# The 400K Cliff: A Context-Discipline Audit for Sonnet 5’s Always-On 1M Window

> Source: <https://pub.towardsai.net/the-400k-cliff-a-context-discipline-audit-for-sonnet-5s-always-on-1m-window-bbc1bf97f780?source=rss----98111c9905da---4>
> Published: 2026-07-07 05:59:12+00:00

Member-only story

# The 400K Cliff: A Context-Discipline Audit for Sonnet 5’s Always-On 1M Window

## Anthropic says the era of context rationing is over. The degradation data, the new tokenizer, and my own workflow audit say the discipline didn’t die — it moved from surviving the window to protecting the margin.

On June 30, Anthropic made a decision for you. When Claude Sonnet 5 shipped, the 1M-token context window stopped being a variant you opt into. On the API there is no 200K version of Sonnet 5, no `[1m]`

suffix, no usage credits to enable. Every session gets the full million, and Claude Code quietly auto-compacts around 967K if you ever get there.

Most people read that as the end of an era. The era of `/clear`

rituals, of `.claudeignore`

files pruned like bonsai, of restarting sessions every thirty minutes out of buffer anxiety. Anthropic's own launch material for 1M context, back when it went GA for the 4.6 models in March, framed the change exactly that way: the summarization tricks and context-clearing workarounds that long-context work used to require were no longer necessary.

Here is the contrarian claim, stated early so you can disagree efficiently: **the always-on million makes context discipline more valuable, not**…
