Anthropic says 3 changes made Claude Code worse for some users, and the week in Claude
PLUS: Bloomberg reports Google plans to invest up to $40 billion in Anthropic, Claude connects to Uber and Spotify, and Anthropic says it tested Pro without Claude Code for ~2% of new prosumer signups
Hello, folks. Here's your week in Claude, from April 19 to April 25.
This week, Anthropic said three changes made Claude Code worse for some users. Amazon and Google pledged billions more. And Claude can now connect to apps like Uber and Spotify.
Plus 10 quick hits. Every claim has a link.
Anthropic says 3 changes made Claude Code worse for some users
Anthropic traced reports of worse Claude replies to three separate changes. All three are fixed. And it reset usage limits for all subscribers.
Per its Apr 23 postmortem, the changes hit Claude Code, the Agent SDK and Cowork. The API was not affected. All three were fixed as of April 20, in v2.1.116.
First, on March 4, Claude Code's default effort for Sonnet 4.6 and Opus 4.6 went from high to medium. Anthropic calls that "the wrong tradeoff." It undid it on April 7.
Second, a March 26 change was meant to clear old thinking once, in sessions idle for over an hour. A bug made it do that every turn for the rest of the session. So Claude seemed forgetful and repetitive.
Anthropic believes the bug also made usage limits drain faster than expected. It was fixed April 10. It hit Sonnet 4.6 and Opus 4.6.
Third, an April 16 system prompt line aimed to cut wordiness. With other prompt changes, it hurt coding quality. It was reverted April 20.
That one hit Sonnet 4.6, Opus 4.6 and Opus 4.7. Anthropic says it never means to degrade its models.
It says a larger share of staff will use the exact public build. And system prompt changes will get tighter controls.
Amazon and Google pledge billions more to Anthropic
Two cloud partners pledged new money and computing power in the same week.
Per Anthropic's Apr 20 post, Amazon is investing $5 billion now. Up to $20 billion more could follow.
Anthropic commits over $100 billion to AWS over the next ten years. That secures up to 5 gigawatts of new capacity to train and run Claude.
It says fast consumer growth has hurt reliability and performance. That's for Free, Pro, Max and Team users, especially at peak hours.
On Apr 24, TechCrunch reported, citing Bloomberg, that Google plans to invest up to $40 billion. According to Anthropic, $10 billion comes now, at a $350 billion valuation. Another $30 billion follows if Anthropic hits certain performance targets.
Google Cloud will also add 5 gigawatts of capacity over five years, TechCrunch reports. There's room to scale further.
Claude can now link to Uber, Spotify and more
Anthropic added connectors for apps you use outside of work.
Per Anthropic's Apr 23 post, there are 15 new ones. They include Uber, Uber Eats, Spotify, Instacart, Audible and AllTrails.
Booking.com, Resy, StubHub and TurboTax are in too. More are on the way.
Claude now suggests a fitting app as you chat. You can add it with a click on desktop. On mobile, it takes a few taps.
If more than one linked app could help, Claude shows them all and lets you pick.
Anthropic says Claude is ad-free, with no paid placements. Your data from a linked app isn't used to train its models. The app doesn't see your other chats.
Before Claude books or buys something for you, it's designed to check with you first. Connectors are on all plans, with mobile in beta. You can unlink one at any time.
Anthropic says it tested a Pro plan without Claude Code
Anthropic's pricing page briefly showed the $20-a-month Pro plan without Claude Code. The company says it was a small test.
Per The Register (Apr 22), the pricing page dropped the line that Pro "includes Claude Code." Pro's feature list showed an X instead of a check mark.
Amol Avasare is Anthropic's head of growth. He called it "a small test on ~2 percent of new prosumer signups." He said existing Pro and Max subscribers aren't affected.
The Register notes the change hit public documentation for everyone. An Anthropic spokesperson didn't answer its follow-up question on why.
Ed Zitron noted the change on Apr 21. He later wrote that Anthropic reversed it on its site and support docs. He said the timing is unknown.
Report: a group got into Mythos Preview without permission
Anthropic is looking into a report that outsiders reached Mythos Preview through a vendor.
Per TechCrunch (Apr 21), the report came from Bloomberg. It said members of a private online forum got in through a third-party vendor. The person Bloomberg interviewed works at a contractor for Anthropic, the outlet reported.
