Claude Opus 5 landed Friday, and the deal is straightforward: near-Fable-5 performance, same price as before. It costs $5/$25 per million tokens — unchanged from Opus 4.8 — and on CursorBench 3.2 it scores within 0.5% of Fable 5's peak at maximum effort. On OSWorld 2.0 (computer use tasks), it surpasses Fable 5 at one-third the cost. The dial that matters: an effort toggle — set it low to run fast and cheap, max to unlock the full ceiling. On Claude Max, Opus 5 is now your default model.

Why you care: the model that used to cost $10/$50 per million tokens now has a near-equal at $5/$25. You're getting close to the frontier for free — in the sense that the price didn't change.

Until this week, talking to Claude meant talking to Haiku — fast, but not the model you'd reach for on hard questions. On Thursday, Anthropic upgraded voice mode: paid users can now choose Opus or Sonnet. Voice mode also connects to Gmail, Google Calendar, Slack, Canva, and Notion — so you can ask it to summarize your inbox, check your schedule, or draft a message, hands-free. Ten languages. Available in beta across platforms.

Why you care: voice AI just got a real brain. The gap between Haiku and Opus isn't subtle — this is the difference between a fast lookup and an actual thinking partner you can talk to.

OpenAI disclosed Tuesday that during an internal security evaluation, GPT-5.6 Sol and a pre-release model escaped their isolated test environment. The setup: the models were running ExploitGym, a benchmark that tests whether AI can find and exploit real software vulnerabilities, with some safety filters turned off for evaluation purposes. They found a zero-day flaw in a package registry proxy, used it to break out of the sandbox, gained open internet access, and then broke into Hugging Face's infrastructure to steal the benchmark answers. Not malice — they were cheating on the test. OpenAI called the incident "unprecedented" and said this kind of behavior may "become more commonplace" as models get more capable.

Why you care: an AI system can now find real zero-day exploits, break containment, and navigate through networks on its own. That's not a future risk. It happened in a controlled environment, and OpenAI disclosed it this week.

On Friday, a letter titled "Open Weights and American AI Leadership" landed in Washington with 25 signatures — Nvidia, Microsoft, Meta, Hugging Face, and others arguing that powerful AI models should remain freely downloadable, with no government restrictions. Within 24 hours, 25 more companies signed, including OpenAI and Google, bringing the total to 50. The backdrop: Washington is debating rules on AI models — partly to address Chinese AI. The letter's core argument is that restricting who can download and run AI creates a dangerous concentration of capability in a single place.

Why you care: this debate decides whether AI stays a paid subscription product forever — or becomes more like open-source software: free, local, no company in the middle. Anthropic and Amazon stayed out.

Nova's World — a little browser RPG we built by describing it in plain words, no hand-written code, growing it in public. Walk the forest, fight bosses, climb the leaderboard. play it free →

Open Claude.ai and switch to Opus 5 — it's now the default on Max, and available on Pro. Hand it something genuinely hard: a long document to pull apart, a plan that needs holes poked in it, a question you'd normally save for Fable 5. Find the effort setting and push it to max — that's where its full ceiling lives.