Google's plan to catch up in AI coding

Welcome to Runtime! Today: Google unveils new AI models and developer tools as it seeks to get in on the agentic coding race, believe it or not, another supply chain attack targeting npm packages just dropped, and the latest funding rounds on enterprise tech.

Please forward this email to a friend or colleague! If it was forwarded to you, sign up here to get Runtime each week, and if you value independent enterprise tech journalism, click the button below and become a Runtime supporter today.

Flash dance

After fumbling around for a few months following the late 2022 launch of ChatGPT, which caught it off guard despite having invented much of the basic technology underpinning that service, few companies have benefitted from the generative AI boom as Google. Revenue from Google Cloud nearly tripled from the first quarter of 2023 to earlier this year, but over the last six months developers went crazy for AI coding tools from Anthropic and OpenAI, and that space could be the most promising opportunity for this technology in the long run.

Google introduced the latest edition of its Gemini models Tuesday at Google I/O, and upgraded several developer tools with what the company described as "a major leap forward in building more capable, intelligent agents" in a blog post. Gemini 3.5 Flash is available right away, while a more powerful version called Gemini 3.5 Pro will arrive in June.

• Google compared Flash 3.5 quite favorably to Anthropic and OpenAI's state of the art models, although it's not like they would have included any other type of comparison in the keynote.
• But the company focused on "output tokens per second" as a key metric for Flash 3.5, claiming it is "4 times faster than other frontier models" while downplaying the fact that it costs much more than its predecessor, Gemini 3.1.
• That speed allows the model to process longer-running tasks — such as orchestrating several different agents — much more quickly than other models, reducing the number of tokens needed to deliver the same results.
• And that could save tokenmaxxing enterprises a lot of money: Pichai told attendees that some of Google Cloud's largest AI customers could save up to $1 billion a year by using Flash 3.5 alongside other frontier models, but those customers should expect their mileage may vary.

A month ago at Google Cloud Next, a Bloomberg story detailing how internal politics were responsible for Google's lack of progress on AI coding tools was the talk of the cocktail party circuit that week in Las Vegas. The company bristled at that suggestion, but introduced several new services Tuesday that suggest it took some of those concerns to heart.

• The Antigravity development environment is now available directly to Google Cloud customers through the Gemini Enterprise Agent Platform, which allows those customers to use the same compliance and billing tools with Antigravity that they do for the rest of their Google Cloud workloads.
• Antigravity, originally designed as a standalone IDE by the former Windsurf team acqui-hired by Google in July 2025, also now comes with a command-line interface that allows developers to build agents without having to download a desktop app.
• For smaller teams or individual developers, Google's Gemini AI plans are now significantly cheaper at a time when other AI labs are increasing prices.
• And Google expanded the number of people allowed to try out its CodeMender AI security agent, which is a rival of sorts to Anthropic's Mythos Preview, to help find and plug software vulnerabilities in their code.

Google spent a lot of time at Cloud Next emphasizing its "full stack" approach to enterprise AI, as one of the few companies developing frontier models on its own cloud infrastructure services powered by custom AI chips. But developer tools are part of that stack, and courting the same developers who nearly knocked Anthropic off the internet in a rush to use Claude Code will go a long way toward fulfilling that promise.

• As Google tested Antigravity 2.0 with Flash 3.5, it was able to triple the amount of tokens it processed every day in just two months, according to Pichai.
• If it can deliver that kind of performance at prices that Anthropic and OpenAI will find hard to match, developers might as well give it a try.

As the worm turns

It's getting a little hard to understand how open-source developers can continue to rely on package managers like npm and PyPI when they are under near-constant attack from malicious publishers. The same team behind the wave of "Mini Shai-Hulud" attacks struck again Monday night, this time targeting developers working with the open-source AntV data-visualization tool.

Developers use software packages to add open-source components to their own applications, but carefully crafted malicious packages can add a lot more if they are downloaded and installed before they are detected. "Once installed, the payload scrapes plaintext secrets from CI/CD runner memory, steals local cloud and SSH credentials, and plants backdoors in VS Code and Claude Code configurations," Aikido Security said in a blog post similar to the ones it has had to write several times over the past few weeks as these attacks have spread. 

In this edition, hundreds of compromised packages flooded npm in just a few minutes, making this attack "one of the larger npm supply chain incidents Socket has investigated recently," the security company said in its own blog post. The attacks work by compromising the accounts of package maintainers who regularly publish updates, and maybe it's time for these services to give those maintainers better security protocols to follow.

Enterprise funding

Decart raised $300 million in new funding to continue developing its AI infrastructure technology, which optimizes AI inference across several different chips from Nvidia, AWS, and Google.

Sigma landed $80 million in Series E funding for its agent analytics platform, which helps companies make sure their agents are behaving as designed.

Viktor scored $75 million in Series A funding for its "virtual coworker" technology, an agent that works inside tools like Slack or Microsoft Teams to help actual employees carry out their work.

Relay raised $50 million in new funding for its financial-services software, which was designed for small businesses.

Dust landed $40 million in Series B funding for its agent-development platform, which was designed to encourage teams to build agents together to solve common problems.

Ocean launched with $28 million in funding for its email security technology, which uses AI agents to detect phishing attempts.

The Runtime roundup

A contractor working for CISA leaked a treasure trove of sensitive agency security credentials on GitHub, which Krebs on Security said "represents one of the most egregious government data leaks in recent history."

Google and Blackstone will invest $5 billion to create an AI neocloud running Google's TPU chips in standalone data centers.

Thanks for reading — see you Thursday!