Today's Overview
A model costing $0.02 per API call just scored 78.2% on SWE-bench Verified-the industry benchmark for AI coding agents-outperforming Claude Opus 4.5 at a fraction of the cost. The win came not from a better model, but from better architectural context fed through an MCP (Model Context Protocol). Researchers at Xanther showed that providing agents with a structural map of a codebase-understanding which modules depend on which-cuts token usage by 20% and lets cheaper models solve harder problems. This changes the maths for startups building coding tools. You don't need GPT-5-level intelligence; you need intelligent retrieval.
Canvas outage hits colleges mid-exams
The learning management platform Canvas went offline Thursday after a cyberattack, the same threat actor responsible for a breach disclosed a week earlier. Schools and colleges across the US postponed final exams. Canvas came back online Friday, but the incident exposed a harder problem: educational infrastructure sits on a thin edge between availability and trust. Instructure disclosed that attackers accessed usernames, email addresses, student IDs, and messages-data that affects millions of students. The ransomware group ShinyHunters claimed responsibility and stated they accessed records from 8,800 schools.
Cloudflare lays off 1,100 because AI got efficient
Cloudflare CEO Matthew Prince announced the company's first large-scale layoff, removing 1,100 positions-roughly 10% of headcount-because AI efficiency improvements made support roles redundant. Revenue hit record highs, and the company's profit margin expanded, yet jobs disappeared anyway. This pattern is accelerating: Expedia reported that over 30% of customer service interactions now run on AI. Oracle cut hundreds of engineering roles and refused to negotiate severance with laid-off workers. The efficiency gains are real. The human cost is equally real.
Three separate stories, one emerging reality: AI is no longer experimental. It's operational. It cuts costs. It disrupts work. And it forces hard questions about who benefits from the efficiency.