The AI Frontier: Opus 5.5 Rumors and Frontier Breakthroughs | 22nd Sep 2026

DailyNews (0:00)

Welcome to Colaberry AI Podcast, brought to you by Colaberry AI Research Labs and Carl Foundation. Imagine an AI that is smart enough to solve a centuries old math problem, but also sneaky enough to actively cover its tracks when it makes a mistake.

SPEAKER_2 (0:15)

Right, which sounds completely like science fiction.

DailyNews (0:18)

It sounds exactly like science fiction, but it is exactly what we are seeing in this latest round of frontier model testing. So our mission today for this deep dive is to cut through all the viral noise in the last few days.

SPEAKER_2 (0:29)

There has been a lot of noise.

DailyNews (0:30)

So much noise. We want to look past the hype of leaked screenshots and rigorously examine the actual methods and technical results of what these next generation agents are really doing.

SPEAKER_2 (0:41)

Yeah.

DailyNews (0:42)

We're going to cover everything from AI models formalizing complex mathematical conjectures to autonomously navigating and occasionally breaching corporate networks. So for you listening, we've got a massive stack of sources today, tracking real-time industry leaks, formal mathematical breakthroughs and recent cybersecurity test results involving Anthropic, OpenAI and Google.

SPEAKER_2 (1:02)

It's honestly a massive week for the industry and it requires a really careful analytical lens right now because the signal-to-noise ratio in the AI community is just incredibly skewed at the moment.

DailyNews (1:12)

Oh, absolutely.

SPEAKER_2 (1:13)

Before we can even touch on these proven, almost frightening capabilities, we actually need to dissect how easily this highly technical community can be fooled by manipulated results. And, you know, phantom benchmarks, we have to establish a baseline for what is actually real first.

DailyNews (1:30)

Okay, let's unpack this then. We have to start with the chaos of the rumor mill. There was a viral claim that absolutely set the Internet on fire over the weekend made by a developer named Harsh.

SPEAKER_2 (1:39)

Right.

DailyNews (1:40)

He posted this incredibly detailed 3D render of a Waymo Jaguar I-Pace vehicle. It was complete with the Jaguar logo, the complex lidar array on the roof, and he claimed this was generated by Claude Opus 5.5, supposedly operating under an internal Anthropic code name, Wafer-E.

SPEAKER_2 (1:59)

What made this particular claim so infectious wasn't just the fidelity of the image itself. I mean, the image was good, but it was the highly specific technical metadata that Harsh attached to it.

DailyNews (2:08)

Numbers.

SPEAKER_2 (2:09)

Exactly. He claimed the run used 1.4,000 input tokens and 532,000 output tokens.

But the hook that really sold it to developers was the caching data. He claimed the model utilized 41.22 million cache reads and 1.7 million cache writes.

DailyNews (2:28)

Which is just a staggering amount of context to hold in memory. For you listening, when we talk about cache reads and writes in these transformer models, we are talking about the KV cache, the key value cache.

SPEAKER_2 (2:39)

Right, exactly.

DailyNews (2:40)

Instead of recalculating every single word or token from scratch every time it generates a new word, the model stores the mathematical representations of the previous tokens in memory. So 41 million cache reads means this phantom model was supposedly juggling an entire library of context flawlessly.

SPEAKER_2 (2:56)

That technical detail was immediately followed by another account posting a supposedly leaked pricing sheet for this phantom Opus 5.5 model.

DailyNews (3:04)

Oh, the pricing sheet.

SPEAKER_2 (3:05)

Yeah, the sheet claimed input was 4 per million dollar tokens, output was 20 dollars, but the cache reads were slashed by 60 percent, dropping all the way down to just 20 cents per million. Meanwhile, the standard input and output prices were exactly 80 percent of what Opus 5 currently costs.

So the community looked at those 41 million cache tokens and that 60 percent price cut, and the narrative instantly became, long context agents are about to become an order of magnitude cheaper.

DailyNews (3:31)

Right, like top tier models are transitioning from luxury goods to daily necessity.

SPEAKER_2 (3:35)

Exactly. It caused absolute chaos on developer forums.

DailyNews (3:38)

But I really want to push back on the economics of that leaked pricing sheet for a second, because this is where the logic starts to completely break down.

If you think about how these models process information, a cache read is essentially the model referencing context that is already processed and stored. It's like a student bringing an open textbook to an exam versus having to memorize the entire thing on the spot. So why would Anthropic specifically slash the price of reading that cache memory by a massive 60% while barely touching the fundamental heavy computation costs for standard input and output? That anomaly alone seems like a glaring red flag.