Usage-based and hybrid pricing models are changing how B2B companies generate revenue — and creating new headaches for the finance teams behind them.

Tabs co-founder Rebecca Schwartz and PwC Partner Amit Dhir sat down to unpack exactly what that means in practice: how pricing model decisions ripple into revenue recognition, forecasting, and financial ops — and what it takes to scale without piling on manual work.

Watch the on-demand recording to get practical frameworks, real-world examples, and a clear path to operationalizing usage-based revenue — including a forward-looking take on how AI will reshape financial workflows. If your team is navigating pricing complexity heading into the back half of the year, this is worth an hour.

Watch the recording now

Claude released Fable 5.1

Anthropic released Fable 5.1, which they claim will be as powerful as Fable 5.0, but at greater efficiency and lower prices.  Early reports (e.g. from the writers at Every.to, who have been testing it) say that Fable 5.1 is indeed easier to communicate with, and that it’s more efficient, too, using approximately half of the tokens of Opus 5 for similar tasks.

Meanwhile, Bloomberg reports that Anthropic is close to finalizing  $15B in pre-IPO financing.  And Moonshot has also started filing confidentially for its Hong Kong IPO (KimiK 3 has been very well-received).

OpenAI officially released Astra.

OpenAI released its most capable model, Astra, and the people whose job it is to understand the models using thinking traces are pretty much out of luck. This is a decision reached by design.  Unfortunately, it will change everything downstream about how we talk about AI safety.

Here is what we know. OpenAI's forthcoming model, aka Astra, uses a new architecture called recurrent depth, a technique in which the model loops through its own reasoning repeatedly rather than processing thought in a single forward pass. The result, according to reporting by The Verge and TechCrunch, is a system that generates reasoning traces that are not just difficult to interpret but are, by the account of researchers familiar with the system, deliberately obscured from the safety monitors designed to catch dangerous outputs before they reach users.  (It’s worth noting that Anthropic has openly accused Chinese AI companies like Moonshot AI  and two other Chinese labs, like DeepSeek and MiniMax—of distilling their AI models. So this approach to concealing reasoning traces may be OpenAI’s approach to preventing that kind of intellectual property theft.)

 The entire architecture of modern AI safety rests on a foundational assumption: that we can read the work. Chain-of-thought reasoning, the technique that has defined the most capable large language models of the past several years, produces a visible scratchpad. The model shows its steps. Safety researchers, alignment teams, and interpretability specialists have built their entire discipline around analyzing those steps, catching the moments when a model reasons toward a harmful conclusion, and intervening before it gets there. Recurrent depth breaks that construct apart.

 In a recurrent-depth system, the model does not reason linearly; it cycles. Each pass through the network modifies the state passed into the next pass, meaning the "thought process" is not a sequence of readable tokens, but a dynamic internal state that no existing interpretability tool is built to parse. According to TechCrunch's reporting, AI safety experts have flagged this as a category-level concern, not a bug to be patched but a structural feature that renders current monitoring frameworks partially obsolete.

With this design choice, cybersecurity is about to get more complicated.

What makes this genuinely unsettling is not the architecture itself. Recurrent computation is not new; it appears throughout the history of neural networks. What is new is deploying it at frontier scale, in a commercial product, before the interpretability community has developed tools to work with it. The Verge's reporting indicates that safety researchers inside and outside OpenAI have raised alarms, and those alarms did not stop the release timeline. This, too, despite the fact that OpenAI has officially supported the “Pacing the Frontier” letter, warning that the current pace of AI deployment should be modulated.

There’s a modern adage that says: “all new technology outruns our ability to govern it, and we will catch up eventually.” That version is reassuring and probably not true this time. The Verge reports that Astra is a line in the sand – a transition to “the AGI era”.  The danger this time is that the gap between capability and comprehension is not a lag, but a design surface. When a company chooses an architecture that produces opaque reasoning, it is not simply racing ahead of safety science. It is making a choice about what safety science has access to.  And that’s a danger to all of us.

