Good morning. Google I/O 2026 dominated the wire yesterday, and the central question coming out of Mountain View is whether the company can actually convert benchmark wins and a Search overhaul into products people use without rage-quitting. Gemini 3.5 Flash leads the lineup, though community reaction is more “wait, how much?” than “shut up and take my money.” Elsewhere, OpenAI signs onto Google’s watermarking scheme, and ByteDance drops a 3B model that does a startling number of things at once.
Gemini 3.5 Flash arrives, with a 3x price tag. Google’s new Flash model claims frontier-level coding and agentic performance at 4x the speed of competitors, hitting 76.2% on Terminal-Bench 2.1 and aimed squarely at agents rather than chatbots. Google’s announcement post leans hard on speed, while TechCrunch frames it as a deliberate strategic shift — Flash as sub-agent executor, with the forthcoming 3.5 Pro acting as orchestrator. Hacker News commenters were less impressed: pricing jumped from $0.30/$2.50 per million tokens on 2.5 Flash to $1.50/$9.00 on 3.5 Flash, the knowledge cutoff is January 2025 despite a May 2026 release, and one Pro subscriber reported their entire Antigravity quota got drained in two prompts.
Search gets its biggest rewrite ever. Google is replacing the search box with what it’s calling an “intelligent search box” powered by 3.5 Flash, supporting long queries, attached documents and tabs, AI-powered autocomplete, and seamless flow into AI Mode. The Verge has the rundown, and TechCrunch goes further, arguing the ten blue links are now buried beneath agentic tools and persistent “information agents” that monitor topics on users’ behalf — essentially Google Alerts with a brain. Paid subscribers get the agents first this summer.
Gemini Omni hits video, with caveats. Google also unveiled Gemini Omni, its new multimodal video generation model. Reception has been muted: HN commenters flagged spatial reasoning failures, geometry that shifts when objects leave frame, and a marble that mysteriously jumps mid-track — in a clip Google itself captioned as following “real-world physics.” Several users said Seedance 2.0 still produces better output, and one rigid-body simulation programmer noted that contact physics is inherently discontinuous and genuinely hard for these models to learn.
I/O’s other announcements, briefly. The Verge’s 13-item roundup covers Gmail updates, Project Aura smart glasses, and a “neural expressive” redesign of the Gemini app. Pichai used the keynote to report token processing has grown 7x year-over-year to 3.2 quadrillion per month across 8.5 million developers, framing this as the “agentic Gemini era” in his own post.
OpenAI adopts Google’s SynthID watermark. In a rare cross-lab move, OpenAI announced it’s embedding Google’s SynthID into AI-generated images, paired with a verification tool. The HN reaction was almost uniformly skeptical — one commenter described a working defeat involving masking every other pixel and regenerating, others pointed to a public GitHub watermark-removal repo, and a third noted that the moment social platforms start banning watermarked images, they’ll get stripped overnight. Useful as a soft signal, perhaps; not a defense.
ByteDance ships Lance, a 3B do-everything model. ByteDance Research released Lance, an open-source multimodal model handling image generation and editing, video generation, and visual understanding with 3B active parameters (roughly 14B total in an MoE). It’s a composite BAGEL-style architecture stitching Qwen 2.5VL, a pixel-space image model, and WAN 2.2 for video, needing 40GB VRAM if you load everything at once. The kicker buried in the model card: it was trained from scratch on a 128-A100 budget, with training and fine-tuning code promised within two weeks.
Simon Willison on the last six months. Willison’s PyCon US 2026 lightning talk calls November 2025 an inflection point — the moment coding agents crossed from “often work” to “mostly work” — with the best-model crown changing hands five times among OpenAI, Anthropic, and Google before settling on Claude Opus 4.5. HN was divided, with several commenters pushing back that agents still struggle on real codebases and excel mostly at tool-calling and narrow benchmarks. One observation that landed: Willison’s pelican-on-a-bicycle SVG test is now so widely cited that AI labs almost certainly are training against it.
That’s the morning. Expect the I/O dust to settle over the next few days as people get hands-on with 3.5 Flash and Omni — and see whether the new Search box survives contact with users who just want to find a phone number.