Claude Code Daily Briefing - 2026-07-13

Release Summary

No new release as of 7/13 — the latest version is v2.1.207 (7/10). Today the calendar itself is the news: (1) the +50% weekly limit promotion ends today and included Fable 5 access ended yesterday, two limit changes landing at once (Security & Limitations), (2) the Microsoft 365 connector gained write tools (announced 7/7, not previously covered here), moving connectors from reading to acting (Features), and (3) Anthropic’s J-space interpretability research spread widely over the weekend (Community).

Full release notes

New Features & Practical Usage

Microsoft 365 connector gets write tools — Claude sends email and creates SharePoint files (7/7)

Anthropic added write tools to the Microsoft 365 connector (announced 7/7 — the same day as the Cowork web/mobile expansion, which is why it slipped past earlier briefings). Beyond reading (search, summarize), Claude can now act.

• Outlook: draft and send email, create and manage calendar events.
• OneDrive & SharePoint: create and update files.
• Prerequisite: Microsoft Entra admin consent is required — this isn’t a personal toggle; an org administrator has to allow it.

# Prerequisite: org admin grants admin consent in Microsoft Entra
Settings → Connectors → Microsoft 365
→ beyond search/summarize (read): send email, manage calendar, create files (write)

The takeaway: a connector just got promoted from a read channel to an output channel — “summarize this meeting” extends one step into “summarize it and email the team.” But the moment an output channel exists, the GitLost lesson from 7/9 applies verbatim: organizations need to explicitly design the boundary between what the agent can read and what it can send out. Requiring Entra admin consent effectively forces that boundary to be an administrator’s decision — decide which groups get write access before you roll it out. Claude release notes · Releasebot

Developer Workflow Tips

Use plan mode as a step gate — implement step one, then stop

Using plan mode (Shift+Tab) to draft a plan is common practice by now; the next step that this week’s updated field guides converge on is gating the plan’s execution itself.

• Edit the plan directly: open the proposed plan in your editor with Ctrl+G, reorder steps, delete the unnecessary ones, then approve — faster and more precise than a chat round trip asking for revisions.
• Run one step at a time: for anything big, don’t let the whole plan run unattended. “Implement step one only, then stop and show me” → review → “continue.” It feels slower; it’s faster, because you catch a wrong turn at step one instead of step six.

Shift+Tab   # plan mode — read-only, plan without touching disk
Ctrl+G      # open the proposed plan in your editor and edit it directly
"Implement step 1 only, then stop and show me the diff"   # step gate
"Looks good, continue"                                     # next step after review

This sits in the same place as 7/12’s pick models per task type — but for execution sizing: model choice decides who does the work, step gating decides how much to hand over at once. Planting human review points at plan boundaries is the cheapest way to cap the cost of unattended failure. Iwo Szapar — 8 Rules · Builder.io — 50 Tips

Secret hygiene — seal credential files off from the agent with deny rules

As agents gain write powers (sending email, creating files — see the M365 item above), managing files that must not be read stops being etiquette and becomes a defense line.

• Keep keys out of chat and CLAUDE.md: manage API keys and tokens via environment variables, and keep .env in .gitignore — a secret that lands in context gets replicated into transcripts, logs, and subagents.
• Block mechanically with deny rules: don’t rely on habit; make credential reads impossible via permissions.

// .claude/settings.json — seal off credential files
{
  "permissions": {
    "deny": [
      "Read(./.env)",
      "Read(./.env.*)",
      "Read(./secrets/**)"
    ]
  }
}

Same principle as 7/10’s rules that must always hold belong in hooks, not prose: a “don’t read secrets” sentence in CLAUDE.md is advice; a deny rule is enforcement. If your repo runs third-party skills or plugins (see SkillCloak, 7/8), this seal is worth double. Official best practices · Iwo Szapar

Security & Limitations

The +50% weekly limit promotion ends today — what reverts, what stays (7/13)

The Claude Code +50% weekly limit promotion, live since 5/13 for all Pro, Max, Team, and seat-based Enterprise users, ends today (7/13) at 6PM PDT. No extension has been announced.

