Claude Next Leak and Enterprise AI Agents | 13th Oct 2026

SPEAKER_1 (0:00)

Welcome to Colaberry AI Podcast, brought to you by Colaberry AI Research Labs and Carl Foundation.

SPEAKER_2 (0:05)

Glad to be here.

SPEAKER_1 (0:07)

So, imagine typing just a simple four line text prompt, right? You walk away from your terminal for a couple of hours, and you come back to find that a machine has coded this perfectly mathematical 30 second animated movie of a human's entire life.

From scratch.

SPEAKER_2 (0:25)

Yeah, completely from scratch. It's wild.

SPEAKER_1 (0:28)

Right? Like, no video generator, no pixel diffusion, just pure raw code executing a visual masterpiece.

SPEAKER_2 (0:36)

And that's exactly why we're talking about this today.

SPEAKER_1 (0:38)

Exactly.

Because we are looking at an engineering landscape right now that is just rapidly becoming autonomous, kind of opaque, and honestly, incredibly fascinating. So, welcome to today's Deep Dive. For you listening, we're glad you're here.

SPEAKER_2 (0:53)

Absolutely.

SPEAKER_1 (0:53)

Our mission today is to bypass all the social media hype. We are analyzing recent leaks and telemetry data around this unreleased Anthropic model. The community has been calling it Claude Next or SODA, alongside Google's totally new Gemini agent architecture.

SPEAKER_2 (1:11)

Yeah, we're really getting into the weeds on this one.

SPEAKER_1 (1:13)

We are.

We're looking at the actual technical mechanics, the API routing methods, and the enterprise deployment results that developers are seeing on the ground today. So spatial reasoning, autonomous coding, multimodal orchestration, the works.

SPEAKER_2 (1:28)

Which is so crucial right now, because the telemetry we're seeing in the wild, it completely challenges our standard assumptions about how these language models parse requests and generate outputs. We're witnessing a fundamental shift in algorithmic capability. It's not just text anymore.

SPEAKER_1 (1:43)

Well, let's start right at that grassroots level. What software engineers are actively seeing in their own IDEs.

SPEAKER_2 (1:49)

Yeah, let's get into it.

SPEAKER_1 (1:50)

Over the last week or so, developers using Claude code, which is Anthropics command line coding tool. They've been selecting the standard Opus 5.5 model for their tasks.

SPEAKER_2 (2:00)

Which is normal.

SPEAKER_1 (2:01)

Totally normal. But the outputs returning to their terminals are showing these behavioral metrics that point to a vastly more powerful unreleased model.

SPEAKER_2 (2:10)

Yeah. And that's the anomaly the community is calling Claude Next, or some internal testers are calling it SODA, which stands for state of the art.

SPEAKER_1 (2:18)

Ah, SODA.

SPEAKER_2 (2:19)

Right.

SPEAKER_1 (2:19)

Okay. Makes sense.

SPEAKER_2 (2:20)

And the primary methodology this anomaly uses to demonstrate its processing power, that's what we really need to pay attention to.

SPEAKER_1 (2:27)

Because it's not generating native video, right?

SPEAKER_2 (2:29)

Exactly. It's bypassing native text and pixel generation entirely. It's producing highly complex visual media, but it's executing it through raw JavaScript and scalable vector graphics, or SVG.

SPEAKER_1 (2:42)

Which is just, it completely upends how we usually think about rendering limits. Like, there's this one tester, goes by Ched Lua, who gave it that four-line prompt I mentioned.

SPEAKER_2 (2:52)

Right. The Lifecycle one.

SPEAKER_1 (2:53)

Yeah. The model used its entire 1 million-token context window and its 128,000-token output limit just to write out a 30-second animated life cycle.

SPEAKER_2 (3:05)

It's staggering.

SPEAKER_1 (3:06)

It is. It rendered an embryo, which then transitioned into a little girl measuring her height against a door frame, and then it concluded with an elderly person sitting in an armchair.

SPEAKER_2 (3:16)

And, you know, for the listener, just picture this, there was no dedicated video generation model involved in this pipeline at all.

SPEAKER_1 (3:23)

Zero.

SPEAKER_2 (3:24)

Just code.

SPEAKER_1 (3:24)

Just code.

SPEAKER_2 (3:25)

And we have to unpack why that approach is such a massive leap in computational logic.

SPEAKER_1 (3:30)

Right. Because usually native models struggle with this stuff.

SPEAKER_2 (3:33)

Oh, immensely. When a dedicated video model generates a scene, it's essentially just hallucinating pixels frame by frame, based on mathematical noise and diffusion.

SPEAKER_1 (3:43)

Which is why things morph and disappear.

SPEAKER_2 (3:45)

Exactly. They lack spatial memory. But by generating an animation using JavaScript and SVG code, this Anthropic model relies on strict deterministic code logic to manage the state.

SPEAKER_1 (3:57)

Meaning it actually knows where things are.

SPEAKER_2 (3:59)

Right. So that recurring armchair in the animation, the model isn't trying to visually remember what a chair looks like.

SPEAKER_1 (4:06)

Oh, I see.

SPEAKER_2 (4:07)

It has literally defined that chair as an algorithmic object within the document object model. It assigned it specific geometric parameters.

SPEAKER_1 (4:16)

And just calls it back.

SPEAKER_2 (4:17)

Exactly. It just calls that object back into the frame.

SPEAKER_1 (4:20)

It's like going to a bakery, right? And ordering a very specific slice of cake. But instead of just handing you the cake, the chef seamlessly builds a fully automated mechanical bakery right in your kitchen.