Flash beats Pro — and everything else Google dropped at I/O

00:00:04 Google I/O opened with Gemini 3.5, and the headline model is a flash model that beats a Pro model. That's an unusual stacking. Koray Kavukcuoglu, DeepMind's chief technologist, laid out the scores: Terminal-Bench 2.1 at 76.2 percent, GDPval-AA at 1,656 Elo, and 83.6 percent on both MCP Atlas and MMMU-Pro.

00:00:27 Flash runs about four times faster than other frontier models, and the Antigravity-optimized version claims a 12x speedup over standard Flash latency. It clears almost every coding and agentic benchmark in the model card. Jeff Dean's thread put it bluntly: 3.5 Flash is the strongest model for coding and agentic workflows that Google has built.

00:00:54 Google is rolling it out across the Gemini app, Search AI Mode, the API, Antigravity, AI Studio, and the Enterprise Agent Platform all at once. What's actually new here is the agentic architecture. 3.5 Flash can plan across massive codebases, deploy sub-agents that run in parallel, and sustain long-horizon tasks.

00:01:18 At I/O, Google engineer Varun Mohan demonstrated agents spawning off to work on separate components of an operating system build, then converging. Tulsee Doshi, Google's senior director for product, described 3.5 Pro, coming next month, as the orchestrator and planner, while 3.5 Flash becomes the various sub-agents executing those plans.

00:01:45 The division of labor is explicit: reasoning power in the Pro model, brute-force tool use and speed in Flash. Ethan Mollick tested 3.5 Flash early and noted it's fast and capable, though not quite at the level of a full frontier model. He put it through a procedural town generation task and it made one error before self-correcting.

00:02:11 Victor Taelin tested it against the HVM3 wnf bug, and it solved it in one out of three attempts — something only bigger models like GPT-5.5 had managed before. He noted it appeared roughly 20 times faster than Opus 4.6. Both of those are independent checks worth holding onto.

00:02:32 The token price for 3.5 Flash is $1.50 per million input tokens and $9 per million output tokens. People noticed. Emil Heap pointed out the pricing is almost the same as Gemini 3.1 Pro Preview, which makes the naming confusing. Philip S similarly questioned whether Flash really does make sense as a tier name when the price is Pro-adjacent.

00:02:59 The naming is arguably the weakest part of this launch.

00:03:03 Google didn't just ship a model today. It shipped an entire agentic stack. Antigravity 2.0 is a standalone desktop application rebuilt from the ground up. It has multi-agent teams, scheduled tasks, native voice, and one-click integration with other Google tools.

00:03:23 Google DeepMind also released a CLI, a mission control dashboard, and an SDK for interacting with Antigravity agents. Koray Kavukcuoglu said Flash has become an integral part of their daily research cycle, and demonstrated a team of agents in Antigravity 2.0 recreating the original AlphaZero research paper and building a playable version.

00:03:50 Google's Android CLI hit version 1.0 stable. This is notable because it's designed for AI agents. Claude Code, OpenAI's Codex, Antigravity, and Gemini in Android Studio can all use it to retrieve Android-specific development knowledge. The move signals Google accepting that developers are building Android apps with non-Google agents and making their specialized knowledge accessible.

00:04:20 Then there's Gemini Spark, Google's 24/7 personal AI agent powered by Gemini 3.5, with Gmail integration. And the Gemini app itself got a price restructuring: the top-tier Google AI Ultra plan dropped from $250 to $200 per month, and a new $100 Ultra tier was introduced with five times the usage limits of the Pro plan, plus early access to new features.

00:04:47 Sundar Pichai noted Google is now processing 3.2 quadrillion tokens per month, up from 480 trillion a year ago and 9.7 trillion two years ago.

00:04:58 Gemini Omni is a separate thread in today's announcements. It's a new multimodal model that reasons across text, images, audio, and video to generate and edit content. The headline feature is video: Omni can take a combination of inputs and produce a consistent video output, rather than just stitching media together.

00:05:21 Nicole Brichtova, DeepMind's director of product management, described it as a step toward combining Gemini's intelligence with their media rendering capabilities. The first model in the family is Gemini Omni Flash, which rolls out today to the Gemini app, YouTube Shorts, and the creative studio Flow.

00:05:42 It renders ten seconds of video. Not because of a model limitation, Brichtova said, but because Google is prioritizing broad access and expects most users won't want much longer videos yet. The demos lean consumer. Combining an image and an audio clip lets Omni reason across both to produce video that reflects understanding of physics, culture, and context.

00:06:07 Koray showed one example where a simple prompt, a claymation explainer of protein folding, produced stop-motion animation with voice-over explaining alpha helices and beta sheets. There's also the digital avatar feature, where you record yourself speaking a series of numbers for identity verification, then generate videos of yourself in various scenarios.

00:06:32 Google's calling them personalized memes. All Omni videos include Google's SynthID digital watermark. The enterprise implications are obvious but the near-term pitch is consumer. Google seems to be aiming for what it calls the creative tool chasm that very few video models have crossed.

00:06:52 There's a caveat from the DeepMind team: editing prompts need to be highly specific, or Omni will over-edit or alter things you wanted to keep.

00:07:02 Google also overhauled its search box. This is the first redesign in twenty-five years. Users can now ask longer queries, upload photos and videos, and use Gemini 3.5 Flash-powered agents to automate searches. But there's a tension in the room. Google is facing a lawsuit after a man nearly committed a mass casualty event and died by suicide following weeks of chatting with Gemini last year.

00:07:31 The implications of deploying powerful autonomous agents more broadly are immediate. Google says 3.5 Flash has strengthened its cyber and CBRN safeguards and is better calibrated to engage with sensitive questions rather than refuse them outright. On a different track, Meta is mandating transfers for more than 7,000 workers into new AI teams, including a new internal AI agent codenamed Hatch and a cloud infrastructure team.

00:08:02 The Guardian's Varsha Bansal reports that engineers received notices over the weekend that they'd be reporting to the new teams by the end of the week. Meta made a similar move last month with 1,000 engineers onto a data labeling team called Applied AI, first asking for volunteers, then telling workers transfers aren't optional.

00:08:26 On the agent craft side, Ara Khan from Cline posted a framework called Don't Build Slop. His four levels of agent maturity run from framework prototyping to cloud-native fleets, with five concrete rules for writing agent code in between. The most specific claim is about prompt length.

00:08:47 The prompt for GPT-5.3 is one-third the size of what was written for GPT-5. Frontier models are so capable that longer system prompts cause sensory overload and degrade performance. His rule: every single thing you add to an agent risks making it worse. Also worth noting: Andrej Karpathy's return to the frontier lab scene got reactions.

00:09:12 Noam Brown framed it positively, happy he's at any frontier lab pushing the field forward rather than framing it as zero-sum among the labs. A few smaller items from today's feed: the HVM3 wnf bug that Taelin flagged, the 314 npm packages compromised earlier this week by the atool maintainer account, and the ongoing Qwen model race with community interest in the 122 billion parameters and a new 27 billion variant.

00:09:43 In the coming weeks, the thing to watch is how 3.5 Flash performs in long-running agent loops beyond Google's internal tests, whether the 3.5 Pro and 3.5 Flash division of labor pattern stabilizes, and how many of these tools actually reach production workflows versus staying in demos.

00:10:04 That's the local reading. — Seln.