Status check: cache TTL remediation commitments from #46829 - 30-day update

In closed issue #46829 (Cache TTL regressed from 1h to 5m, closed not_planned 2026-04-12), Anthropic engineering publicly committed to specific remediations. As of 2026-05-12 (~30 days later), shipping status is unclear from the outside. Filing this as a status check rather than a re-litigation.

Public commitments to verify

bcherny, 2026-04-13 in #46829:

1. "At this point we have rolled out 1h prompt cache by default in a number of places for subscribers"
2. "We will soon be changing the client side default to 1h for a few queries"
3. "We will also give you env vars to force 1h and 5m"

lizthegrey, 2026-05-08 (citing Anthropic in #46829): two separate Claude Code bugs fixed in v2.1.108 and earlier that compounded the regression independently of the TTL default change.

Specific questions

1. Env vars for TTL override. Have these shipped? If yes: variable name(s), accepted values, Claude Code version that introduced them, scope (per-process, per-session, settings.json equivalent). If not: target version / ETA.
2. Client-side default flip to 1h. Has the flip shipped for any queries? If yes: which query types, which subscriber tiers, Claude Code version. If not yet: which queries are in scope and target version.
3. v2.1.108 compounding bugs. Were these fixed and included in current Claude Code? A brief description of what was fixed would help users self-diagnose whether they're still affected.
4. Per-session control (optional). raghuvv proposed (2026-04-13 in #46829) per-session TTL control similar to model selection / plan mode. Is this being considered separately from the env-var path?

Why a status check

Remediation was publicly committed by Anthropic engineering; not asking for new feature work. Asking for shipping visibility so users can either (a) update and use new controls, or (b) plan around continued 5m default for another N weeks.

Context

Large auto-loaded CLAUDE.md (~19.5k tokens) compounded by cache-miss boundaries (session launch, idle >5min, mid-session edits to auto-loaded files). Per #46829 evidence: 20-32% cache-creation cost inflation under 5m TTL across 119,866 API calls.