This Week in AI: Four Days to I/O, Eight Models Going Dark, and a $950B Number

Google I/O lands Tuesday, May 19, at 10 AM Pacific. That is the calendar everyone in AI is writing against right now. Android Show went out Monday. Anthropic let the $950 billion number leak Tuesday. xAI is retiring eight models at noon Pacific today, four hours from the time of this post. Apple started rewriting App Store rules for autonomous agents. Amazon killed Rufus. The pre-I/O week was louder than the keynote it is supposed to lead into.

The unifying thread is positioning. Every major lab and platform is trying to set or reset the narrative before Sundar Pichai walks on stage Tuesday and tries to make Gemini 4 the only thing the industry talks about for two weeks. Here is the full roundup.

1. Google Sandbagged Its Own Keynote

On Monday May 12, Google ran The Android Show: I/O Edition and used it to ship the most significant Android announcement in years: Gemini Intelligence. Agentic AI baked into the OS, generative UI widgets you describe in natural language, a Gboard mode called Rambler that takes spoken filler-word ramble and formats it cleanly, and cross-app task automation that copies a grocery list out of Notes and lands it in your shopping app of choice.

The strategic read is unusual. Google front-loaded a feature drop that would have anchored the I/O keynote in any other year. Why give that up a week early? Two reasons. First, Apple WWDC lands June 8, three weeks after I/O, and Google wants Android's agentic story locked in before Apple gets the keynote stage. Second, I/O itself is now reserved for Gemini 4, the model, and Google does not want the model news competing with platform news on the same day.

Initial rollout is Galaxy and Pixel this summer, with watches, cars, glasses, and laptops to follow before year end. The honest read: Google just turned every Android device into an agent runtime, then handed Apple a 27-day window to respond. We covered the timing dynamics of that gap in our analysis of Apple's WWDC counterprogramming math. The window just got harder.

2. Anthropic Priced Talks at $900B to $950B

Bloomberg reported Tuesday that Anthropic is in talks to raise $30 billion at a $900 billion valuation. By Wednesday, the higher end had moved: the New York Times put the range at $30 billion to $50 billion at up to $950 billion. The board reportedly meets this month to set a final number.

We covered the $900 billion thesis when it first surfaced in our piece on Anthropic lapping OpenAI on valuation. What changed this week is the upside: $950 billion is the number that puts Anthropic one funding round away from a trillion-dollar private valuation, which would be a first.

The ARR line is the one that matters. Anthropic has gone from a roughly $9 billion run rate at the end of 2025 to about $30 billion by Q2 2026. That is the fastest ramp from a frontier-model business in history, and it is what is supporting the valuation walk-up. The $40 billion Google equity commitment from April and the $25 billion Amazon commitment are still in the mix; the new round is on top of, not instead of.

3. xAI Sunsets Eight Models at Noon Pacific Today

As of 12:00 PM PT today, xAI is retiring eight models from the API: grok-4-1-fast-reasoning, grok-4-1-fast-non-reasoning, grok-4-fast-reasoning, grok-4-fast-non-reasoning, grok-4-0709, grok-code-fast-1, grok-3, and grok-imagine-image-pro. Requests to those slugs auto-redirect to grok-4.3, billed at the new rate ($1.25 in / $2.50 out per million tokens) regardless of where you came from.

The retirement notice went out May 6. Nine days from email to model going dark. For single-developer projects that is fine. For anyone with a production pipeline, that is a tight window for QA, eval re-runs, and prompt-template adjustments. xAI's consolidation push is real, and it is faster than anything OpenAI or Anthropic has ever asked customers to swallow.

If you are still pointing at any of those slugs, run a final eval on your golden test set against grok-4.3 today. Token-level outputs change between models even when they benchmark similarly. We track the live model catalog at tensorfeed.ai/models and added the deprecation flags to the xAI rows this morning.

4. Amazon Killed Rufus and Replaced It With Alexa for Shopping

On Wednesday May 13, Amazon launched Alexa for Shopping, an Alexa+ powered assistant that replaces Rufus across mobile, desktop, and Echo Show surfaces. Rufus shipped less than two years ago as Amazon's answer to ChatGPT-for-shopping. It is now folded into the consumer Alexa brand and quietly removed from the product nav.

