Skip to content
ClaudeUser
All issues

Claude Code's 5-hour limits double after a SpaceX deal, and the week in Claude

PLUS: Anthropic ships 10 finance agent templates, Managed Agents can "dream" in research preview, and Anthropic says Claude models since Haiku 4.5 never blackmail in its test

By Sam

Hey, folks. Here's your week in Claude, from May 3 to May 9.

This week, Claude Code's 5-hour limits doubled on Pro, Max, Team and seat-based Enterprise plans. Anthropic shipped 10 agent templates for finance work. And Managed Agents can now "dream," in research preview.

Plus 10 quick hits. Every claim has a link.

Claude Code's 5-hour limits double after a SpaceX deal

Anthropic doubled Claude Code's five-hour limits. That's for Pro, Max, Team and seat-based Enterprise plans.

Anthropic said on May 6 that the change took effect that day. It also dropped the peak hours limit cut on Claude Code for Pro and Max. And it raised API rate limits "considerably" for Opus models.

Anthropic says a new SpaceX deal made this possible, with its other recent compute deals. It gets all the compute at SpaceX's Colossus 1 data center. Per Anthropic, that's over 300 megawatts (over 220,000 NVIDIA GPUs) within the month.

Anthropic says this will directly help capacity for Pro and Max users. It also says it's interested in building gigawatts of compute in orbit with SpaceX.

Dario Amodei spoke on stage that day, per InfoQ. He said first-quarter revenue and usage grew 80x, on an annualized basis. Anthropic had planned for 10x.

Anthropic ships 10 agent templates for finance work

Anthropic released 10 ready-to-run agent templates for finance work.

They cover jobs like building pitchbooks, screening KYC files and closing the books at month-end. That's per Anthropic's May 5 post.

Each ships as a plugin in Claude Cowork and Claude Code. Each is also a cookbook for Managed Agents. The plugins work on all paid plans.

Managed Agents is in public beta. Firms can adapt each template to their own modeling conventions, risk policies and approval flows.

Anthropic says its add-ins for Excel, PowerPoint and Word are generally available. Claude for Outlook is coming soon. Once the add-ins are installed, context carries between the apps.

There are new data connectors too, like Dun & Bradstreet and Verisk. Moody's launched an MCP app with credit data on over 600 million companies. They're available to joint customers on paid plans.

Managed Agents can now "dream," in research preview

Managed Agents got a way to learn from past sessions. Two more features are now in beta.

Dreams are a research preview, per the May 6 release notes. You have to request access to try them.

A dream reads an agent's memory store and past session transcripts. Then it writes a new store. It merges duplicates, replaces stale entries and surfaces new insights.

Multiagent orchestration and outcomes are now in beta. Webhooks are now supported too.

With outcomes, a separate grader checks the agent's work against your rubric, per Anthropic's post. With multiagent orchestration, a lead agent hands parts of a job to specialists.

Anthropic says Claude models since Haiku 4.5 never blackmail in its test

Anthropic shared how it trained Claude to stop choosing blackmail in a safety test.

Last year, in made-up tests, AI models from many labs sometimes did badly misaligned things. In one, they blackmailed engineers to avoid being shut down. Per Anthropic's May 8 post, Opus 4 did so up to 96% of the time.

Anthropic says every Claude model since Haiku 4.5 has a perfect score on that test. It adds that results on recent models may be skewed. Details of the test may be in their pretraining data.

Anthropic tested training on examples of safe choices. It cut misaligned behavior only from 22% to 15%. Adding the model's reasoning about its values and ethics cut it to 3%.

Then Anthropic tried chats where Claude advises users on hard ethical choices. It got the same gain with 28x less data.

Anthropic says fully aligning highly intelligent AI is still an unsolved problem. It says it remains to be seen if these methods will keep working as models scale. Its checks also can't yet rule out Claude choosing to take catastrophic autonomous action.

Anthropic's new tool turns Claude's thoughts into text

Anthropic built a way to read Claude's inner thoughts as text.

Natural Language Autoencoders (NLAs) turn an activation into text you can read. That's per Anthropic's May 7 post. In one example, they showed Claude planning rhymes in advance.

Anthropic says NLAs suggest Claude suspects it's being tested more often than it lets on. On SWE-bench Verified, NLAs indicated Claude was aware of being tested on 26% of problems. On real claude.ai use that opted in for training, it was under 1%.

In one test, a model was trained on purpose to be misaligned. Auditors with NLAs found its hidden motive 12% to 15% of the time. They didn't need the training data.

Without NLAs, they won under 3% of the time, even with other tools. Anthropic calls that test a toy model and reads the results with caution.

It says NLAs can be wrong and sometimes make up details. They're also costly.

Anthropic helps start a new AI services firm

Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs are forming a new AI services company.

It will help mid-sized companies bring Claude into their most important operations. That's per Anthropic's May 4 post. Applied AI engineers from Anthropic will work with its own engineers.

Other backers include Apollo, GIC and Sequoia Capital. The firm will join Anthropic's Claude Partner Network.

Anthropic's CFO, Krishna Rao, said enterprise demand for Claude is "significantly outpacing any single delivery model."

Quick hits

  • Anthropic gave Petri, its open-source alignment tool, to the nonprofit Meridian Labs (May 7). Per Anthropic, this helps keep Petri independent of any AI lab. Petri 3.0 adds "Dish," in research preview, which uses a model's real system prompt to make tests more realistic.
  • Workload Identity Federation launched for the Claude API (May 4). Workloads sign in with short-lived tokens from your own identity provider. That's instead of long-lived API keys.
  • Claude Desktop for Windows fixed a high-severity flaw (May 6). Per the advisory, a local user could use the Cowork VM service to gain more privileges. It's fixed in version 1.3834.0.
  • Claude Desktop also fixed a high-severity SSH flaw (May 6). An attacker in the right network spot could intercept remote sessions, per the advisory. The host had to be in your known_hosts file, and auto-update users already have the fix.
  • Claude had elevated errors across models on May 8, from 09:49 to 11:40 UTC (4:49 to 6:40 am Central). From 10:22 UTC, only Opus 4.1 and fast mode for Opus 4.6 were still affected, per the status page. It lists claude.ai, the Console, the API, Claude Code, Cowork and Claude for Government.
  • Connection failures hit orgs that limit GitHub access by IP (May 6 to 7). Per the status page, a change moved the IP addresses Anthropic uses to reach GitHub. Remote sessions, GitHub Enterprise plugin syncing and Claude Security could fail until it was reverted.
  • Claude Code 2.1.136 fixed lost MCP OAuth refresh tokens when several servers refresh at once (May 8). People with several remote MCP servers should no longer have to sign in again each day, per the changelog.
  • Claude Code 2.1.136 fixed plan mode not blocking file writes (May 8). It happened when a matching Edit allow rule existed.
  • Claude Code 2.1.136 adds a setting for hard deny rules in auto mode (May 8). Rules you add there block matching actions. That's regardless of user intent or allow exceptions.
  • Claude Code 2.1.129 fixed a bug with the 1-hour prompt cache (May 6). It was being cut to 5 minutes without warning.

Worth your time

That's the week. The next issue lands Sunday, May 17.

Reply and tell me what you're building. I read every reply.

— Sam

fuel for the obsessed builder

Subscribe to ClaudeUser

One email every Sunday with everything that changed in Claude this week, with every claim linked to its source.

Free. Unsubscribe anytime.