The Week in AI

Mark Gurman's report on May 5. Caveat: this is a Bloomberg leak, not an Apple announcement. But Gurman is Apple's most reliable reporter — when he says "people with knowledge of the matter," it almost always happens.

The plan for iOS 27 / iPadOS 27 / macOS 27 (fall 2026): a new "Extensions" system that lets users pick which AI model powers which Apple Intelligence feature. Different models for different tasks. Claude for Writing Tools. Gemini for Image Playground. Something else for Siri. Your call.

AI providers integrate by shipping App Store apps that hook into Extensions. And the juicy bit: each model can have its own Siri voice. You'll literally be able to give Claude one voice and Gemini another.

Apple's been moving this direction quietly — ChatGPT integration since iOS 18, an upcoming Gemini deal with Google. Extensions is the full opening of the platform. Watch WWDC in June for the official reveal.

Announced May 12 at The Android Show. "Gemini Intelligence" is umbrella branding for a suite of features: screen understanding, cross-app task completion, automation. Not a new product — a feature suite. Rollout summer 2026 on flagship Pixel + Samsung.

This is Google's pre-I/O counter to Apple. Two operating systems picked different strategies in the same week. Apple says "you pick your AI." Google says "we picked Gemini for you, and it's now everywhere."

The biggest AI governance lawsuit ever filed just finished its trial. Judge Yvonne Gonzalez Rogers (Northern District of California) sent the case to a 9-person advisory jury Thursday afternoon. Deliberations begin Monday, May 18.

What Musk wants: $150 billion in damages (to go back to the OpenAI nonprofit), Sam Altman and Greg Brockman removed from leadership, and the entire for-profit conversion unwound. What OpenAI wants: dismissal. The jury advises. The judge ultimately decides liability and remedies.

Monday: Satya Nadella testified for 3 hours. Said Microsoft's investments were approved by OpenAI's nonprofit board, Musk never raised concerns with him directly, and Altman's 2023 ouster was a "blip."

Tuesday: Sam Altman on the stand. Said Musk originally asked for 90% of OpenAI's equity, proposed merging the company with Tesla, and in a 2017 conversation suggested passing control to his children if he died. Altman on whether he's trustworthy: "I believe so."

Wednesday: OpenAI's chief safety researcher described a 2018 meeting where Musk called him a "jackass" over safety priorities. OpenAI tried to enter an actual gold-plated jackass-rear-end trophy as physical evidence. The judge declined.

Musk's lawyer Steven Molo, on Altman's credibility: "Imagine you're on a hike about to cross a bridge, and a woman at the entry says, 'Don't worry, the bridge is built on Sam Altman's version of the truth.' Would you walk across that bridge? I don't think many people would."

OpenAI's lawyer William Savitt: "Mr. Musk looks at all of this and shouts, 'They stole a charity.' That's sort of all he does: shouts." OpenAI's frame: Musk wanted control, didn't get it, left in 2018, started xAI, now suing.

Musk himself was absent for closings. Reportedly in China.

OpenAI's joint venture (announced in Ep40) shipped product. It's a services entity that embeds Forward Deployed Engineers inside customer organizations — an Andreessen-Horowitz-style consulting model, not a product portal. Real engineers walking into Fortune 500 offices to redesign workflows around ChatGPT.

Advertisers can now buy contextual ads inside ChatGPT responses. The format is search-style contextual CPC, no video yet. Revenue target: $2.5 billion this year. $100 billion per year by 2030. For context: $100B/year would put ChatGPT ad revenue near Google Search's ad revenue today. That's the bet.

Same underlying model, two names. A defensive-cyber variant of GPT-5.5 with stronger account controls + permissions, for authorized red-teaming and pen-testing teams. EU got it. Pentagon already had it (per Ep40). Anthropic's Mythos still gated. AI for cyber is bifurcating geopolitically.

Starting June 15, 2026, every paid Claude plan gets a separate monthly credit that's ONLY for programmatic use of Claude (Claude Code, claude -p, GitHub Actions, third-party apps built on the Agent SDK like Conductor or OpenClaw).

Right now, all your Claude usage shares ONE bucket — chat at claude.ai and any automation behind the scenes eat from the same pool. After June 15, your chat limits stay the same and you get a NEW separate bucket for the code-driven stuff. The credits: $20/mo for Pro, $100/mo for Max 5x, $200/mo for Max 20x, $20/seat for Team Standard, $100/seat for Team Premium.

The catch: credits don't roll over, and you have to click a button in an email (arrives before June 15) to claim it. Once claimed, it refreshes automatically each billing cycle.

Why are they doing this? Not generosity. It's margin protection. Heavy-automation users were getting near-unlimited compute on subscription plans. The split caps that and pushes serious automation to the API. For most viewers watching this show that's free unused capacity. For indie devs who built third-party apps on subscription auth, the economics just changed.

Anthropic dropped a new product tier specifically for small businesses. Plus they're hitting the road with a free half-day AI fluency training tour — first stop Chicago on May 14, 100 small business leaders per city. This is the first time Anthropic has made SMBs an explicit segment with their own pricing + training.

Technically May 6 at Code w/ Claude SF, but heavy coverage spilled into this week. Dreaming = a scheduled background process that reviews past agent sessions, extracts patterns, and rewrites the agent's memory store. Self-improvement loop without retraining. Harvey (legal AI) saw a 6x increase in task completion rate after enabling Dreaming.

First public product preview from the lab Mira Murati started after leaving OpenAI as CTO. The product: an interactive AI that listens and talks at the same time — simultaneous, not turn-by-turn. Thinking Machines has been stealth since 2024 with $2B at a $12B valuation. The frontier-lab map just expanded past Anthropic / OpenAI / xAI / Google.

Part of Project Glasswing (Anthropic's initiative that launched in April with 12 partners including Apple, AWS, Google, Microsoft, JPMorgan).

A 27-year-old vulnerability in OpenBSD — a system famous for security hardening. A 17-year-old remote code execution exploit in FreeBSD (CVE-2026-4747) that Mythos discovered AND exploited fully autonomously — no human involved after the initial "go find bugs" prompt. This week added new findings in Apple's Mac M5 hardware defenses (patches in progress).

The concerning data point: fewer than 1% of vulnerabilities Mythos has found have been patched. Defenders got the tool first. The race is to patch before adversary versions arrive.

An AI co-editor that builds video drafts on the timeline from a text prompt, places sound effects frame-accurately, and lets users interrupt to take back control. Part of the ElevenCreative subscription — GA, not beta.

Better reasoning, more natural back-and-forth. Closes the reasoning gap that's been the gating problem for voice AI. Edge claimed over Sesame and xAI's Grok Voice Mode.

ByteDance's video model. Native 4K (not upscaled). Competitive with Sora 2 and Veo 3.1.