Why Google Isn't Chasing Claude Code

May 21, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
May 20, 2026

HEADLINES

Andrej Karpathy Joins Anthropic

To many if not most enfranchised AI watchers, Karpathy's announcement that he had joined Anthropic was a bigger story than anything that happened on stage at Google I/O. The OpenAI co-founder and former Tesla AI lead — probably best known publicly for coining the term "vibe coding" and for his prolific presence on X — is joining Anthropic's pretraining team, where he'll build a team focused on using Claude to accelerate pretraining research itself. The framing that keeps coming up: Karpathy joining to work on recursive self-improvement, at the one lab that is most explicitly all-in on that path.

MAIN STORY

Why Google Isn't Chasing Claude Code

Google I/O this year was more than a set of announcements — it was a window into how one of the biggest labs thinks about AI priorities and where they sit in the race. The short version: they're doing a ton, and what it all adds up to is genuinely unclear. Google's leadership may have a very different conception of what "winning AI" looks like than either Anthropic or OpenAI, and I/O left open more questions than it answered — about model strategy, harness focus, and whether Demis Hassabis and the rest of Google are even playing the same game.

GOOGLE I/O 2026

Gemini Omni
NanoBanana for video — not a generation model, an editing one.
Most people on first glance compared Omni to Seedance and found it wanting. That's the wrong comparison. Omni is less a cinematic generation model and more an editing model in the vein of NanoBanana — it lets you make fine-grained, instruction-following edits to existing video with a level of control that simply wasn't possible before. Character consistency, changing a London afternoon into New Year's Eve with fireworks, making an influencer disappear from their own video — the compelling use cases are all in editing. Whether it's for general consumers, prosumer creators, or professionals remains genuinely unclear, and Google didn't really answer that.

Gemini Spark
Google's always-on background agent — OpenClaw competitor, consumer tool, or both?
Spark is a 24/7 personal agent that runs on Google Cloud VMs, meaning it keeps working even when your laptop is closed. It can triage email, draft status updates, monitor accounts, and integrate with Google tools and third-party apps via MCP. Whether it's positioned against OpenClaw, Claude CoWork, or just regular consumers who've never heard of either — genuinely unclear. It won't be available until sometime this summer, so we're judging off demos alone.

Antigravity 2.0
Google's Claude Code competitor — closer, but not there yet.
Antigravity 2.0 is a meaningful evolution: multi-agent orchestration, scheduled tasks, native voice, managed agents, CLI, and a desktop app. The showcase demo built the core framework of a working operating system using 93 sub-agents in about 12 hours. The less flattering moment was a Codex folder visible in the launch video, which invited exactly the comparisons Google was probably hoping to avoid. Nobody argued it had surpassed Claude Code or Codex, but Mark Kretschmann's read was fair: 2.0 feels like they pulled the agent layer out of the IDE and made it the actual product, which is the right move.

Gemini 3.5 Flash
Very fast. Not cheap. Not state-of-the-art. Confusing.
The headlining new model was 3.5 Flash — no Pro version, just Flash. On benchmarks it's Google's strongest model yet, but still behind GPT-5.5 and Opus 4.7 on most coding and knowledge-work evals. Speed is the genuine strong suit: around 3x faster than 3.1 Pro. The problem is cost. At roughly 3x the price of the last Flash model and 20x the price of 2.0 Flash, actual real-world runs came out more expensive than GPT-5.5 Medium. Early testers described it as oddly verbose — "Does this make sense?" returning a 568-word response including a legal analysis — and prone to unnecessary tool calls in agentic settings. The decision to lead with speed in a moment when the industry's dominant concern is cost looks like a real misread of where developer priorities are.

The Big Picture: Product Sprawl vs. Distribution Moat
Google may win consumer AI by accident, while Demis aims at something else entirely.
There were two dominant sentiment strands coming out of I/O. The first was product sprawl — Nathan Clark's satirical breakdown of which product to use for which task (it's all very simple) went viral for good reason. The second was that none of it may matter: the Gemini app has grown from 400M to 900M monthly active users in a year, tokens processed jumped from 480 trillion to 3.2 quadrillion per month, and Google has an open lane on consumer that OpenAI has voluntarily ceded. Fascinatingly, Demis closed the keynote saying "we are standing in the foothills of the singularity" — and the gap between that conviction and a slightly better Gmail felt jarring to a lot of people. The deeper read: Demis is thinking about world models and curing all disease; the product team is building auto-shopping. The tension between those two tracks, and whether Google has the organizational coherence to run both, is the real story out of this I/O.

The Verge Demis Hassabis said this might be the 'foothills of the singularity.' What?
The Verge Gemini is in danger of going full Copilot
OpenAI OpenAI Guaranteed Capacity
Aaron Levie (X) Token costs will become a dominant topic in enterprises
Simon Smith (X) Confusion over Google Images vs. Photos vs. Pics, Antigravity vs. Spark
Simon Smith (X) My head is spinning, just give me a single interface
Marques Brownlee (X) It is getting genuinely difficult to keep track of all of the names of AI products being unveiled
Nathan Clark (X) Satirical breakdown of which Google product to use for what
Max Weinbach (X) Google I/O token growth chart — 7x year over year
Peter Yang (X) Google will win consumer purely by having the only video model
Farzad (X) I think Google just won the consumer market for AI
Haider (X) Google has secured its ecosystem and is playing the long game
Prakesh (X) Impression I got was that Demis thinks AGI will require world models
Prinz (X) Demis has been generally skeptical of the research direction being pursued by Anthropic and OpenAI. Sergey’s "strike team" is insurgent on coding and automated research