Claude Code is having a very bad day

Anthropic’s worst day this year. Claude Code’s usage quotas are burning out in hours, not days—one Max 5 subscriber paying $100 per month exhausted an entire monthly allotment in a single session, and Pro users at $200 per year report getting roughly 12 usable days out of every 30-day cycle. Anthropic acknowledged that “people are hitting usage limits way faster than expected,” and cache bugs in the CLI are inflating token costs by 10–20x—downgrading to v2.1.34 reportedly helps, but peak-hour sessions between 8 AM and 2 PM ET consume more “usage units” per prompt, effectively creating surge pricing. A March promotion that doubled off-peak limits expired on March 28. Then came the second hit: security researcher Chaofan Shou discovered that a .map file in Claude Code’s npm package pointed to unobfuscated TypeScript source on Anthropic’s cloud—the complete proprietary CLI source code, exposed to anyone who looked. The post went viral on X with 19,000 likes and 9 million views, and it’s the second Anthropic leak in one week after the Mythos CMS incident. The company that positions itself as the responsible AI lab is now fielding two credibility crises simultaneously—one about reliability, the other about security.

While Anthropic patches leaks and quotas, Apple is building the tollbooth. iOS 27 will introduce “Extensions”—a framework that lets users install any third-party AI chatbot directly inside Siri, with a dedicated App Store section and Apple’s standard 30% commission on every AI-driven transaction. ChatGPT, Gemini, Anthropic, Meta, and xAI are all eligible. iOS 26.5 beta shipped with no Gemini features—everything deferred to the iOS 27 announcement expected at WWDC on June 8. It’s a concession that Apple can’t build the best model and a bet that it doesn’t need to—owning the distribution layer and taxing every AI provider at 30% is the same playbook that grew Apple Services into a $96 billion annual revenue stream. The labs racing to build the smartest model may find that the biggest margin goes to the company that controls where 1.5 billion users encounter it.

And in Europe, the infrastructure race is being financed differently. Mistral raised $830 million in debt—not equity—from 7 banks including BNP Paribas, HSBC, and Bpifrance to buy 13,800 Nvidia GB300 GPUs and build a 44-megawatt facility in Bruyères-le-Châtel near Paris, targeting operations by Q2 2026. Mistral has now raised $2.9 billion total, making it the best-funded LLM builder in Europe, and it aims for 200 megawatts of capacity across the continent by end of 2027. Debt financing means Mistral keeps its equity intact while locking in hardware at today’s prices—a bet that GPU access will be harder to secure in 12 months than the interest payments will cost. Three different strategies for the same industry: Anthropic scales a developer tool and stumbles on demand, Apple taxes everyone else’s AI, and Mistral borrows $830 million to own the physical layer.

Latest from PulseMark

AI Agent Framework Comparison 2026: Claude SDK vs LangGraph AutoGen’s 56k GitHub stars made it look like the winner—then Microsoft moved it to maintenance mode. We compared Claude Agent SDK, LangGraph, and CrewAI on protocol support, production readiness, and why CrewAI is the only major framework with both MCP and A2A. Read the analysis →