Google's July drop: a new robotics brain, three faster Gemini models, and an AI code optimiser
Gemini Robotics ER 2 gives machines a physical-world brain, while three new Flash models chase cheaper, quicker AI agents and AlphaEvolve opens to all Cloud customers.
Key points
- Google launched Gemini Robotics ER 2, an "embodied reasoning" model that lets robots understand their surroundings and work through multi-step physical tasks.
- Three Gemini models arrived in July 2026: 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber, tuned for cheaper, quicker AI agents.
- AlphaEvolve, Google's code-optimising AI, is now generally available to Google Cloud customers on the Gemini Enterprise Agent Platform.
- Gemini Spark can now log into a user's accounts, with permission, to handle web errands like booking apartment viewings.
- Google joined BlackRock, Ford and Carhartt to launch the Alliance for America's Skilled Trades.
Google used its July 2026 roundup to push AI out of the browser and into the physical world. The headline act, for anyone who watches robots for a living, is Gemini Robotics ER 2.
ER 2 stands for "embodied reasoning," version 2: the software brain that lets a robot look around a room, understand what it sees, hold a conversation about it, and plan a sequence of physical actions. We covered the model's launch on 30 July in two separate pieces, one on its dexterity and multi-robot teamwork and one on how it splits high-level planning from low-level motion control.
Google calls it its most capable robotics model yet and pitches it at developers building machines meant to be genuinely useful. What Google's blog post doesn't answer is which partner robots are running ER 2 autonomously versus teleoperated (a human steering quietly from off-camera). That distinction matters more than any benchmark.
What are the three new Gemini models?
They're cheaper, faster versions of Gemini aimed at developers running lots of AI agents at once. An AI agent is software that carries out multi-step tasks on its own, such as booking a trip or filing an expense.
Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber each target a different point on the cost-speed curve. Google says they improve token efficiency (how much text they process per cent), latency and reliability. Those three qualities matter when you're running thousands of agents in production and the bill scales with every word.
| Model | Positioned for |
|---|---|
| Gemini 3.6 Flash | General agent workloads at scale |
| Gemini 3.5 Flash-Lite | Cheapest, fastest tier |
| Gemini 3.5 Flash Cyber | Security-focused tasks |
Why does this matter for people who don't build robots?
The same July drop quietly pushed AI deeper into everyday chores. Google expanded Gemini Spark so it can use logged-in accounts and saved passwords, with permission, to handle web errands: scheduling apartment viewings, researching flights, starting the booking process.
That's a real behavioural shift. Handing over credentials so software can click around the web on your behalf is a fundamentally different relationship than typing a question into a search box. Read the permissions screen carefully before saying yes.
Google also rolled Gemini Notebook (formerly NotebookLM) into the main Gemini app and Search, backed by a secure cloud computer. On Samsung's Galaxy Z Fold8 Ultra and the Flip8, Android 17 now lets iPhone switchers move data across wirelessly without a separate app.
What about the business and creative side?
AlphaEvolve, Google's Gemini-powered code optimiser, is now open to all Google Cloud customers on the Gemini Enterprise Agent Platform. Feed it a working algorithm and a goal; it searches for a faster or cheaper version and returns readable code you can inspect. AI2Day first covered AlphaEvolve on 22 July 2026.
On the creative side, Google Vids gained Gemini Omni for editing clips with everyday language, plus personal avatars you can drop into your own videos. Lyria 3.5, the newest music generation model, powers Google Flow Music with better vocals and lyrics.
Google's decision to anchor an AI roundup with a skilled-trades alliance is the detail worth sitting with. Software fixes fewer problems than the industry likes to admit, and somebody still has to wire the building that runs the data centre.
