🧠 Google's Robot Has A Lightbulb Problem

Gemini Robotics 2 landed yesterday, and it is really three models working together: one that turns instructions into motion, one that plans and reasons, and a lightweight version that runs on the robot itself with no internet.

Basically, you tell an Apptronik Apollo 2 humanoid to put a watering can on the bottom shelf, and it walks over, picks it up, walks to the shelf, and places it. Previous versions handled tabletop work only. This one uses its hands, legs, fingers, and of course its (artificial) intelligence. What we thought was super cool was how its five-fingered hand ties trash bags 44% of the time and handles a ziplock 40% of the time.

The reasoning model is live in Google AI Studio today. The rest is early-access partners only.

Starting July 30, GPT-5.6 Luna costs 80% less and Terra costs 20% less. Luna now sits at $0.20 per million input tokens and $1.20 per million output.

Here is why that matters even if you never touch the API. Those lower prices also change how your usage gets counted inside Codex and ChatGPT Work, so the same subscription now stretches noticeably further.

OpenAI also swapped Priority Processing for a new Fast mode, which runs GPT-5.6 Sol up to 2.5x faster at double the price, with identical output quality. See more here.

Inkling-Small is out, and the full weights are on Hugging Face. It runs 276B total parameters with only 12B active at any moment, which is the trick that keeps it cheap to run.

The part that raised eyebrows: it matches its much larger sibling at a quarter of the size, and beats it on reasoning. Humanity's Last Exam went from 29.7% to 31.6%, and it clears 80% on SWEBench Verified.

It reads images and audio natively, holds a million tokens of context, and lets you dial reasoning effort up or down per task. You can fine-tune it on Tinker or just chat with it in the Tinker Playground. Learn more about this here.

If you have ever shipped a voice agent, you know the demo works and the real caller breaks it. Cekura runs your agent through thousands of simulated calls first, with different accents, personas, interruptions and background noise, then monitors real calls once you go live. Plugs into Vapi, Retell, LiveKit, Pipecat and ElevenLabs. Free tier gives you 300 credits to start.

2. 🔐 Prefactor

Every auth system you have used was designed for a human typing a password. Prefactor is built for agents instead, handling agent login, scoped short-lived credentials, delegated access and MCP support, with an audit trail for all of it. It wraps around your existing provider rather than replacing it. Worth a look if your agents are starting to touch real customer systems.

Claude cannot browse a page continuously on its own. Visualping can, and it now runs as a connector, so Claude can hand the job off and get notified when something moves.

Here’s how you connect the two:

1. Create a free Visualping account, then add https://visualping.io/mcp/sse as a custom connector inside Claude.

2. Tell Claude which page to watch and describe the exact change you care about in plain English.

3. Say how often to check and where the alert should go, email or SMS.

4. Claude sets up the monitor and Visualping runs the checks in the background.

5. Come back anytime and ask Claude to list, pause or delete your monitors.

Great for tracking job openings at a specific company, competitor pricing pages, restocks, ticket drops, or a menu that changes daily.

Here's a quick demo.

Its finally Friday, which means the PAI team is all geared up to go live this morning! 🙌✨

Today’s session shows you how to build a lead generation and outreach system in real time. You'll see exactly how to pull leads into your business and automate the follow-up.

Because we build in real-time, you get to see how we think through the problem, where we get stuck, and what the final outcome is. It's the difference between reading a recipe and watching someone cook it while talking through their decisions.

🗓️ When: Today @ 10AM PT (1PM ET)
 📹️ Where: Virtual