Big Trends & Ecosystem Shifts 🌎
OpenAI opened GPT-5.6 to the world on Thursday, ending a staggered launch it had run under White House pressure. The Sol, Terra, and Luna split covers frontier reasoning, balanced everyday work, and cheap-and-fast, with Sol tuned for biology, chemistry, and cybersecurity. All three are already available in GitHub Copilot.
SpaceXAI shipped Grok 4.5, its first model since going public and buying Cursor, and the first the two built together. Musk pitches it as Opus-class but faster and cheaper, at $2 and $6 per million tokens versus Opus 4.8's $5 and $25. It targets coding, agents, and knowledge work, and it is live now in Grok Build and in Cursor on every plan.
Anthropic and AE Studio published GRAM, a way to route dual-use knowledge like virology, cybersecurity, and nuclear physics into modules you can switch on or off per user. Deleting a module erases the capability about as well as never training on it, with general performance intact, and it resisted adversarial fine-tuning. It is early and not yet in production, but it argues for access control over blanket export bans.
Developer Tools 🛠️
Wiz disclosed GhostApproval, a symlink trick that gets AI coding agents to write files outside their sandbox, with a path to remote code execution. The agent reads that project_settings.json is really a shell config, then still shows you a clean prompt to approve the edit. It hits Amazon Q, Claude Code, Augment, Cursor, Antigravity, and Windsurf. AWS, Cursor, and Google patched, and Anthropic called it outside its threat model.
New reports suggest Perplexity has an internal coding tool, Teammate, that its engineers have used since May. It is model-agnostic and built for long-horizon work, owning projects, investigating issues, and monitoring services. A launch is unconfirmed, but the intent is clear: CTO Denis Yarats told staff that by year end engineers should stop looking at code and let AI do it.
Till next time,