Why this is in the vault
Anthropic is permanently changing Claude Code's weekly usage limits effective September 14 — a temporary 50% boost expires and the new permanent baseline is only 25% above the pre-boost floor, which reads as a 17% cut from what sessions feel like right now.
Mapping against Ray Data Co
Ray (this COO instance) runs on Claude Code continuously — cron skills, always-on channel monitoring, cert study, vault ingestion, investing research fleets. A 17% drop in felt weekly capacity starting September 14 is a direct operational constraint: session-budget planning for heavy fan-out work (deep-research nightly runs, family-research-round, /check-board) should assume tighter headroom starting mid-September, not the current temporary-boost baseline. Worth a light pre-mortem before the 14th on which recurring skills are most session-hungry, so nothing silently starts failing mid-run when the ceiling drops. Secondary relevance: the Grok Bot item (Stripe Link single-use virtual cards, per-purchase approval, merchant/amount/budget rules) is a live analogue to [[project_rdco_spending_authority]] — RDCO's own <$20/mo autonomous spend + founder-gate-above threshold is the same shape of problem (agent transacts, human sets rails) that xAI just shipped for consumer shopping.
Curation section
- Anthropic raises Claude Code weekly limits 25% permanently (Sept 14) — but a 50% temporary boost expires first, so the felt change from today's level is a 17% reduction. Applies to Pro, Max, Team, and seat-based Enterprise. Anthropic says better usage-visibility tooling is coming.
- Claude Code desktop app can resume CLI sessions —
/resumein the desktop app picks up any terminal session with full conversation history intact; lets you start in terminal and finish in the GUI. - Grok Bot can shop and pay online — Stripe Link plugin issues single-use virtual cards per purchase, agent never sees the real card number, human approves each spend (merchant/amount shown), configurable site allowlist and budget caps. US only, mobile approval coming.
- Signals roundup: Anthropic says Claude can autonomously fix misalignment in models up to 4.7x larger than itself; Pangram Labs' AI text detector claims a 99.99% true-negative rate; a new video pretraining method reportedly matches Meta's V-JEPA 2 at up to 20x less compute; Toby Ord published a paper arguing a true intelligence explosion is mathematically harder than commonly assumed; a group of fired developers built an open-source "AI CEO" running on 8 Claude agents.
⚠️ Sponsorship
Three paid placements inside this issue: OpenRouter ("one API for every model" — unified multi-provider routing, direct alternative to calling providers individually), Teleport (agent session observability/risk-scoring, webinar plug), and Granola (AI notetaking, Apple Watch feature push). None of the three sponsors are the subject of the Claude Code limits story or the Grok Bot story — the editorial content and the ad placements are cleanly separated in this issue, no detected bias in how the Anthropic story is framed.
Related
- [[2026-08-10-alphasignal-claude-code-cross-session-auto-mode]]
- [[2026-07-04-anthropic-max-plan-tos-productized-agentic-use]]
- [[2026-08-13-alphasignal-grok46-deepseek-v4pro-claude-cowork]]
- [[project_rdco_spending_authority]]