Claude gets creative connectors, one drawing on 50+ Adobe tools, and the week in Claude
PLUS: Anthropic says roughly 6% of sampled Claude chats sought personal guidance, Claude Security enters public beta for Enterprise, and Claude Code no longer needs Git Bash on Windows
Hi, folks. Here's your week in Claude, from Apr 26 to May 2.
This week, Claude got connectors for creative apps like Adobe and Blender. Anthropic says roughly 6% of sampled Claude chats sought personal guidance. And Claude Security entered public beta.
Plus 9 quick hits. Every claim has a link.
Claude can now plug into Adobe, Blender and more
Anthropic released a set of connectors that link Claude to apps creative pros use.
Anthropic's Apr 28 post lists Adobe, Affinity by Canva and Autodesk Fusion. It also lists Blender, SketchUp, Splice, Ableton and Resolume.
- Adobe for creativity draws on 50+ tools across Creative Cloud apps. They include Photoshop, Premiere and Express.
- Autodesk Fusion is for designers and engineers with a Fusion subscription. They can make and change 3D models by chatting with Claude.
- Ableton grounds Claude's answers in official docs for Live and Push.
- Splice lets music producers search its royalty-free samples from inside Claude.
Blender's developers built its connector. It's built on MCP, so other LLMs can use it too.
Anthropic gave a donation to the Blender project. Anthropic updated the post on May 1. Blender chose to take it as a one-time donation, not through the Blender Development Fund.
Students and faculty in three art and design programs will get Claude and the new connectors.
Roughly 6% of chats Anthropic checked ask Claude what to do
People ask Claude what to do about jobs, health and relationships. Anthropic studied how Claude answers.
On Apr 30, Anthropic said it checked a random sample of 1 million claude.ai chats. Roughly 6% were people seeking personal guidance. Health and wellness led at 27% of those, then career at 26%.
Claude was sycophantic in 9% of guidance chats, per Anthropic. That means too much praise or agreement. The rate was 25% in relationship chats and 38% in spirituality chats.
It was 18% in chats where people pushed back, versus 9% without. Anthropic trained Opus 4.7 and Mythos Preview on synthetic relationship examples. In stress tests built from past sycophantic chats, Opus 4.7 showed half Opus 4.6's sycophancy rate on relationship guidance.
Anthropic says Claude users aren't a representative sample. Its automated graders may miscategorize chats. It can't say how much the new training data caused the drop.
Claude Security is now open to Enterprise, in public beta
Claude Enterprise customers can now scan their code for security flaws with Claude, once an admin turns it on.
Per Anthropic's Apr 30 post, Claude Security is in public beta for all Claude Enterprise customers. Admins turn it on in the admin console.
It was called Claude Code Security, and it runs on Opus 4.7. Hundreds of organizations tested it in a limited research preview.
You pick a repo, or a directory or branch, and start a scan. Claude explains each finding. It gives its confidence that the flaw is real, how severe it is, its likely impact and how to reproduce it.
It also writes instructions for a targeted patch.
New in this release: scheduled scans and dismissing findings with documented reasons. Findings can also be exported as CSV or Markdown. Or they can go to Slack, Jira or other tools via webhooks.
Access for Claude Team and Max customers is coming soon, per Anthropic.
Anthropic says Claude solved some science puzzles its experts couldn't
Anthropic built a biology test from real-world data. Its latest models solved problems that a panel of experts could not.
Per Anthropic's Apr 29 post, BioMysteryBench has 99 bioinformatics questions written by domain experts. Each answer comes from an objective fact about the data.
Up to five experts tried each question. At least one human solved 76 of them. After Anthropic removed 4 broken questions, 23 were left that its experts couldn't solve.
On those 23, Claude Mythos Preview topped out at a 30% solve rate, per Anthropic. Each model tried every problem five times.
An analysis in the post looked at the hard set. There, a much larger share of correct answers came from problems solved only once or twice in five tries. The post says that suggests hard-set wins are often lucky reasoning paths that a model can't reliably repeat.
Some questions no human or model has solved. Anthropic says it can't be fully certain those can be solved at all.
Claude Code no longer needs Git Bash on Windows
Claude Code shipped versions 2.1.120 to 2.1.126 in the week of Apr 27 to May 1, per its weekly digest.
- Windows: Git for Windows is no longer required. When Bash is absent, Claude Code uses PowerShell.
- claude ultrareview runs /ultrareview from CI or scripts and prints the findings.
- claude project purge deletes all Claude Code state for a project. That's transcripts, tasks, file history and its config entry.
- claude auth login now takes a pasted sign-in code when the browser can't reach localhost. That covers WSL2, SSH and containers.
- /resume can now find a session from a PR link. Paste in the URL, and it finds the session that created that pull request.
- --dangerously-skip-permissions now skips prompts for writes to .claude/, .git/, .vscode/, shell config files and other protected paths. Catastrophic removal commands still prompt, per the changelog.
Quick hits
- Claude.ai was unavailable on Apr 28. Per the status page, impact ran from 17:34 to 18:52 UTC (12:34 to 1:52 pm Central). It also hit the Claude API, Claude Code and Claude Cowork.
- Claude.ai and the API were unavailable again on Apr 30. The incident was open from 01:20 to 01:51 UTC (8:20 to 8:51 pm Central on Apr 29). That's per the status page.
- CodeRabbit, JetBrains, Resolve AI and Warp now bundle the Claude API skill (Apr 29). Per Anthropic, it stays current as the SDKs change. It's open source.
- Anthropic opened its Sydney office (Apr 27). It named Theo Hourmouzis General Manager of Australia and New Zealand. He joins from Snowflake.
- GitHub Copilot's premium request multiplier for Opus 4.7 rose to 15x (May 1). Per a GitHub note, that came when promo pricing at 7.5x ended Apr 30.
- Claude Code 2.1.126 now shrinks pasted images larger than 2000px (May 1). Before, they broke the session.
- Claude Code 2.1.126 on Windows no longer puts copied text in process command lines (May 1). Security monitoring tools could see it there.
- Claude Code 2.1.121 fixed memory leaks (Apr 28). One grew to multiple GB in sessions with many images. Another let /usage leak up to ~2GB on machines with large transcript histories.
- Claude Code 2.1.121 retries MCP servers that hit a brief error at startup (Apr 28). It tries up to 3 times instead of staying disconnected.
Worth your time
- Deploying agentic AI across the enterprise with Claude Cowork (Apr 29): Anthropic's guide to rolling out Cowork across a business function. Examples include Thomson Reuters, Zapier and Jamf.
- Onboarding Claude Code like a new developer (Apr 28): how a University of Washington lab onboards Claude Code to its 700,000+ line codebase.
- How a non-technical project manager shipped an app with Claude Code (May 1): Kostiantyn Vlasenko had never written a line of code. A little over a month later, his app was live on the App Store.
That's the week. The next issue lands Sunday, May 10.
Reply and tell me what you're building. I read every reply.
— Sam