AI News Summary — Week 21.2, 2026

This week: Google goes all-in on agentic AI, a high-profile defection shakes up the AI talent landscape, and the industry tackles AI image attribution. Let's dive in.

Google Unveils Gemini 3.5 and Gemini Spark

Google announced Gemini 3.5 on May 19, positioning it as a major leap forward in building capable, intelligent agents. The release includes Gemini 3.5 Flash, available immediately to billions of users globally.

Key Highlights

Agentic Performance: 3.5 Flash outperforms its predecessor on challenging coding and agentic benchmarks, including Terminal-Bench 2.1 (76.2%) and MCP Atlas (83.6%)
Speed: 4x faster output tokens per second than other frontier models
Cost Efficiency: Often less than half the cost of other frontier models for agentic tasks
Enterprise Adoption: Partners like Shopify, Macquarie Bank, Salesforce, and Databricks are already using 3.5 Flash for automated workflows

Enter Gemini Spark

Perhaps the most intriguing announcement is Gemini Spark, Google's new personal AI agent. Built on 3.5 Flash, Spark runs 24/7 to help users navigate their digital lives — taking action on their behalf while remaining under user direction.

The rollout starts with trusted testers, with Google AI Ultra subscribers in the US getting Beta access next week. This positions Google directly against Apple's Apple Intelligence and Microsoft's Copilot in the personal AI assistant race.

The Bigger Picture

Google is emphasizing that 3.5 Flash is now the default model for the Gemini app and AI Mode in Search globally. The company is clearly betting big on agentic AI — models that don't just answer questions, but execute tasks.

Andrej Karpathy Joins Anthropic

In a major talent move, Anthropic announced it has hired Andrej Karpathy, the renowned AI researcher who co-founded OpenAI before leading AI at Tesla.

Karpathy will join Anthropic's pretraining team, which builds the foundation models that power Claude. This is a significant hire for Anthropic, which has been positioning itself as the "safe" alternative to OpenAI for enterprise customers.

Why It Matters

Karpathy is one of the most respected figures in AI — he taught the famous CS231n Deep Learning course at Stanford, built Tesla's autopilot team from scratch, and was among OpenAI's earliest employees. His move from OpenAI to Anthropic signals that the "safety-first" narrative is resonating with top talent.

This follows Anthropic's $200M partnership with the Gates Foundation announced two weeks ago. The company is on a hiring and partnership offensive.

OpenAI Adds Watermarks to AI Images

OpenAI announced it is joining the C2PA open standard and partnering with Google to embed invisible SynthID watermarks in AI-generated images.

The Implementation

C2PA Metadata: Standardized provenance information embedded in image files
SynthID Watermarks: Invisible watermarks that persist through common edits (cropping, filters, etc.)
Verification Tool: A public tool to verify if an image was generated by OpenAI's tools

The measures currently apply only to DALL-E generated images. This is part of a broader industry push to address concerns about AI-generated deepfakes and misinformation — though critics note that watermarks can potentially be stripped by bad actors.

Pentagon Explores Anthropic Alternatives

Bloomberg reported that the Pentagon is testing OpenAI and Google AI models as it seeks alternatives to Anthropic for defense applications.

The move comes as Anthropic has faced scrutiny over its safety commitments, and the Defense Department appears eager to diversify its AI partnerships. This marks an interesting dynamic in the AI arms race — the biggest tech companies are competing for defense contracts while balancing safety concerns.

What's Hot This Week

Agentic AI is here: Gemini 3.5 Flash is built specifically for agentic workflows, promising speed and cost efficiency
Talent matters: Karpathy's move to Anthropic is a vote of confidence in the safety-focused approach
Attribution arms race: Watermarks are becoming standard, but effectiveness remains debated
Defense diversification: Pentagon testing multiple AI providers beyond Anthropic

That's a wrap for this week. Stay tuned for more AI developments!

Sources: Google Blog, TechCrunch, CNBC, The Next Web, Bloomberg