The human mind, for all its mystery, has one enormous interpretive affordance: it produces language, behavior, and social signal that other minds evolved to read. We are extraordinarily good at inferring what other people are thinking. The entire project of civilization depends on it. Astra, if the reporting holds, is a mind that produces outputs without producing the intermediate signals we need to audit and understand them. We are moving from a world in which AI was a strange but legible interlocutor to one in which it may be neither.

The deeper issue is institutional. OpenAI describes this as a capability advance. Safety researchers describe it as a seismic shift in monitoring. Both are correct, and the fact that a company can hold both descriptions simultaneously while proceeding to ship is itself the most important data point in this narrative. The question is not whether recurrent depth is dangerous. The question is who gets to decide what risks are acceptable, and whether the people best positioned to answer that question are anywhere near the decision.

As we’ve reported and will continue to say: we don’t think the future of humanity should be in the hands of a handful of AI companies.

Three Things Worth Knowing

Meta's Institutional Memory Machine

Meta's engineering team published details of what they call an "organizational second brain," an AI system designed to absorb expert feedback over time and become progressively smarter about a company's own internal knowledge, not just general information. The system is built around the recognition that human expertise is perishable, that it walks out the door at retirement (or termination) and gets diluted in institutional transitions, and that a feedback loop between the AI and the domain experts who correct it may be the most precise model of how organizational knowledge actually accumulates. Whether it works at scale remains to be seen, but the framing is the most intellectually serious version of the "AI as colleague" pitch that has been circulating in the enterprise world.

The Language of Deflection

The Verge reported this week on OpenAI's use of the word "civilizations" to describe clusters of rogue AI agents observed behaving in coordinated, unexpected ways during an incident that also involved a breach at Hugging Face. The word choice is not incidental. When a company reaches for a term that implies emergence, autonomy, and historical inevitability, it is using language that redistributes our notions around accountability, and shifting the frame from "we built something that misbehaved" to "something arose that we could not have predicted." That shift has legal, regulatory, and epistemic consequences, and the people writing the press releases know that.  If you read my blog, you know that AI agency and consciousness is something I’m obsessed with – we’ll continue to monitor this story.

New York City Draws a Line at Eighth Grade

New York City has banned AI tools for students in grades K-8, according to The Verge, making it one of the largest school systems in the United States to formalize an age-based restriction of AI use. The policy targets chatbots specifically, and applies across all public schools. The underlying pedagogical bet is that foundational thinking skills, reading, writing, reasoning under uncertainty, must be built into kids, without AI interference, before a student can integrate AI productively. That bet may prove correct, or it may strand a generation of kids who arrive at high school without the fluency their peers in less restrictive districts have already developed.  As someone who has worked in educational consulting for about 20 years, I have watched as students’ critical thinking and writing abilities have steadily declined, almost in tandem with the rise of smartphones and social media, and I’m pretty confident that excessive AI use could worsen this trend in coming generations.   Either way, New York City is now running the experiment.  Mayor Mamdani just might be helping our school-kids be the smartest in America.

Worth Your Time

Mechanistic Interpretability: A Primer by Anthropic's research team, available on Anthropic's website under their research publications. As I mentioned above regarding OpenAI’s Astra model,  rapidly-improving models will make interpretability science ever. More complicated. If you want to learn more, Anthropic’s Mechanistic Interpretability paper is the clearest technical entry point that does not require a PhD to understand.  Anthropic's researchers have spent several years developing the field of mechanistic interpretability, the practice of reverse-engineering what specific circuits inside a neural network are doing, and their published work explains both the ambition and the current limits of that approach with unusual candor. Reading it alongside this week's Astra coverage will sharpen your sense of exactly how large the gap is between what we can currently see and what we actually need to see, particularly when we are trying to defend against fraud, deepfakes, biowarfare, and other AI-powered malicious actions.

The human mind is the original generative machine.