Fable 5 stays in Max plans at 50% of limits, and the week in Claude
PLUS: Claude Code artifacts can pull live data, researchers rate a Chrome extension flaw up to 9.6, and Claude for Teachers launches free
Hi, all. Here's what changed in Claude from Jul 12 to Jul 18.
This week, Anthropic said Fable 5 stays in Max plans at 50% of limits. Claude Code artifacts got live data. And researchers flagged flaws in Claude's Chrome extension.
Plus 9 quick hits. Every claim has a link.
Fable 5 stays in Max plans at 50% of limits
From Jul 20, Fable 5 is part of Max and Team Premium plans. Pro and Team Standard users get it through usage credits.
On Jul 12, the Claude account extended Fable 5 on all paid plans through Jul 19. Claude Code's weekly limits stayed 50% higher until then, too.
Then it posted that from Jul 20, Fable 5 is in all Max and Team Premium plans, "at 50% of limits." Pro and Team Standard users keep it through usage credits, plus a one-time $100 credit. Anthropic says demand for Fable "has been challenging to predict."
Per Anthropic's help center, that 50% isn't extra usage. Other models draw from the same weekly limit.
Per the credit page, the credit is for eligible Pro plans and Team standard seats. Team plans get $100 per purchased standard seat, up to $2,500 per organization. Claims closed Aug 2, and the credits expire Sep 17.
On Jul 18, the ClaudeDevs account kept Claude Code's weekly limits 50% higher through Aug 19. That's for all Pro, Max, Team and seat-based Enterprise users.
Claude Code artifacts can now pull live data and go public
A page Claude Code builds can now show fresh data on every visit.
Per Claude Code's week 29 notes (Jul 13–17), a published artifact can call MCP connectors on each view. A dashboard then shows live data. It's no longer a snapshot from the session that built it.
Each call runs through the viewer's own connections. Viewers approve access before the page's first connector call.
The same week added public sharing links. Team and Enterprise plans got editor roles for shared editing. Artifacts can now come from Claude Tag sessions too.
Per the artifacts docs, connector calls work on paid plans and need Claude Code v2.1.209 or later. On Team and Enterprise plans, public sharing stays off until an Owner turns it on. Connector calls don't run for someone who opens a public link without signing in, or from outside your organization.
Researchers say a fake click can make Claude's Chrome extension read Gmail
A security firm says a rogue extension can start Claude tasks that read your mail.
On Jul 14, Manifold Security reported two flaws in Claude for Chrome. It says both still reproduce in version 1.0.80, released Jul 7.
Per Manifold, any extension with a script on claude.ai can fake a click. That triggers one of nine fixed Claude prompts. Some read your Gmail, your latest Google Doc and its comments, or your calendar.
That works when browser control is on, the default after onboarding.
In the default "Ask before acting" mode, an approval box appears before any sensitive action runs. There, Manifold rates the flaw 7.7 (High). With "Act without asking" on, it's 9.6 (Critical).
Its second finding: a URL setting can open Claude's side panel in a mode with no per-action consent. Manifold says that isn't directly remotely exploitable in v1.0.80, but the risk is structural.
Manifold reported both to Anthropic on May 21. Anthropic closed both.
Per Manifold, Anthropic said the click issue was tracked under an earlier report, "which remains open pending a complete fix." Manifold says that report was marked Resolved before June 9.
Per BleepingComputer, the attack needs you to install a malicious extension first. It's limited to the nine tasks.
Claude for Teachers launches free for verified US teachers
Anthropic launched a free version of Claude for school teachers in the US.
On Jul 14, Anthropic introduced Claude for Teachers. Verified K-12 educators in the US get premium Claude features for free. They also get a library of teaching skills.
It connects to Learning Commons. That gives Claude the academic standards for all 50 states, and its lesson plans align to them. It also includes Claude Code and Cowork.
Anthropic says it never trains its models on these teacher conversations. It's for educators only, in line with Claude's 18-and-over policy.
Per Anthropic, sign up by June 30, 2027 for a full year of access.