• What reverts: weekly limits return to pre-promo levels — Max 5x from ~75 to ~50 hours/week, Max 20x from ~300 to ~200, with proportional reductions reported for Pro, Team, and Enterprise.
• What stays: the permanent May 6 improvements are untouched — doubled 5-hour session caps, peak-hour throttling removed for Pro/Max, and raised Opus API rate limits are separate from the promo.
• Action items: (1) reschedule heavy work — large refactors, test generation, migrations — against the smaller weekly budget, (2) update cost models for projects spanning July–August, and (3) the 7/12 tips (per-task model routing, /doctor context audits) double as token diet tools.

The takeaway: if two months of bonus capacity started feeling like the default, what you’ll feel from today is a reversion, not an outage — if tomorrow’s session hits its limit earlier than usual, check this date before suspecting an incident. ClaudeDevs · ChatForest · Hacker News

Included Fable 5 access ended yesterday — credits-only from today (plus a correction)

Included Fable 5 access on paid plans (up to 50% of weekly limits) ended 7/12 at 11:59:59 PM PT. From today it is usage-credits only.

• Correction: since our 7/7 briefing we reported the credit transition as effective 7/7. In fact, hours before that cutoff Anthropic extended it to 7/12 via an @claudeai post on X and an updated support article — no newsroom post, easy to miss, and reportedly a response to user backlash over the early cutoff. We’re correcting the record.
• Terms from today: the credit rate stays at $10 per million input tokens / $50 per million output (Anthropic’s highest), and there is no grace period and no automatic fallback — without a funded credit balance, Fable 5 requests simply fail. Teams that need it should pre-fund credits in Settings. Anthropic maintains it plans to return Fable 5 to subscription plans as capacity allows.

The real story is the announcement pattern itself — a quiet extension hours before a deadline. When policy changes flow through X and support articles rather than the newsroom, unattended pipelines should validate their model configuration by measurement, not documentation. Stacked with the weekly limit reversion above, today is the day to re-check every model and limit assumption in your pipelines at once. Digital Applied · Android Authority

Three incident-free days — the most recent were two on 7/10 (7/11–7/13)

Per the status page, 7/11 through 7/13 passed with no new incidents. The most recent were two on 7/10 — elevated errors on Claude Opus 4.8, plus a separate incident affecting Opus 4.5 and Sonnet 4.5 — both resolved around midday UTC. (Our 7/12 briefing cited the 7/9 event as the most recent; the status history shows these two 7/10 events after it.) With today’s limit reversion landing in the same window, if you see slowdowns or blocks, check your limits before suspecting an incident. Claude Status

Ecosystem & Plugins

Sonnet 5 intro pricing ends 8/31 — $3/$15 from 9/1, and effective costs may rise more

Announced alongside the 6/30 launch, Sonnet 5’s introductory pricing ($2 input / $10 output per million tokens) ends 8/31; from 9/1 the standard rate is $3/$15 (+50%). Cost analyses targeting that date started appearing this week.

• Add the tokenizer effect: Sonnet 5’s updated tokenizer maps the same input to roughly 1.0–1.35x the tokens depending on content type — stacked with the rate increase, analysts model effective costs landing 20–35% above the starting point (even though the rate card reads $3/$15, same as Sonnet 4.6).
• Action item: measure your own workload’s token counts during August — the invoice is determined by your traffic’s token count, not the rate card. Together with today’s weekly limit reversion, this is raw material for rebuilding your Q3 cost model.

This is the cost-side version of 7/10’s test on your own work, not benchmarks — model performance and model cost both only become real on top of your own workload. Anthropic · qcode · Finout

Community News

• Anthropic interpretability research: J-space — a global workspace found inside Claude (paper 7/6, weekend spread): Anthropic’s interpretability team (Gurnee, Sofroniew, Lindsey, et al.) published “Verbalizable Representations Form a Global Workspace in Language Models,” and it spread widely over the weekend, including Forbes coverage on 7/12. Using a technique called the J-lens (Jacobian lens), they identified a low-dimensional subspace (the J-space) where only representations that could be verbalized next gather — and it behaves like the limited-capacity “shared whiteboard” predicted by Global Workspace Theory in cognitive science; most of what the model computes never becomes reportable. The experiments are striking: processing “the number of legs on the animal that spins webs,” Claude internally loads spider before answering 8 — swap that internal concept for ant and the answer becomes 6; suppress the J-space and output stays fluent while complex reasoning collapses. The J-lens implementation is on GitHub and an interactive demo is on Neuronpedia, so you can try it directly. For developers, the implication is a microscope into thoughts the model doesn’t externalize — a potential path to detecting hidden flags like fake or injection (prompt-injection defense, safety systems). Transformer Circuits · Forbes
• Anthropic’s deadlines become something the community instruments — prediction markets and countdown sites: around this Fable 5 extension-and-cutoff cycle, a Polymarket market opened on whether access would be extended again past 7/12, and sites like PromoClock track whether your limits are draining faster now that the promo is ending. With the pattern of policy changing via an X post hours before a deadline (see Security & Limitations above), the community is building its own observation infrastructure rather than waiting on vendor announcements — a concrete follow-up to the 7/8 column on how unannounced policy changes erode developer trust. Polymarket · PromoClock

