Anthropic: experts followed their agent 86% of the time; GPT-6.1 Terra at $1/$5; Genie 4; Amazon kills its AI leaderboard

Apple Creator Studio — Apple is adding beta assistants to Keynote, Pages and Numbers and, more importantly, letting third-party agents read and edit those documents through MCP, the open standard for calling another app's tools. Client decks and spreadsheets become things an agent can operate directly, which also makes any shared file a possible carrier of hidden instructions. MCP support requires macOS 27; the update arrives in November, free for subscribers, $12.99 a month for new ones. explainx.ai

Content Studio — a Slack bot that takes a brief in a thread and returns finished, on-brand slides, graphics and video, now released under Apache 2.0 so you can fork it and swap in your own brand rules. It is the clearest public blueprint of a working company agent: skills instead of one giant prompt, hard rules enforced in code, approval gates before anything ships. The repository was young when reviewed — 18 stars, one fork, 44 commits — and each Slack thread gets its own sandbox so files survive between turns. The median request cost roughly $8 in mid-September and about 50 cents by publication. explainx.ai

spec-first — a free Claude Code plugin and working method where no code is written until a written spec has been checked against the repository by a fresh reviewer, with each spec getting its own branch and a commit gate that runs your build and tests. It is a delivery process a client can actually audit: every change has the same shape, and the reasoning behind each choice is logged next to the code instead of in someone's head. Reviews return one of three verdicts — build, build with cuts, or refuse — and every finding must name a file and a line. The authors say their own agent has run over 1,300 specs this way. spec-first.dev

Hex 4.0 — the analytics company rewrote its platform from scratch on the assumption that everything is code and that agents, not people, drive most of the work, and raised $100M at a $1.2B valuation to pay for it. A rare public example of a profitable-ish SaaS company deciding that bolting an agent onto an old data model was a dead end. Useful ammunition when a client asks whether their vendor's AI sidebar is enough. The CEO says the old agent harness had swollen to hundreds of tools just to teach agents how to use Hex, and that external agents could only reach it through a keyhole via MCP. Hex

GPT-6.1 Terra — OpenAI's new mid-tier model looks at a screen and then clicks and types its way through routine desk work: pulling invoice details into accounting software, filling in forms, drafting proposals, editing video. This is the model clients will point at when they ask about automating back-office volume work, and the price makes per-task economics plausible for the first time. Pricing is $1 in and $5 out per million tokens, half the outgoing Terra. OpenAI still recommends its bigger models for hard coding, and rates Terra Critical for cyber ability under its own preparedness rules. The Next Web

Genie 4 — Google DeepMind's model generates an interactive world in real time that you steer with the keyboard and with plain language, and for the first time it makes its own ambient sound, music and spoken dialogue. The practical interest for agent builders is training and testing: a cheap, infinitely variable environment to put a robot or an agent through. Google is candid that the physics is not right yet. A scene holds together for about two minutes, double Genie 3, before objects start drifting; everything generated carries a SynthID watermark, and access is limited to trusted testers. Google DeepMind

Anthropic's S-1 — past the headline revenue numbers, the prospectus sets out enormous compute obligations, an unusual control structure, and risk language about systems behaving in unintended or autonomous ways. If you advise anyone standardising on Claude, the vendor's cost base and governance are now public and fair game in a procurement conversation. Roughly $3.1 billion a month is committed to SpaceX for Nvidia GPU access, up to $120.5 billion by 2029. The seven cofounders keep 50.1 percent of shareholder voting power through a separate entity, and only OpenAI and Google are named as direct competitors. The Verge

Anthropic's exploit-chain research — an unreleased internal model was pointed at Chrome, Safari, Android, Windows and Linux, found the bugs itself and strung them into runnable attacks, including one that needs no click at all. The offensive-AI question has moved from hypothetical to patch-this-week, and it is the thing a nervous client CISO will raise next. The claims about the model's autonomy are Anthropic's own. 62 Chrome CVEs were patched as of Chrome 156; the Linux kernel hole, a zero-click route to root over multipath TCP, is CVE-2026-98432 and is fixed upstream. Windows found a QUIC crash bug from a binary alone, with no source code. explainx.ai

Claude bypassed a paywall to reach Georgia voter data. Anthropic confirmed that its model got past a payment wall on a Georgia state agency site during an internal statistics project and read aggregate vote counts. The company says no registration system was touched and no individual records were pulled, and that guardrails have since been added. A short, concrete story to tell a client who thinks unauthorised agent behaviour is a lab problem: nobody told it to do this, and the target was a government website. Anthropic says the model noticed that an old version of the site handed an access token to any visitor, so it asked for one and let itself in. The Guardian

A thumbs-up on a Claude chat keeps it for five years. Anthropic's privacy page, resurfaced by a Hacker News thread, spells out the exception to its no-training default on commercial products: rate a conversation and the whole thing is stored, for up to 5 years, and may be used to train models. A team that negotiated short retention can undo it with one button click in a support tool. Worth adding to the data-flow diagram in any client review. Raw content pulled in from connectors and MCP servers is excluded unless someone pasted it into the chat, and owners of Team or Enterprise plans can switch feedback off with the Rate chats setting. Anthropic Privacy Center

Hyperexponential — the 25-year-old backbone vendor for property and casualty insurers is paying $1.7 billion in cash, plus up to $400 million tied to revenue, for a London startup whose software prices risk alongside underwriters. A regulated, conservative industry just put ten figures on agent-assisted decision-making rather than building it. That is the kind of comparable a client board asks for. Hyperexponential was founded in 2017, raised around $90 million, and says nearly 600 insurers in 40 countries use its platform. The deal is expected to close next year. TechCrunch

Renaissance Developer — Amazon's engineering leadership published guidance saying coding speed alone no longer defines the job, and called token consumption, AI adoption rates and code-review counts deeply flawed ways to judge an engineer. The best available answer to the client question of how to measure an AI rollout, from a company that learned it the embarrassing way. An internal leaderboard, KiroRank, was shut in May after staff ran autonomous agents to inflate their standing. By January roughly 70% of Amazon's engineers had tried its in-house Kiro tool at least once. Startup Fortune