Memory goes everywhere, Mythos 5 powers Claude Security, Claude Code gets restricted mode

Anthropic shipped the biggest memory update since the architecture rebuild in July. Memory now works across both chat and Claude Cowork in the cloud — Claude can pick up where users left off across tasks and full conversations, not just chat threads. Everything Claude remembers is listed under Topics in Settings > Memory, where users can edit or delete any entry individually. A new sensitive topics setting keeps topics like health, beliefs, and relationships out of memory unless you explicitly turn on Include sensitive topics in memory. Memory is on by default for Free, Pro, and Max plans and off by default for Team and Enterprise organizations, where admins control the setting.

Claude for Teachers, which launched in July for individual verified US K-12 educators, expanded this week to institutional access for schools and districts. School administrators can now provision Claude for Teachers for entire campuses or districts through a dedicated admin flow, giving all verified educators in the organization access to the Learning Commons connector and standards-aligned teaching skills without each teacher needing to verify individually. The institutional tier also includes usage dashboards for school IT admins.

Claude Security scans now run on Claude Mythos 5 for all Claude Enterprise customers — the first time the restricted Mythos-class model has been available through a self-serve product without Glasswing membership. Enable in the admin console, connect a GitHub repository, and Mythos 5 traces data flows across files and Git history — not just pattern matching — and returns each finding with a CWE category, confidence rating, severity rating, and suggested patch for human review. Results are scoped: the model cannot be steered into writing exploits from the scan interface, only defensive outputs surface. Scans are billed as standard token usage under the existing Enterprise plan with no separate add-on.

Alongside the Claude Security expansion, Anthropic launched the Defender Advantage Fund — 0xDAF — committing $35 million in Claude credits to organizations working on open-source software security. Grants prioritize three areas: patching live vulnerabilities in widely used projects, building scanning and patching automation other projects can replicate, and pursuing security approaches that make projects resistant to entire classes of attack. Anthropic is starting with a small number of larger pilot grants before scaling. Applications are open at anthropic.com/0xdaf.

The Cyber Verification Program, which gives vetted defenders reduced safeguards for authorized security research, will expand in coming weeks to cover broader dual-use capabilities on Claude Opus and Sonnet — with Mythos-class access to follow for the most vetted organizations. Organizations already enrolled need take no action; Anthropic will reach out directly. Separately, Anthropic is widening Mythos 5 access through Project Glasswing for operators of critical infrastructure that meet strict security control requirements, building on the ~200-partner expansion announced in June.

A dense Claude Code release: --restricted (or CLAUDE_CODE_RESTRICTED=1) removes built-in tools that run commands or code and WebFetch, keeps file tools inside the working directory, refuses bypassPermissions, and ignores user and project settings files — a clean isolation mode for untrusted environments. Cross-session messaging via SendMessage and ListAgents now works between sessions. /usage-credits lets Enterprise members on AWS Marketplace, self-serve Enterprise, and Enterprise trials request higher usage limits directly from their admin. experimental.cacheTtl in agent frontmatter sets a per-agent prompt cache TTL when no subagent TTL is configured. The /web-setup flow now warns when the GitHub CLI token lacks the workflow scope needed for large repository pushes.

The 50% Claude Code weekly usage limit boost that has been in place since May 13, 2026 — extended multiple times, most recently to August 31 — expires tonight. Anthropic cited strong demand and ongoing data center capacity expansions as reasons for the repeated extensions but confirmed August 31 is the final date. Standard limits return from September 1 for Pro, Max, Team, and Enterprise plans. If your team's workflows were shaped around the elevated limits, audit typical weekly usage today to identify what to prioritize or route to the API.

Claude Sonnet 5's introductory pricing of $2 per million input tokens and $10 per million output tokens expires tonight, August 31, 2026. Standard pricing of $3/$15 per MTok takes effect September 1 — unchanged from Sonnet 4.6 in list price, but remember Sonnet 5's updated tokenizer maps the same content to 1.0–1.35x more tokens than Sonnet 4.6, so the effective per-task cost increase will be larger than the 50% list-price rise alone suggests. Batch API pricing rises from $1/$5 to $1.50/$7.50. Prompt cache write pricing rises from $0.25 to $0.375, and read hits from $0.20 to $0.30.