Claude Code weekly limits go up 25% for good on Sep 14, and the week in Claude
PLUS: a judge rules the Pentagon's label on Anthropic illegal, Claude in Chrome reaches every paid plan, and 10,000 free or discounted Claude seats for scientists
Hello again. Here's your week in Claude, from Aug 23 to Aug 29.
This week, Claude Code limits got a lasting raise, starting Sep 14. A judge ruled for Anthropic against the Pentagon. And Claude in Chrome came to every paid plan.
Plus 13 quick hits. Every claim has a link.
Claude Code weekly limits go up 25% for good on Sep 14
Anthropic is keeping part of its Claude Code usage boost.
On Aug 29, Anthropic's ClaudeDevs account posted the news. Standard weekly limits in Claude Code will rise 25% for good. The change starts Sep 14.
Per the post, it covers Pro, Max, Team and seat-based Enterprise plans. Until then, the current 50% increase stays in place.
Anthropic's help article says the 25% is on top of the limits from before the promotion. That article also lists Ultra. The 50% boost never changed 5-hour limits.
A judge says the Pentagon's risk label on Anthropic was illegal
On Aug 27, a federal judge ruled that the Pentagon's label on Anthropic was illegal. The label called Anthropic a supply-chain risk.
Per TechCrunch, U.S. District Judge Rita Lin called the label "unlawful retaliation" in violation of the First Amendment. She also called it "arbitrary and capricious."
She said Anthropic was denied due process, too.
Earlier this year, Defense Secretary Pete Hegseth and President Trump gave Anthropic the label. They ordered all federal agencies to stop working with it.
Per TechCrunch, the fight began over Anthropic's hard lines on fully autonomous weapons and mass surveillance of Americans. The Pentagon denied it would use Anthropic's models for anything but lawful purposes.
Per The Register, the court looked at Claude models already inside Pentagon systems. It found Anthropic can't remotely access, modify, update or disable them. Lin called the alleged danger "entirely unfounded."
Lin wrote that the Department of War is "undisputedly free to select the AI vendor of its choice." Per TechCrunch, a second Anthropic suit in Washington, D.C. is still ongoing.
Claude in Chrome now works on every paid plan
Claude in Chrome is generally available on every paid Claude plan. Claude can now take browser actions without asking you each time.
On Aug 26, Anthropic opened Claude in Chrome to all paid plans. It can read and type text, click links and fill out forms. It uses your existing logins.
Claude now approves actions it judges safe, the same way auto mode works in Claude Code. You can switch this off in settings.
A classifier checks each action against what you asked for. If it doesn't match, it's blocked.
In Anthropic's current test, attacks came from professional red-teamers. Before extra safeguards, attacks that reached the model worked 3.8% of the time on Opus 5. Per Anthropic, none worked on Sonnet 5, Opus 5 or Mythos 5 with probes and the safety classifier on.
With the same safeguards, 0.3% of attacks worked against Fable 5. Anthropic says all successful breaks were in low-severity scenarios, and it's working to fix them. It calls prompt injection "a moving target."
Claude in Chrome doesn't run on other Chromium browsers or on mobile yet. On Enterprise plans, admins can limit it to approved domains.
Claude now shares one memory across chat and Cowork
What Claude remembers in chat now carries into Cowork, and back.
On Aug 25, Anthropic said Cowork now has memory. It's the same one you use in chat. Per the release notes, memory now works across chat and Cowork in the cloud.
Claude now adds topics to memory as you chat. It no longer waits for a chat to end. You can pause or reset memory at any time.
Everything it remembers is listed under Topics in Memory settings. You can read, edit or delete each item there.
By default, Claude doesn't store sensitive topics, like health or beliefs. You can choose to turn on "include sensitive topics in memory." Then you'll see a notice each time Claude saves one.
Some things are never stored, even with that setting on. They include sensitive ID numbers, criminal history and immigration status.
Memory is on by default on Free, Pro, Max and Ultra plans.
Cowork gets its own browser, separate from yours
Claude now has a browser built into Cowork in the desktop app.
Per Anthropic's Aug 26 post, a browser opens in the side panel when a task needs a website. Claude never sees your tabs, bookmarks or passwords.
You can bring your logins over site by site. That works from Chrome, Edge or Firefox on macOS, and from Firefox on Windows and Linux. Banking, email and single sign-on sites are left out unless you choose to include them.
It's rolling out over the coming week to Pro, Max, Ultra and Team plans. It's in the desktop app on macOS, Windows and Linux (in beta). Once it reaches you, it's on by default.
On Enterprise, admins can turn it on now. If you already use Claude in Chrome, it stays your default.
Anthropic says its safeguards "meaningfully reduce the risk but can't eliminate it." It recommends starting on sites you trust.
Salesforce and Anthropic launch Claudeforce, with a 37-skill Claude plugin
Salesforce is bringing its sales data and actions into Claude through a plugin.
On Aug 26, Salesforce and Anthropic announced Claudeforce, an expanded partnership. It launches with Salesforce in Claude, a plugin with 37 prebuilt sales skills. They include meeting prep, deal health review and pipeline review.
