The Claude ecosystem had a turbulent week. While Anthropic continued shipping at a furious pace โ Claude for Word in beta, the Advisor Tool in public beta, and 10+ Claude Code releases โ the conversation shifted to uncomfortable questions. Users reported significant performance degradation in Opus 4.6, Anthropic lost an appeals court bid to block the Pentagon’s “supply chain risk” designation, and the US Treasury banned Anthropic products. Meanwhile, a CoreWeave infrastructure deal and early exploration of custom AI chips signal the company is preparing for compute demands it cannot yet fully meet. Here is this week’s Claude Weekly.
The Performance Backlash
The week’s most closely watched story was not a product launch โ it was a trust problem.
Fortune and VentureBeat reported growing user backlash over perceived Claude Opus 4.6 performance decline. Claude Code lead Boris Cherny confirmed that Anthropic reduced the default “effort” level to “medium” to manage token consumption โ a change made without public announcement.
Independent analysis found a ~67% decrease in reasoning depth across 6,800+ Claude Code sessions. A separate cache TTL controversy โ users claiming a silent reduction from 1 hour to 5 minutes โ drew 724 upvotes on Reddit with analysis citing 120,000 API calls.
For enterprise teams, this matters because it raises questions about model consistency guarantees. Paid Claude Code users now default to effort level “high,” but the transparency gap has not been addressed. With IPO preparations underway and revenue doubling quarterly, user trust is the one asset that cannot be repurchased.
Pentagon, Treasury, and a Lobbying Hire
Three developments defined Anthropic’s regulatory week โ and none of them were good.
The D.C. Circuit Court denied Anthropic’s emergency injunction to pause the Pentagon’s “supply chain risk” designation (originally issued March 4). Two Trump-appointed judges sided with the government’s national security interests. The underlying dispute: Anthropic’s refusal to permit Claude for domestic mass surveillance or fully autonomous weapons. Defense Secretary Hegseth accused Anthropic of seeking “veto power over U.S. military decisions.”
Days later, Anthropic hired Ballard Partners, a Trump-linked lobbying firm, for DoD procurement advocacy. And the US Treasury Department banned Anthropic products โ a signal of governance headwinds even as enterprise adoption accelerates.
Separately, Treasury Secretary Bessent and Fed Chair Powell summoned Wall Street CEOs for an urgent meeting on cyber risks arising from Claude Mythos Preview capabilities. The full Pentagon case remains undecided. Outcome could define the AI-military relationship for the entire industry.
Claude for Word: Competing on Microsoft’s Turf
Claude for Microsoft Word launched in beta on April 11, targeting legal, finance, and HR document workflows. The key differentiator: Claude’s edits appear as Word Tracked Changes, preserving the review workflow lawyers and compliance teams already depend on.
Features include sidebar-based editing with formatting preservation, 50+ page document handling, and cross-app context sharing with Excel and PowerPoint add-ins. Available for Team and Enterprise plans via the Microsoft Marketplace.
This is Anthropic competing directly with Microsoft Copilot on Microsoft’s home turf. For enterprise IT directors evaluating document AI, it is the first real alternative inside the Office suite itself.
Claude Code: The Densest Week on Record
Claude Code shipped 10+ releases (v2.1.92 through v2.1.105) with four headline features:
- Ultraplan (research preview): Cloud-based codebase planning from the terminal. Three parallel exploration agents plus a critique agent draft plans in a web session for collaborative review before execution.
- Monitor Tool (v2.1.98): Background watchers that stream events into conversations โ tail logs, watch CI, auto-fix crashes without sleep loops.
- /autofix-pr: Automatic PR fixing from the CLI. Claude watches CI and review comments and pushes fixes until green.
- Routines (April 14): Repeatable, structured workflows defined directly in Claude Code.
Other notable additions: Focus view (Ctrl+O), Perforce mode for enterprise SCM, guided Bedrock/Vertex AI setup wizards, and a 60% speed improvement for the Write tool.
One security item: CVE-2026-35020 disclosed an OS command injection vulnerability in the Claude CLI via the TERMINAL environment variable. Update to the latest version.
Advisor Tool: Multi-Model Workflows as a First-Class API
The Advisor Tool entered public beta on April 9, introducing a new server-side tool that pairs a fast executor model (Sonnet or Haiku) with a high-capability advisor (Opus) for mid-generation guidance.
The numbers justify the design: Haiku with an Opus advisor more than doubled its standalone SWE-bench score (19.7% to 41.2%) at 85% less cost per task than Sonnet alone. For teams running long-horizon agentic workloads, this is the most cost-efficient approach Anthropic has shipped. Requires beta header advisor-tool-2026-03-01.
Infrastructure: CoreWeave, Custom Chips, and the Compute Squeeze
CoreWeave signed a multi-year deal to power Claude workloads with Nvidia GPUs in U.S. data centers, sending CoreWeave stock up 10%. This sits alongside the previously announced Google-Broadcom 3.5 GW TPU capacity commitment for 2027.
Reuters also reported that Anthropic is in preliminary stages of exploring custom AI chip design โ a move to reduce reliance on third-party silicon as demand outpaces supply. No timelines or partners were disclosed.
The infrastructure picture connects directly to the performance backlash: surging adoption (from $19B ARR in March to $30B in April) has outrun compute capacity, and the effort-level throttling may be an operational consequence.
MCP Ecosystem Crosses 97 Million Monthly Downloads
The Model Context Protocol hit 97 million monthly SDK downloads with 5,800+ servers in the wild. The Linux Foundation’s Agentic AI Foundation now governs the protocol.
Key developments this week: AWS Bedrock added stateful MCP client capabilities enabling bidirectional agent conversations. Cloudflare published an enterprise MCP reference architecture for securing non-human identities at scale. WordPress Core committed to MCP integration, and Shopify launched an AI Toolkit supporting Claude Code. MCP is no longer a developer tool โ it is becoming enterprise infrastructure.
What to Watch
- Performance transparency. The effort-level and cache TTL controversies are the most immediate risk to user trust. A public, data-backed response from Anthropic would go a long way โ silence will not.
- Pentagon supply chain risk. Full case undecided. The Ballard Partners hire signals a sustained fight. Combined with the Treasury ban, Anthropic faces a regulatory environment that could constrain government and defense sector adoption.
- Managed Agents at enterprise scale. Rakuten deployed five agents across departments in one week. Watch for SOC 2 and HIPAA compliance certifications โ those will determine whether regulated industries can adopt.
That is this week’s Claude Weekly. For the first time, the dominant story is not what Anthropic shipped but how it communicated โ or failed to communicate โ the trade-offs behind a product growing faster than its infrastructure can comfortably support. Check back next week for the latest.