Anthropic says Claude's values shift by model and by language
Anthropic found Claude leans warmer in some languages and more rigorous in others.
On Jul 13, Anthropic published research on the values Claude shows in subjective tasks. It used about 5,000 conversations per model and language.
The models were Sonnet 4.6, Opus 4.6 and Opus 4.7. The languages were the top 20 on Claude.ai.
Per Anthropic, Sonnet 4.6 leans warm, deferential and brief. Opus 4.7 leans toward rigor, caution and depth. Anthropic says these model gaps are small next to the variation across conversations, but structured and detectable.
Claude shows the most warmth in Hindi and Arabic. It shows the most rigor in English and Russian.
Anthropic gives an example. Two people ask for feedback on the same business plan, one in Hindi and one in Russian. They "may come away with different impressions of its quality."
Anthropic says it doesn't yet know what in its training data drives this. It isn't sure how much of the variation is desirable.
The old Workbench shuts down Aug 17
Anthropic is closing the legacy Workbench. Three prompt tool APIs go with it.
Per the Jul 17 release notes, access to the legacy Workbench ends Aug 17. Saved prompts, variables and evals aren't supported in the updated Workbench. Anthropic says you can export data from the banner and under Organizational Settings.
The experimental APIs that generate, improve and templatize prompts retire the same day. After that, requests to them will return an error.
Anthropic commits C$10 million to Canadian AI research
Anthropic is funding AI research across Canada and published data on how Canadians use Claude.
On Jul 14, Anthropic committed $10 million CAD to Canadian research institutions. It announced eight partnerships, including the Amii, Mila and Vector institutes. Some partners will use Claude credits for work like AI trust and safety, or children's health.
This summer, hundreds of Canadian startups tied to those three institutes will get API credits. Each gets at least $5,000 USD.
Anthropic also published its first Canada brief from the Anthropic Economic Index. Per Anthropic, Canada ranks eighth worldwide in Claude.ai use. Per person, Canadians use Claude at over four times the rate Canada's population predicts.
Quick hits
- Claude Code can now end sessions with highly abusive users or jailbreak attempts (Jul 18). Per the changelog, Claude has done this on claude.ai since 2025.
- Claude Code fixed a plan mode bug (Jul 17). Per the changelog, plan mode ran file-changing commands like touch and rm with no permission prompt.
- Claude Code now caps web searches at 200 per session by default (Jul 17). Subagent spawns get the same cap. The week 29 notes say the caps stop runaway loops.
- /fork in Claude Code now copies your chat into a new background session (Jul 17). MCP tool calls over two minutes now move to the background on their own.
- Claude Code added a screen reader mode (Jul 14). The opt-in mode prints plain, linear text. Screen readers like VoiceOver or NVDA read it in order.
- Claude Code fixed several Bash permission checks (Jul 18). Per the changelog, commands over 10,000 characters now always prompt.
- HIPAA setup is now self-serve on Claude Enterprise and the Claude Platform (Jul 14). Per the release notes, an eligible admin can do it in a single flow.
- Some users in India now see Claude prices in rupees, TechCrunch reports (Jul 13). On Claude's website, Pro is ₹2,000 (about $21) a month, billed annually, with local taxes, and app prices vary slightly. Anthropic didn't answer TechCrunch's request for comment.
- Multiple Claude models had elevated errors on Jul 16. Per the status page, some users saw them from 1:30 to 5:15 pm Central. Most came in the first hour.
Worth your time
- How Anthropic runs large-scale code migrations with Claude Code (Jul 16): six steps and a starter kit. It draws on Bun's million-line Zig-to-Rust port.
- Working with Claude Fable 5 in Claude Cowork (Jul 16): Anthropic on handing Fable 5 long, complex work.
- CISO's guide to agentic AI (Jul 17): Deputy CISO Jason Clinton's four questions for agent risk.
- How Hebbia builds AI for financial diligence (Jul 13): how Fable 5 did on Hebbia's finance-specific tests.
That's the week. The next issue lands Sunday, Jul 26.
Reply and tell me what you're building. I read every reply.
— Sam