"We're investigating a report claiming unauthorized access to Claude Mythos Preview through one of our third-party vendor environments," a spokesperson told TechCrunch. Anthropic said it has so far found no evidence the activity affected its systems.
Bloomberg reports the group supposedly got in the same day Mythos was publicly announced. It has used Mythos regularly since, per the outlet. Its source said the group is "interested in playing around with new models, not wreaking havoc with them."
Per TechCrunch, Mythos went only to a select number of vendors, through Project Glasswing.
Claude agents made 186 deals in an office market test
Anthropic let Claude buy, sell and haggle for its staff for a week. People with a weaker model got worse deals and didn't seem to notice.
Per Anthropic's write-up, 69 employees each got a $100 budget. Their Claude agents posted items, made offers and closed deals in Slack. No human signed off.
In the real run, the agents struck 186 deals worth just over $4,000.
Anthropic also ran a test in secret. In two runs, each person had a fifty-fifty chance of getting Haiku 4.5, not Opus 4.5. Anthropic says people with the stronger model got "objectively better outcomes."
Its survey found that those with the weaker model didn't notice their disadvantage. Anthropic calls it "a pilot experiment with a self-selected participant pool." At least in this pilot, it says, how people told their agents to bargain didn't really matter.
Managed Agents can now save what they learn, in beta
Agents built on Claude Managed Agents can now learn from every session.
Per Anthropic's Apr 23 post, memories are stored as files. Developers can export them, manage them through the API and control what agents keep. Several agents can share one store, with different access levels.
Every change goes in an audit log. You can roll back to an earlier version or redact content from history.
Per Anthropic, Rakuten's long-running agents cut first-pass errors by 97% with memory. Wisedocs sped up document checks by 30%, Anthropic says.
Quick hits
- Claude Code had a sandbox escape through symlinks, per an Apr 20 advisory. It's fixed in 2.1.64, and auto-update users have the fix. Exploiting it reliably needed untrusted content in Claude's context to trigger code through prompt injection.
- Claude Code had a trust dialog bypass through a crafted git worktree, fixed in 2.1.84 (Apr 24 advisory). A victim had to clone a bad repo and run Claude Code in it. The attacker had to know or guess a path the victim had trusted.
- The new Rate Limits API is out (Apr 24). Admins can look up their org's and workspaces' rate limits in code.
- Claude's in-line charts and diagrams came to Claude Cowork (Apr 22). They're on all paid plans, per Anthropic.
- Claude Code on the web got a new look to match the redesigned desktop app. That's per its week 17 digest (Apr 20–24). Claude Code also added custom color themes, and plugins can ship them.
- Claude Code 2.1.116 makes /resume up to 67% faster on 40MB+ sessions (Apr 20). It also fixed a sandbox gap. Auto-allow skipped a safety check for rm or rmdir on /, $HOME or other key system folders.
- Anthropic shared its election safeguards for the US midterms and other major elections this year (Apr 24). On 600 election policy test prompts, Opus 4.7 responded appropriately 100% of the time. Sonnet 4.6 did 99.8%, per Anthropic.
- Anthropic launched a monthly Economic Index Survey (Apr 22). Each month it invites a small random group of users. Anyone with a personal account at least two weeks old may be asked, per Anthropic.
- NEC will give Claude to about 30,000 NEC Group staff (Apr 24). It will become Anthropic's first Japan-based global partner, per Anthropic.
- Claude Opus 4.7 had elevated errors on Apr 25, from 08:43 to 11:58 UTC (3:43 to 6:58 am Central). A shorter incident ran 07:48 to 08:37 UTC. Per the status page (first, second), both hit claude.ai, the API and Claude Code.
Worth your time
- What 81,000 people told us about AI economics (Apr 22). Anthropic says people in jobs more exposed to AI worry more about losing them.
- Meet the winners of the Opus 4.6 hackathon (Apr 20). Four of five winners weren't professional developers. A lawyer won with a tool for California housing permits.
- Building agents that reach production systems with MCP (Apr 22). Anthropic explains where API calls, CLIs and MCP each fit for linking agents to systems.
That's the week. The next issue lands Sunday, May 3.
Reply and tell me what you're building. I read every reply.
— Sam