The rebrand matters less than the architecture: Amazon is consolidating consumer AI under a single Alexa+ brand and pushing it into every shopping touchpoint. That is the same play Google ran by collapsing Bard, Search Generative Experience, and Assistant into Gemini. Two of the three biggest consumer AI surfaces (Google search and Amazon shopping) now run under a single, unified, agentic AI brand. Apple is the one that has not done this yet, and that is the gap WWDC will have to close.

5. Apple Started Rewriting the App Store for Agents

Reuters and Bloomberg both reported this week that Apple is drafting App Store guidelines for autonomous AI agents, including third-party agents that act inside other apps. The framing is security and privacy first: agents will need explicit permission grants per app and per data class, and Apple is reportedly building an agent-action receipt system so users can audit what an agent did on their behalf.

This is the WWDC pre-leak. Apple does not draft App Store policy in May for fun. The architecture has to ship in iOS 27 to be useful, and that means it lands at WWDC on June 8. The relevant question is not whether Apple has an agent story; it is whether Apple has a model good enough to run the agent layer Apple is now writing the rules for. Our working assumption is still that Siri becomes the router and a third-party model (Anthropic, OpenAI, or both) does the heavy lifting behind it. Three weeks until we know.

6. The Snap-Perplexity $400M Deal Collapsed

Snap confirmed this week that its previously announced $400 million distribution and search partnership with Perplexity has been terminated before broad rollout. Both sides framed it as a quiet wind-down. The market read is that Perplexity could not commit enough exclusive feature surface to justify Snap's payment schedule, and Snap could not afford to keep paying without a clear search-share story.

This is the second high-profile AI distribution deal to fall apart in 2026. The pattern is consistent. Consumer platforms want default-search economics from AI partners; AI companies want optionality across distribution. The middle ground (paid distribution with non-exclusive surface) is not pricing right for either side. Expect more of these unwinds before the model gets fixed.

7. CAISI Pre-Launch Testing Stays the Story Underneath

One follow-up from last week's roundup worth flagging. The Center for AI Standards and Innovation pre-launch evaluation agreements with Microsoft, Google, and xAI signed last week are now visibly shaping release timing. Two frontier labs we talk to have quietly confirmed that they are building two-week CAISI review cycles into their launch calendars. That is the new baseline. Anyone planning a pre-I/O surprise drop has to clear it through Commerce first, and those windows do not move for marketing calendars.

The chart to watch: whether Gemini 4 ships at I/O on Tuesday or whether Google announces it for a date two to four weeks out. A slipped ship date would be the first visible proof that pre-launch evaluation is changing the cadence. A same-day ship would mean the review window started a couple of weeks ago and we just did not see it.

Our Take

The story of the week is consolidation under brand. Google folded its agent story into Android and Gemini. Amazon folded shopping AI into Alexa+. Apple is preparing to fold agents into the App Store. xAI is folding eight model SKUs into one. Anthropic is folding its valuation into a number that signals it is a trillion-dollar candidate. The fragmented era (separate brands for each AI surface, separate model SKUs for each task) is closing fast. The big platforms have decided the consumer cannot keep track of more than one AI per company.

Practical implications. One, if you are building on xAI, you have hours, not days, to finish the grok-4.3 migration. Two, if you are an Android developer, the Gemini Intelligence APIs are worth reading this weekend, because the questions at the I/O developer keynote on Tuesday will assume you have. Three, if you are pricing an AI distribution deal, the Snap-Perplexity unwind is the new precedent: write the deal assuming non-exclusive, or do not write it at all.

Tuesday is the hinge. If Gemini 4 lands with the 2 million token context window, native omnimodal handling, and the agentic harness Google has been telegraphing, the rest of the industry has to respond inside two weeks: Apple at WWDC, OpenAI on a likely GPT-5.6 drop, Anthropic with whatever Claude line update they have queued. If Gemini 4 underdelivers, the trillion-dollar Anthropic valuation talk gets a lot louder, and the policy and platform stories from this week become the durable storylines of the month instead.

See you Monday after the keynote.