You know, I sat down last night to write this, thought to myself "let's dive deep into something," and didn't have to look far.

I save the important tweets in bookmarks and Gemini 3.8 Live was the latest, thought I'd cover it. But I realized way too much has happened this week, AGAIN.

And it's genuinely tough to keep up.

So yes, another roundup edition today, except I did try Gemini 3.8 Live myself. Let's dive in.

GPT-Live-1 is live, and it's not cheap.

OpenAI's full-duplex voice model hit the API and became the default in ChatGPT Voice. It listens and talks at the same time instead of waiting for you to finish.

Gemini 3.8 Live does the same job at roughly $1.38/hour.

Claude stopped asking which mode you want.

If you've been switching between the two with no clear rule, same, and this fixes it.

Your bank account is the next battleground.

OpenAI launched ChatGPT for Financial Services for bankers. Four days later Claude Money showed up in Anthropic's iOS app, connecting to your bank accounts.

OpenAI's ahead on consumer, its Plaid-linked tools have been live since May.

Meta wants a subscription.

Meta One bundles Instagram, Facebook, WhatsApp and Meta AI from $2.99/month, ₹79 in India.

Two days later its Muse agent got a Mac app, and both Muse and rival Instinct learned to make phone calls the same day.

ElevenLabs skipped the model fight and sold the product.

While Google and OpenAI shipped models, ElevenLabs shipped Reception, an AI receptionist for small businesses.

Everything else worth knowing, in a line each.

Someone finally scored the AI assistants you text, on daily tasks. 116 assistants, 215 tasks, open source. Muse currently leads at 9.3, Grok Bot at 7.3.

A frontier model that doesn't write text. You feed it a state and a typed question, it returns a decision in one pass. Claims 40 to 200x an LLM's speed.

More on this on Monday.

You drop in a few photos and the occasion, and about a minute later it writes a 30-second song about what's in the pictures and sings it in your language. English, Hindi, Hinglish, Arabic, Spanish, Turkish.

Google shipped two models: Gemini 3.8 Live, the fast one for normal conversation. Gemini 3.8 Live Extended Thinking, the same model with a reasoning budget you can turn up.

I set up AI Studio's playground, picked the Zephyr voice, and gave it a problem.

Google says 3.8 Live handles 97 languages and can switch between them mid-conversation.

I wanted to break that claim because switching languages usually means restarting the whole context. I started in Hindi, asked it for ways to grow a small D2C brand.

It gave me three directions and asked which one to go deeper on. I picked automation and got specific.

It worked out I meant cart abandonment on its own, then walked me through reminders, personalised offers, a cleaner checkout.

Then mid-sentence I switched to Kannada.

Instead of resetting, it carried the same problem forward into personalisation, returns, retargeting and promo experiments, just in Kannada now.

The idea survived the language switch.

Most of us don't think in one language.

Every AI until now made you translate that thought into whatever language you were prompting in. 3.8 Live lets the languages live inside one conversation without dropping the thread.

That's a more useful version of "multilingual."

If you only try one thing this week, try this: ask it something that needs a lookup and listen for whether it keeps talking while the lookup runs or goes quiet.

That background tool-call while speaking is the whole reason to leave the old pipeline behind, and no benchmark can show it to you.

I'll admit I didn't expect to arrive at this conclusion but here we are. This is the week voice clicked for me, mainly because the language-switch worked, and in a way I'd use.

But I still catch myself typing instead of talking, and I don't think that's the model's fault anymore. It's the wrapper around it.

The model got good, while the experience of using it hasn't caught up.

I keep thinking about Instinct and Openclaw. They're fundamentally the same but Instinct has more adoption because it just felt better to use.

Same lesson, over and over: the model is table stakes now, and the winner is whoever wraps it best.

Which is the whole reason Reception stuck with me this week.

Google and OpenAI shipped the better engine while ElevenLabs shipped what a dentist can turn on by pasting a URL, and sold it for the price of a phone plan.

I'm not sure the labs have fully clocked it, maybe they have and they're happy being the layer underneath. I don't have a clean answer here.

So try it yourself this week: talk to Gemini 3.8 Live, then ask whether it was the model that won you over or the way it felt to use.

Reply and tell me. I think your answers will say more about where this is going than any leaderboard does.

Until next time,
Vaibhav 🤝🏻

If you read till here, you might find this interesting

#AD 1

A PM can summarize research, draft a PRD, and mock up a prototype before lunch. The hard part starts when the team has to decide what actually gets built.

Jira Product Discovery gives product teams one place to capture insights, prioritize ideas with consistent frameworks, and build living roadmaps stakeholders can rally around.

And because it’s connected to Jira, the context behind every decision stays with the work—so developers and their agents know not just what to build, but why.

AI helps PMs move faster. Jira Product Discovery helps the whole team build with confidence.

Get Started for Free.

#AD 2

Be the smartest person in the room. 1440 navigates 100+ sources to deliver a comprehensive, unbiased news roundup — politics, business, culture, and more — in a quick, 5-minute read. Completely free, completely factual.

Sign up now!