AWS Begins Transitioning Pentagon Workloads Off Claude
Amazon Web Services confirmed it is moving Department of Defense workloads off Anthropic’s Claude following the Pentagon’s designation of Anthropic as a supply-chain risk. The transition applies specifically to defense use cases — AWS stated clearly that non-defense customers and partners can continue using Claude across all workloads without disruption. AWS is actively supporting affected defense customers as they migrate to alternative models running on AWS infrastructure.
This marks the most concrete commercial fallout from the Pentagon standoff to date. While Anthropic’s direct DoD revenue is relatively small, the AWS channel represents significant enterprise reach. The bigger risk is optics: cloud buyers in adjacent regulated sectors — federal civilian agencies, defense contractors, intelligence community primes — may now second-guess Claude deployments. The March 24 court hearing will be critical. A favorable ruling could reverse the supply chain designation and stop the bleeding before it spreads further.
Cowork Persistent Agent Threads Now Rolling Out — Max Plans First
Anthropic began rolling out persistent agent thread support in Cowork today. The feature lets users manage long-running tasks from both mobile and desktop, maintaining context across sessions rather than starting fresh each time. Max plan subscribers get access starting March 18, with Pro plan access following over the next two days. Team and Enterprise plans are included in the rollout.
Persistent threads are a foundational building block for real agentic workflows. Right now, Cowork tasks die when a session ends. With persistent threads, you can kick off a multi-step research or file management task, close the app, come back hours later, and pick up where you left off. This moves Cowork from a powerful single-session tool toward something that can genuinely run background work on your behalf. Expect more agentic features to layer on top of this infrastructure in coming weeks.
Claude Code Voice Mode Goes Live for 5% of Users — 20 Languages Supported
Anthropic’s Claude Code voice mode is now active for approximately 5% of Pro, Max, Team, and Enterprise subscribers, with a progressive rollout planned through the rest of March. The feature activates via the /voice command and uses a push-to-talk model: hold the spacebar to speak, release to send. There is no always-on listening. The feature ships with support for 20 languages, up from the initial 10 at launch. Voice mode is included at no extra cost for all eligible plans. Users can remap the push-to-talk key in keybindings.json.
Push-to-talk is the right call for a coding context. Always-on voice in a dev environment would be a privacy and noise nightmare. The spacebar mechanic is intuitive and low-friction — you can dictate a complex refactor request without taking your hands off the keyboard for a mouse click. The real test will be latency and accuracy when describing code-specific tasks like “extract this into a generic utility function and update all call sites.” Early impressions from the 5% cohort should surface on X and Reddit this week.
The /loop Command Turns Claude Code into a Recurring Monitor
Claude Code’s new /loop command lets you run any prompt on a repeating interval within the current session. Syntax: /loop 5m check the deploy runs your prompt every 5 minutes. It works like a lightweight in-session cron job — no shell scripting or external schedulers needed. The loop runs until you stop it or close the session.
This is deceptively powerful for developer workflows. Watch a build pipeline and alert on failures. Poll an API endpoint for a state change. Re-run a flaky test every minute until it passes. The constraint is that it’s session-scoped, so it stops if you close Claude Code — but paired with the new persistent agent thread infrastructure in Cowork, you can see where Anthropic is heading: agents that run continuously, not just on demand. The /loop command is available now in the latest Claude Code release.
Opus 4.6 Output Limit Jumps to 64K Tokens — Upper Bound Now 128K
Anthropic increased the default maximum output token limit for Claude Opus 4.6 to 64,000 tokens, up from the previous default. The upper bound for both Opus 4.6 and Sonnet 4.6 is now 128,000 tokens. These increases apply automatically — no API changes required. The higher limits are especially relevant for long code generation, document drafting, and multi-step agent outputs where previous limits forced chunking workarounds.
TIME: Anthropic Is Now “The Most Disruptive Company in the World”
TIME magazine published a deep-dive profile of Anthropic this week, naming it the most disruptive company in the world for 2026. The piece covers the Pentagon standoff, Anthropic’s rapid revenue growth (reportedly approaching $20B ARR), and CEO Dario Amodei’s vision for AI safety-first development at frontier scale. The profile notes that Claude is now the only AI model deployed on classified Pentagon systems — a fact that makes the supply chain designation battle even more consequential.
Opus 4.6 Elevated Errors Overnight — Fourth Incident This Month
Claude experienced elevated API error rates on Opus 4.6 overnight on March 17–18, marking the fourth service incident in March. Anthropic implemented a fix and the incident was resolved by 00:25 UTC on March 18. This incident was narrower in scope than earlier outages this month, affecting primarily Opus 4.6 API calls rather than Claude.ai broadly. Status updates were posted to status.claude.com throughout the incident.
The AWS Move Tells You Where Enterprise Risk Is Heading
AWS shifting defense workloads off Claude is not just about the Pentagon — it’s a signal to every regulated enterprise buyer watching from the sidelines. AWS is the most conservative cloud provider in the world when it comes to compliance optics. When Amazon says “we are helping customers migrate,” that language is designed to protect them from liability, not from Anthropic. It tells you that in regulated sectors, the supply chain designation now creates real procurement risk even for customers with zero defense ties.
Anthropic’s best move is winning in court on March 24. A judge blocking or reversing the designation removes the compliance question entirely. If that doesn’t happen, expect Microsoft, Google, and Oracle to issue similar “defense customers should transition” guidance quietly over the following weeks. The window for Anthropic to contain the commercial fallout is narrowing fast.