Latest from the Academy

Cursor Composer 2.5 reaches 62 on Artificial Analysis Coding Agent Index at $0.07/task — 60× cheaper than Claude Code's frontier tier

Open Notebook's full REST API + Pandoc Typst + Kokoro TTS replaces NotebookLM for any team running batch Markdown-to-PDF-and-audio pipelines in 2026

Continue.dev v1.4 ships MCP JSON import — you can drop your Claude Desktop or Cline config directly into .continue/mcpServers/ and it just works

Claude Computer Use now has two separate execution contracts — API route for builders, Cowork for delegation — and conflating them produces either over-engineered or insecure systems

A silent March 2026 TTL change added $949 in excess cost across 119,866 calls — and most teams still don't know about it.

Context engineering subsumes prompt engineering — the prompt is one instruction file inside a larger control system you now have to design first

NotebookLM is becoming the polished Google learning Studio; Open Notebook is the controllable local automation layer, so the winning setup routes jobs instead of choosing one product.

Claude for Small Business is an SMB activation stack, not just a connector bundle.

MCP registry risk comes from packaging trust, not just model behavior: one-click installs now move faster than provenance, signing, and review.

The practical split is now clear: Opus 4.7 for repo-scale refactors, GPT-5.5 for terminal-heavy agents, with both routable through Bedrock

OpenAI GPT models now run on AWS Bedrock under IAM auth — not API keys — with data that never leaves your AWS region

Cursor 3.2 and Claude Code both landed parallel-subagent runtimes in April 2026 — same architecture, opposite ergonomic bet