Per Salesforce, an admin connects it once, and every seller on the team gets access. It's open to select pilot customers now. Salesforce expects an open beta in September 2026.
Through the partnership, Claude is the default model for Slack, per Salesforce. All Salesforce developers and knowledge workers will get Claude Code and Claude Enterprise.
Anthropic opens 10,000 free or discounted Claude seats for scientists
Scientists can now get Claude for free, or at a discount, for one year.
On Aug 27, Anthropic launched a Claude team plan for scientists with 10,000 seats. Standard seats are free. Premium seats with 5x usage limits cost $15 per month.
To qualify, you must be a principal investigator or equivalent. It has to be at an academic or nonprofit research institution. You can then add the researchers in your lab.
Anthropic says it intends to extend the program well beyond 10,000 seats over the coming months. Its AI for Science program offers up to $50,000 in credits per project. Any researcher can apply.
For now, researchers in biology and chemistry are still limited to Opus-class models. Fable models will keep blocking professional biology and drug development queries. Anthropic cites their potential dual-use risks.
Anthropic says Claude found fixes for 10 kinds of AI safety flaws
Anthropic had Claude work on its own to make AI models behave more safely.
In an Aug 28 report, Claude trained models to do better on public tests. The tests cover 10 kinds of alignment failure, like deception and sycophancy.
Per Anthropic, Claude found fixes for all 10 that improved the target tests without hurting capabilities. The methods also worked on models up to 4.7 times larger.
Claude also outscored 28 human safety researchers who had up to eight hours. Anthropic notes the humans couldn't iterate, so it sees this "less as a direct comparison."
An Opus 4.8 monitor found cheating attempts in 39 of about 1,600 research transcripts (2.4%). Anthropic is "cautiously optimistic" it caught most of the cheating.
Anthropic lists limits. The failures it studied were narrow compared with those in production.
It only checked a limited set of capabilities. Accepted methods may have hurt others it didn't measure. It also didn't test whether the gains last after heavy RL training on other tasks.
Quick hits
- Claude models had elevated errors overnight Aug 23 to 24, including Opus 5 and Fable 5. Per Anthropic's status page, they ran from 04:50 to 07:36 UTC on Aug 24. That's 11:50 pm to 2:36 am Central.
- An upstream cloud provider issue hit Cowork and Claude Code on the web on Aug 28. Some sessions could fail to start or drop mid-task, and could be retried. The incident was open from 17:22 to 20:21 UTC.
- The Model Hardware Standard is in research preview (Aug 27). It's a shared spec for AI agents to safely run devices, like microscopes and robotic arms. It's open to a first group of labs and manufacturers, ahead of an open-source release.
- Anthropic launched a $5 million grant program (Aug 25). It funds independent, open-source research on how AI affects users' wellbeing. Applications are due Sep 21.
- Claude for Teachers is now a free Enterprise offering for US K-12 schools and districts (Aug 28). Qualifying schools and districts that sign up by June 30, 2027 get a full year free. Its data isn't used for model training.
- Bain joined the Claude Partner Network at its top tier (Aug 25). That tier requires 1,000 certified practitioners. Bain rolled Claude out to all 19,000 of its employees.
- You can now create personal and service account API keys in the Claude Console (Aug 27). Per the release notes, they act as you or a service account. They stop working when that account leaves the org.
- The Compliance API's session endpoints are out of beta (Aug 26). That covers Cowork and Claude Code sessions.
- Claude Code 2.1.248 adds a --restricted mode (Aug 27). It removes built-in tools that run commands or code, plus WebFetch, unless named in --tools. It keeps file tools inside the working directory.
- Claude Code 2.1.251 fixes a symlink hole (Aug 28). File tools could follow a symlink swapped inside the working directory after the permission check. That could read or write outside the approved location.
- Claude in Chrome actions now always go through Claude Code's permission checks (2.1.251, Aug 28). That includes sessions with telemetry off.
- Claude Code's native install download shrank to about 75 MB from 340 MB (2.1.243, Aug 25). That's on Linux x64. Native builds also use roughly 40 to 70 MB less memory per session.
- Opus 5 is now Claude Code's default model for seat-based Enterprise (2.1.251, Aug 28). That matches other premium plans.
Worth your time
- Outside researchers studied real Claude chats (Aug 26): three outside groups ran studies. Stanford's SALT Lab found over half of the chats involved people handing consequential tasks to AI. Each group studied roughly 250,000 chats from April to May 2026.
- How Anthropic employees use Claude Tag (Aug 28): three ways its teams tag @Claude in Slack threads. The post shares their prompts.
- How Warp builds self-improving agents on Claude (Aug 26): Warp uses skills to keep feedback to its agents. Feedback usually vanishes when a session ends.
- An Anthropic field marketer's weekly sales digest (Aug 24): he built a weekly Slack update for each rep. He made it with Claude Code.
That's the week. The next issue lands Sunday, Sep 6.
Reply and tell me what you're building. I read every reply.
— Sam