Managed Agents Reach Public Beta: Cron Schedules and Credential Vaults
Anthropic moved two Managed Agents capabilities into public beta on the Claude Platform: scheduled deployments and secure access to CLI tools and authenticated services. A scheduled deployment hands an agent a cron schedule — when it fires, the agent spins up a fresh session, finishes the task, and shuts down, with no scheduler for you to build or host. Owners can pause, resume, archive, or fire an extra run on demand.
The credential handling is the clever part. The agent’s sandbox holds only a placeholder; the real API key is attached at the network boundary, and only on requests to domains the customer allowlists. Even if the agent is hijacked by a prompt injection, the actual secret never enters the model’s context. Anthropic attached no separate price — scheduled runs bill under existing Claude Platform usage. Rakuten is already running weekly and monthly spreadsheet-to-report jobs on it.
Project Glasswing Expands to ~150 Orgs, Adds Claude Security
Anthropic widened Project Glasswing — the vetted-access program that gates its highest-capability models — to roughly 150 new organizations, extending Claude Mythos Preview access to a larger circle of cyber defenders and critical-infrastructure operators. The expansion lands the same week Fable 5 went public, reinforcing the “same model, different doors” structure: the public gets the safeguarded version, while the unrestricted twin stays inside Glasswing.
The program also picked up Claude Security, a capability aimed at scanning codebases and proposing patches. Pitching the unrestricted model to defenders as a security tool is the clearest expression yet of Anthropic’s bet that frontier capability is safest when it reaches the people defending systems before it reaches everyone else.
Four Days Out: Sonnet 4 and Opus 4 Retire From the API June 15
The clock is short now. Claude Sonnet 4 (claude-sonnet-4-20250514) and Opus 4 (claude-opus-4-20250514) retire from the API at 9am PT on Monday, June 15. After that, requests to those model IDs — including the claude-sonnet-4-0 and claude-opus-4-0 aliases — return errors, and any production app still pointing at them breaks.
For most users the fix is a one-line model-string swap: move to claude-sonnet-4-6 and claude-opus-4-8, with prompts, tool-use conventions, and JSON formats stable across the 4.x line. Consumer Claude.ai and Claude Code managed environments aren’t affected — Anthropic handles model selection there. If you own API code, this weekend is the last comfortable window to test and ship the change.
SpaceX Compute Deal Feeds Higher Limits for Pro and Max
Anthropic signed an agreement to take the full compute capacity of SpaceX’s Colossus 1 data center, and it’s framing the deal in user terms: more headroom for Claude Pro and Claude Max subscribers. The pitch is that raw capacity is now the binding constraint on what a frontier lab can offer, and locking up a data center translates fairly directly into higher usage ceilings.
For developers, the read is about reliability and rate limits as much as raw access. Capacity deals like this are what let Anthropic bundle a model as expensive as Fable 5 into paid plans, even briefly, and what determine whether your agent workloads hit throttling during peak hours. Infrastructure isn’t glamorous, but it’s increasingly the story behind every limit you feel.
Anthropic Files Confidentially for an IPO
Anthropic confirmed it has confidentially submitted a draft S-1 to the SEC, teeing up a public-market test of the AI boom. Reporting around the filing puts the company at a roughly $965B valuation after a $65B raise, with annualized revenue cited near $47B from Claude subscriptions and platform usage — numbers that would make it one of the largest tech IPOs in recent memory.
A confidential filing doesn’t lock in a timeline, but it signals intent and starts the regulatory clock. For customers, the relevant subtext is durability: an Anthropic answerable to public shareholders is a different kind of vendor than a private lab, with more disclosure and more pressure on the unit economics behind the models you depend on.
Claude’s Legal Push Keeps Pulling In the Incumbents
Anthropic’s Claude-for-legal play — 20+ MCP connectors and 12 practice-area plugins spanning corporate, employment, privacy, IP, regulatory, and litigation work — is still reshaping legal-tech alignments. The connectors wire Claude into the tools firms already run, from DocuSign, Ironclad, and Box to research platforms like Thomson Reuters and the Free Law Project, plus AI assistants Harvey and Solve Intelligence.
The notable part is who’s leaning in. Harvey, Relativity, Everlaw, and Thomson Reuters are integrating deeply rather than competing head-on, betting that sitting inside the Claude ecosystem beats sitting outside it. It’s a preview of how vertical AI is consolidating: the model becomes the platform, and the established vendors become connectors on it.
“I Don’t Know” Is the Whole Problem With Military AI
Asked whether Claude played any role in a strike that killed an estimated 120 children at an Iran elementary school, Anthropic’s CEO said he doesn’t know. That answer isn’t evasion so much as a confession about the architecture of selling AI to defense buyers: once a model is deployed inside classified systems, the vendor loses visibility into how it’s used. Usage policies and acceptable-use terms are paper; they don’t generate telemetry from a battlefield.
This is the unresolved tension under every defense AI deal in the industry, not just Anthropic’s. Labs want the revenue and the “democracies should lead” framing; they also want to claim their safeguards mean something. But a safeguard you can’t observe being honored is a hope, not a control. Expect “I don’t know” to become the line that critics, and eventually regulators, refuse to accept.
Scheduled Agents Are the Quiet Half of the Autonomy Story
Fable 5 got the headlines, but Managed Agents on a cron timer may matter more to how work actually changes. A model you prompt is a tool; a model that wakes up at 2am, runs a task against live systems, and goes back to sleep is closer to a coworker. The credential-vault design — secrets attached at the network edge, never in context — is a genuinely thoughtful answer to the obvious objection, that an autonomous agent with real keys is an autonomous agent that can be tricked into misusing them.
The thing to watch is the gap between “runs unattended” and “runs correctly unattended.” Scheduling removes the human from the loop by design, which is the point and the risk in one move. The teams that win with this won’t be the ones that automate the most; they’ll be the ones that build the tightest observability around what their agents did while no one was watching.