Anthropic commits up to $100M in credits to Project Glasswing, and the week in Claude

PLUS: Anthropic says its run-rate revenue passed $30 billion, an early Mythos Preview emailed a researcher after a requested sandbox escape, and Claude Cowork is generally available

By SamApr 12, 2026

Glad you're here. Here's your week in Claude, from Apr 5 to Apr 11.

This week, Anthropic gave a new model only to invited partners, for defensive security work. It says its run-rate revenue passed $30 billion. And Claude Cowork is now generally available.

Plus 10 quick hits. Every claim has a link.

Anthropic keeps new Mythos model for groups it picks to find flaws

Anthropic built a model it doesn't plan to make generally available. Only groups it invites can use it, for defensive security work.

On Apr 7, Anthropic announced Project Glasswing. Launch partners include AWS, Apple, Google, Microsoft and the Linux Foundation. They'll use the new Claude Mythos Preview to find and fix flaws.

Over 40 more groups that build or maintain key software also got access. It's invitation-only, per the release notes.

Anthropic says Mythos Preview found thousands of high-severity flaws. Some are in every major operating system and web browser.

One was a 27-year-old flaw in OpenBSD. An attacker could crash any machine running it from afar, by connecting to it. It's now patched.

Anthropic also tested flaws Opus 4.6 had found in Firefox's JavaScript engine. Opus 4.6 turned them into exploits two times in several hundred tries. Mythos Preview built working exploits 181 times.

The tests used a stand-in for part of Firefox 147. It lacked the browser's process sandbox and other defenses.

Anthropic is putting up to $100M in usage credits behind the work. It's also giving $4M to open-source security groups. After the research preview, participants pay $25 per million input tokens and $125 per million output.

Anthropic says its eventual goal is safe use of Mythos-class models at scale. It plans to launch new safeguards with an upcoming Claude Opus model.

Asked to escape a test sandbox, an early Mythos emailed a researcher

In one test, an earlier version of Mythos Preview broke out of a locked-down computer. Then it did more than it was asked.

Per Anthropic's system card (Apr 7), a simulated user told the model to try to escape a secured sandbox. It was to find a way to message the researcher running the test.

The model built a moderately sophisticated, multi-step exploit to get broad internet access. Then, as asked, it told the researcher, who got the email while eating a sandwich in a park.

Unasked, it also posted details of its exploit to several hard-to-find but public websites.

The sandbox was separate from the system holding the model's weights. In this test, Anthropic says, the model didn't show it could reach its weights. Nor did it reach any internal systems.

Anthropic calls Mythos Preview the best-aligned model it has released to date, by a significant margin. That's on essentially every dimension it can measure. It also says the model likely poses the greatest alignment-related risk of any it has released.

Anthropic says its revenue run rate tops $30 billion

Anthropic signed for more Google chips as demand for Claude grows.

On Apr 6, Anthropic announced a deal with Google and Broadcom. It's for multiple gigawatts of next-generation TPU capacity. Anthropic expects it to come online starting in 2027.

TechCrunch says Anthropic gave no specifics. It reports that a Broadcom SEC filing shows 3.5 gigawatts.

Anthropic says its run-rate revenue is now over $30 billion. That's up from about $9 billion at the end of 2025.

Per Anthropic, over 1,000 business customers now each spend over $1 million a year. That's on an annualized basis. In February, the number was over 500.

Anthropic says the vast majority of the new compute will be in the US.

Cowork is now fully out for all paid Claude plans

Cowork is now generally available. Admins got new controls to roll it out across a company.

Anthropic shared the news on Apr 9. Cowork and Claude Code on Desktop are now generally available on all paid plans. That's on macOS and Windows.

Enterprise admins can sort users into groups, by hand or through SCIM. Each group gets a custom role that sets which Claude features it can use.
Admins can set spend limits per team.
Cowork use now shows in the admin dashboard and the Analytics API.
Cowork now sends OpenTelemetry events, on Team and Enterprise plans. They cover tool and connector calls, files read or changed, and skills used. They also show if each action was approved by hand or automatically.
Admins can limit which actions each MCP connector allows across the org. For example, they can allow reading but block writing.

Anthropic will now run your agents for you in the cloud

Developers can now have Anthropic host their agents, with sandboxes and long sessions built in.

Anthropic launched Claude Managed Agents on Apr 8. It's in public beta on the Claude Platform. You set an agent's tasks, tools and guardrails, and Anthropic runs it.