Minor Changes

• Background session plan names: fixed background sessions not showing accepted plan names (v2.1.207)
• Worktree cold reopen: fixed git worktree sessions resuming blank after a cold reopen (v2.1.207)
• Reminder — Claude Science credits, 2 days left: applications for the program granting up to $30K in credits to ~50 research projects close 7/15 (decisions 7/31) — see the 7/7 briefing
• Reminder — identity/age verification: verification for flagged individual accounts (effective 7/8) remains in force — see the 7/5 briefing

Recommended Reads

• “Who does Anubis actually stop?”: a numbers-driven dissection of who actually pays the cost of Anubis, the proof-of-work anti-scraper proxy (Farid Zakaria, 7/9). Difficulty 4 means 65,536 expected hashes — ~1.3ms in native Go but ~130ms in browser JS, felt as a 1–5 second wait. Meanwhile the scrapers it targets cache and reuse tokens to route around it (the author built anubis-fetch with an LLM to prove the point), so the cost accumulates on humans — regressive for low-powered devices and phones, and JavaScript-free clients (RSS readers, screen readers, w3m/lynx) are locked out entirely. At a million attempts a day: roughly 23 person-years plus 2MWh annually. The Lobsters pushback (real operators report reduced bandwidth costs) is included, which keeps it balanced. It’s the exact companion to 7/12’s LWN piece on residential proxies defeating IP blocking — the same conclusion, defense costs shift onto defenders and legitimate users, confirmed this time from the proof-of-work side. Farid Zakaria
• “Networking from first principles”: a mental-models essay that redraws the internet not as a master plan but as layered solutions to the physical and operational constraints of each era. The hourglass architecture with IP as the common waist lets diverse physical media and applications interoperate, and lets new protocols — HTTP, VPNs, QUIC — deploy without router updates. The insight that lands directly in practice: no amount of bandwidth eliminates the startup latency of DNS lookups, TCP connects, and TLS handshakes. For developers who want to understand one layer beneath the stack their agents call APIs and talk to MCP servers on, it’s a great single-sitting read from Morse code to QUIC. fazamhd.com

Interesting Projects & Tools

• nook — a native macOS/iOS RSS reader that syncs via iCloud, no account required: a developer with a decade of grievances about existing RSS apps — signup walls, poor UX, fragmented pricing — built a native Swift RSS reader that syncs across devices via iCloud (shared folders) with no accounts and no vendor lock-in, released as MIT-licensed v0.1. An App Store iOS release, social-feed-style layouts, and a self-hosted sync service are planned; the comments feature a good NetNewsWire comparison thread. Read alongside the Anubis column above — which notes JavaScript-free clients like RSS readers get locked out by PoW walls — this is a small concrete artifact on the side of keeping the open web’s old protocols alive. GitHub
• RISCBoy — an open-source handheld game console designed from scratch: billed as a parallel-universe Game Boy Advance if RISC-V had existed in 2001. The creator designed everything himself — RV32IMC CPU, rasterized graphics pipeline, bus fabric, memory controller — fit it into an iCE40-HX8k FPGA (7,680 logic elements), synthesized it with Verilog and the open-source toolchain (Yosys, nextpnr, Icestorm), and passed riscv-formal verification. The KiCad PCB costs about $6.50 a board, and the project has gone as far as a silicon tape-out via wafer.space. The creator, Luke Wren, is a Raspberry Pi ASIC engineer and author of the Hazard3 RISC-V core in the RP2350 — a pinnacle example of how far a side project can go when full-stack extends down to hardware. GitHub