AI Valley
Posts
Anthropic leaked its most powerful model

Anthropic leaked its most powerful model

PLUS: Language barriers might actually be over

Barsee
30 Mar

_{Sign up}_|_{Follow us on X}_|_Sponsor

Together with

howdy, it’s Barsee again.

happy monday, AI family, and welcome back to AI Valley.

here are the biggest things worth knowing today:

Language barriers might actually be over
Anthropic Claude "Mythos" leak
Plus trending AI tools, posts, and resources

Let’s dive into the Valley of AI…

VOICES

Top brand teams know the secret to standout voice AI. Do you?

Courtesy: Voices.com

The brands making their mark with voice AI treat it like a core identity asset, with the same rigor of logo design and the priority of a product launch. Amplified 2026, Voices’ Annual State of Voice Report, shows exactly how they’re doing it: insights from 700 enterprise leaders and consumers, four defining trends, and the clearest picture yet of where voice AI is headed and what leaders need to do to lock in their brand voice.

Get the playbook your competitors are already using →

_{*This is sponsored}

THROUGH THE VALLEY

1/ Language barriers might actually be over 🎧

Google just expanded Live Translate, its Gemini-powered feature that lets you hear real-time translations through your headphones.

And unlike Apple, it works with basically any headphones.

Courtesy: Google

What’s happening?

Google’s Live Translate can translate speech live into your ears while someone else is talking.

It’s now rolling out on both iPhone and Android across more countries.

Why is this such a big deal?

Because this is one of those AI features that instantly clicks.

You could use it to:

follow conversations with family members
understand train or airport announcements abroad
survive travel without constantly panic-checking your phone

Way more useful than the 900th AI chatbot nobody asked for.

Why is Google’s version more interesting than Apple’s?

Because it feels more human.

Apple’s Live Translation plays everything back in Siri’s voice, no matter who’s speaking.

Google says Gemini preserves the original speaker’s tone, cadence, and emphasis, so it sounds more like the actual person talking.

Which is… a lot less cursed if it works.

So… does it actually work?

Kind of.

It performs well in clear, structured conversations, but still struggles with:

fast speakers
overlapping dialogue
messy real-life conversation

There’s also enough delay to kill the flow a bit, so this still isn’t Star Trek yet.

Why it matters?

We’re not at a perfect universal translator yet.

But we’re getting very close to “good enough to actually use.” And that’s when this goes from demo to daily habit.

2/ Anthropic just leaked its next monster model 👀

Anthropic appears to have accidentally revealed its next flagship AI model after a CMS error exposed thousands of unpublished internal assets.

Buried in the leak: a draft post for a new model called Claude Mythos.

Or… maybe Capybara.

Yes, really.

Courtesy: Anthropic?

What leaked?

A draft blog post describing a new model that would sit above Opus in a completely new tier.

According to the draft, it’s:

larger and more capable than Opus
Anthropic’s most powerful model yet
a major jump in reasoning, coding, and cybersecurity
reportedly presents unprecedented cybersecurity risks

Anthropic later confirmed it’s testing a new model with meaningful advances in those areas.

So this is not just random internet fan fiction.

What’s “Capybara”?

The leak suggests Capybara is the name of a new model tier, built to sit above Opus.

Anthropic reportedly frames it as a move toward more “connective intelligence," models that don’t just solve tasks, but connect ideas across different domains.

And Mythos appears to be the model inside that tier.

So yes, the naming is somehow both ridiculous and very serious.

Why is this a big deal?

The leaked material reportedly describes Mythos as having unprecedented cyber capabilities, to the point that it could help attackers outpace defenders.

That is… a very aggressive warning label.

Why is Anthropic being so cautious?

Anthropic has previously said a Chinese state-sponsored group used Claude Code in cyber operations targeting around 30 organizations.

So a stronger cyber-focused model is… not exactly a chill product launch.

How is it being released?

Very slowly.

Anthropic is reportedly planning a gradual rollout, with early access first going to cybersecurity defenders so they can harden systems before the model, or others like it, become more widely available.

That alone tells you how seriously they’re taking it.

Why it matters?

Anthropic may be building a model that’s not just more powerful, but harder to safely put in people’s hands.

And that’s becoming the real pattern at the frontier.

The models are getting better, but the question is whether they’re getting too dangerous to release normally.

TRENDING TOOLS

PixVerse V6 > Still waiting for Seedance 2.0? V6 matches it — faster, sharper, and more affordable. Ready now ^_(sponsored)
Google Translate > Its Live Translate feature with headphones is now officially live on iPhone, bringing real-time spoken translations to iOS
Bluesky > Just launched Attie, a standalone Claude-powered AI app that lets you build custom social feeds using simple natural-language prompts
Runway Ad Concepter > Turn a single prompt into AI-generated ad concepts in seconds
Cursor > Says Composer 2 is now improving itself every five hours in real time
SUN > Generate on-demand AI audio courses with interactive playback for hands-free learning
Pendium > Helps AI agents recommend you more often by making your product easier for them to understand and surface

WHAT I'M CONSUMING

INDUSTRY MOVES

Brett Adcock, the founder of Figure, just revealed a new AI lab called Hark after 8 months in stealth. So yes, apparently one robotics company was not enough.
Ross Nordeen has reportedly left xAI, meaning Elon Musk’s original founding team is now fully gone. Not exactly the kind of sentence you want floating around while trying to build the future of intelligence.
Anthropic could reportedly go public as soon as Q4. If that happens, one of the most safety-focused labs in AI is about to get introduced to quarterly earnings calls.
OpenAI’s ad pilot has reportedly crossed $100M in annualized revenue. The “there will definitely never be ads” era is looking increasingly decorative.
Midjourney says revenue has now grown past $200M, even as its web traffic keeps falling. Which is a nice reminder that traffic and business are not always the same thing.