This is the number most teams never look at. A long-running agent's cost isn't only tokens through the model; it's the transcript it drags forward, the tool schemas it reassembles every turn, the checkpoints it never garbage-collects. On a session with hundreds of tool calls, that overhead is the bill. Anthropic just proved you can cut it by an order of magnitude without touching the model.
We keep saying it because it keeps being true: the model is a component. The runtime around it — what it remembers, what it rebuilds, what it throws away — is the system you actually ship.