Google's Big AI Test Comes Next Week

May 15, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
May 15, 2026

HEADLINES

Cerebras Delivers a Monster IPO Debut

Cerebras delivered a massive first day of trading, kicking off what looks like a significant IPO season for AI companies. The chipmaker upsized its share offering, raised its price, and still priced above the guided range — red hot demand flowed straight from the private roadshow into public markets, with the opening trade briefly doubling the stock before it settled at a 68% gain. The company went from a $40B valuation at open to briefly touching $100B, closing at $66B.

Predictably, the frothy debut drew contrarians. Jim Cramer urged caution and General Intelligence's Andrew Piccinelli called it a potential market top. But the bull case is simple: CNBC reported 45 buyers for every seller, and with SpaceX paperwork expected next week and Anthropic and OpenAI lining up for year-end IPOs, this sets an interesting floor for what comes next.

Figma and NVIDIA: The Market Is Back on the AI Hype Train

Figma is the latest software company to come back from the dead on AI revenue. After seeing its stock down as much as 50% this year, Figma reported 46% revenue growth in Q1 — accelerating from 40% the prior quarter — and credited its AI features. CFO Pravi Melwani stated you "can't dismiss" the significance of new tools, and the company said 75% of customers are still using AI features after it introduced usage caps in March. The stock jumped 8% in after-hours trading.

Meanwhile, NVIDIA is quietly having a major surge — up 20% over the past seven trading days, pushing the world's largest company close to a $6 trillion valuation. Thursday's session alone added 4.7%. NVIDIA and Micron together have accounted for more than 30% of this year's S&P gains.

OpenAI vs. Apple: A Partnership Gets Rocky

OpenAI is reportedly considering legal action against Apple for breach of contract related to the ChatGPT integration that was announced at WWDC 2024. The deal was always a bit of an afterthought — Sam Altman was at the event but not on stage, and the integration was never core to Apple Intelligence, just a handoff for complex Siri requests. That product didn't exactly age well, and apparently neither did the relationship.

Sources told both The Information and Bloomberg that OpenAI executives feel Apple buried the feature and limited what ChatGPT could do on iPhones. One unnamed OpenAI exec told Bloomberg that Apple said "take a leap of faith and trust us" before the contract was signed, and that the promised boost to user numbers never arrived. OpenAI has now retained outside legal counsel. Separately, The Information reports Apple has largely switched to Claude internally for coding, and is testing native Claude and Gemini integrations with the same system access as ChatGPT — which, to be clear, OpenAI says is unrelated to any potential lawsuit.

Anthropic's $900B Round; Microsoft Cuts Claude Code

Anthropic appears to be setting an IPO price floor with one last venture round. The Financial Times reports terms have been agreed for a $30B raise at a $900B valuation — edging past OpenAI's last mark of $852B and nearly tripling from the $380B Series G in February. Sequoia, Altimeter, Dragoneer, and Greenoaks are co-leading with $2B+ each. Meanwhile, Microsoft is heading in the other direction, canceling Claude Code licenses for its developers at the end of June and shifting them to GitHub Copilot CLI. Claude Code was apparently very popular — perhaps too popular, with cost cited as a factor alongside Microsoft's need to push its own competing tool.

Mythos Finds a MacOS Vulnerability

Security researchers have used Claude Mythos to discover and exploit a vulnerability in Apple's macOS, linking together a pair of bugs to gain access to kernel memory — a significant result given macOS's reputation as one of the more hardened operating systems available. Researchers wrote that Mythos is "powerful: once it has learned how to attack a class of problems, it generalizes to nearly any problem in that class." This follows Mozilla's announcement last week that Mythos helped find and patch 423 bugs in a month — more than the previous 15 months combined. Anthropic has also released an updated Mythos checkpoint with significantly boosted cybersecurity capabilities; the UK AISI found it completed their automated cyberattack benchmark 6 out of 10 times, versus 2 out of 10 for the prior version and 1 for GPT-5.5.

MAIN STORY

Google’s Big AI Test Comes Next Week

Google I/O next week is shaping up to be a pivotal moment — not because Google is expected to drop a state-of-the-art model, but because of what it reveals about how they're choosing to compete. The core question: will Google pick a lane between consumer AI and work AI, or continue trying to do both? Early signals suggest the latter. That's a harder path, but Google may have more to gain than people realize — if they get a few specific things right.

WHAT GOOGLE NEEDS TO DO AT I/O

Codex in the ChatGPT Mobile App
This is the context for everything that follows at I/O.
OpenAI's announcement that Codex is now fully integrated into the ChatGPT mobile app is more than a feature update — it's a continuation of the shift from "doing the thing" to "managing AI agents that do the thing." This isn't just remote control; it's a full-fledged experience where you can initiate new work, review outputs, steer execution, and approve next steps entirely from your phone. The laptop becomes a satellite device. Lapo Chirici's framing cuts to it: the mobile interface isn't a convenience feature, it's an admission that the bottleneck is now human review cycles, not generation speed. Your job is triage, not execution. That shift in modality is what makes the harness question so important for Google.

Work AI vs. Consumer AI: The Divergence
AI is a normal technology for consumers, an abnormal one for work.
The Codex story sets up the broader theme: work AI and consumer AI are diverging fast — in how the technology works, what users want, and how much they're willing to pay. For work users, we can't get model and harness updates fast enough, can't get token access fast enough. For consumers, AI is impressive but growing more normally, and much of the pushback comes from people having AI shoved into places they don't want it. OpenAI chose a lane months ago (work). Anthropic was always there. Apple and Meta are on the consumer side. Google is the one major player still trying to serve both — and I/O next week is going to tell us a lot about whether that's a strategy or a liability.

Gemini Spark: Google's Consumer Play
The same promise Google has made for eight years — hope it works this time.
A screenshot of a Gemini Spark welcome screen has been circulating, describing it as an always-on AI agent ready to help with your inbox, online tasks, and more — learning from connected apps, chats, tasks, and personal intelligence over time. The consumer bull case is real: Google has 20 years of context about its users that no one else has, and Spark sounds like it's finally trying to capitalize on that. The bear case is equally real: Peter Gostev from Arena AI noted that "I feel like I've seen that line from Google for about eight years with the product name changed once in a while." The question isn't whether Google can build it — it's whether they actually will this time.

Gemini 3.2 Flash: Google's Work AI Opportunity
15-20x cheaper than GPT-5.5 at comparable performance could be a very big deal.
Early reports suggest new Gemini models are coming but won't be state-of-the-art — somewhere between Claude 4.7 and GPT-5.5, and even that might be optimistic. That sounds disappointing until you look at the Flash story. Rumors suggest Gemini 3.2 Flash (possibly renamed 3.5) is hitting 92% of GPT-5.5's coding and reasoning performance while being 15-20x cheaper on inference, with sub-200ms latency. There are a lot of companies right now trying to decide if they're comfortable running Chinese open-source models for cost reasons. If Google can offer Opus 4.5/4.6 class performance at a fifth of the price, that is a genuine lane — and they should triple down on it.

The Harness Question: Where Is Google's Codex?
Gemini CLI, AI Studio, Jules — Google needs to pick one.
Even if Google lands the model story, there's still the harness problem. Ethan Mollick asked this week when Gemini is going to join the Cowork and Codex race to build a local app that isn't just for developers. Antigravity, Google's answer to that, hasn't posted updates in a month. Meanwhile OpenAI ships every Thursday and Anthropic has Claude Code. If Google can leave I/O with a clear, consolidated agentic harness — not four competing products — that would be as significant as any model announcement.