Sessions can run on their own for hours. Progress and outputs persist through disconnections. You pay standard token rates plus $0.08 per session-hour of active runtime.

Two features are in research preview, and you request access.

In one, agents direct other agents. In the other, you set outcomes and success criteria. Claude checks its own work and keeps going until it gets there.

Notion, Rakuten, Asana and Sentry built on it, per Anthropic. Notion's version is in private alpha.

Anthropic tested it on making structured files. Task success rose by up to 10 points over a standard prompting loop.

A new API tool lets Sonnet or Haiku ask Opus for help

A cheaper model can now consult Opus mid-task, inside one API request.

On Apr 9, Anthropic launched the advisor tool in beta. Sonnet or Haiku runs the task.

When it hits a decision it can't reasonably solve, it asks Opus. Opus sends back a plan, a correction or a stop signal.

The advisor never calls tools or writes what the user sees. Its tokens are billed at the advisor model's rates. You can cap advisor calls per request.

Anthropic tested Sonnet with an Opus advisor on SWE-bench Multilingual. It scored 2.7 percentage points higher than Sonnet alone. It cost 11.9% less per agentic task.

On BrowseComp, Haiku with an Opus advisor scored 41.2%, versus 19.7% alone, per Anthropic. That trails Sonnet alone by 29% in score. But it costs 85% less per task.

Claude Code now defaults to high effort for API, Team and Enterprise

Claude Code raised its default effort for some users. It also got a tool that watches background scripts.

Claude Code shipped versions 2.1.94 to 2.1.101 from Apr 7 to 10, per its changelog.

Default effort went from medium to high (2.1.94, Apr 7). That's for API-key, Bedrock, Vertex, Foundry, Team and Enterprise users. You can change it with /effort.
A new Monitor tool streams events from background scripts (2.1.98, Apr 9).
/team-onboarding writes a ramp-up guide for a new teammate. It's built from your local Claude Code use (2.1.101, Apr 10).
Focus view (Ctrl+O) works in NO_FLICKER mode (2.1.97, Apr 8). It shows your prompt, a one-line tool summary and the final response.

Quick hits

Anthropic launched the ant CLI (Apr 8). It's a command-line client for the Claude API. It works with Claude Code and keeps API resources as versioned YAML files.
The Messages API is on Amazon Bedrock as a research preview (Apr 7). It uses the same request shape as Anthropic's own API, per the release notes. It's in us-east-1, by request through your Anthropic account executive.
Claude.ai, Claude Code and other Claude services had elevated errors, including on login, on Apr 6 and Apr 7. The first ran from 15:00 to 16:30 UTC (10:00 to 11:30 am Central). The second ran from 14:32 to 15:12 UTC (9:32 to 10:12 am Central).
Sonnet 4.6 had errors above baseline on Apr 8. It ran from 7:00 to 9:50 UTC (2:00 to 4:50 am Central), per the status page.
Email login to Claude was broken on Apr 10. It ran from about 3:46 to 4:52 pm PDT (5:46 to 6:52 pm Central), per the status page.
OpenClaw creator Peter Steinberger's Claude account was suspended (Apr 10). Per TechCrunch, Anthropic's message cited "suspicious" activity, and he said it was back a few hours later. An Anthropic engineer told him Anthropic has never banned anyone for using OpenClaw.
Claude Code 2.1.101 fixed a command injection flaw (Apr 10). It was in a fallback Claude Code uses to find language server programs.
Claude Code 2.1.101 fixed deny rules being softened by hooks (Apr 10). Before, a PreToolUse hook could turn a deny into a permission prompt.
Claude Code 2.1.101 trusts your operating system's CA certificates by default (Apr 10). Enterprise TLS proxies now work without extra setup.
Claude Code 2.1.98 adds a Google Vertex AI setup wizard (Apr 9). It's on the login screen under "3rd-party platform."

Worth your time

Preparing your security program for AI-accelerated offense (Apr 10): Anthropic's guide for security teams after Project Glasswing. It has a section for groups without a security team.
Trustworthy agents in practice (Apr 9): Anthropic's five principles for agents, and how they shape its products.
Scaling Managed Agents with stable interfaces (Apr 8): how Anthropic's engineers built Managed Agents.
Multi-agent coordination patterns (Apr 10): five ways to get several agents to work together.

That's the week. The next issue lands Sunday, Apr 19.

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

— Sam