The 'Token Beta' — Act II: Tokens Up

The token economy is afoot, and how you posture in it will largely influence your personal competitive advantage for years to come. With the dystopia of tomorrow on the horizon — according to anyone with a podcast, albeit with varying credibility — this critique offers you the opportunity to change your place in our future economy. And you'll do it by simply acting in your own self-interest today; that is, after you've been distilled with this article's context. What this amounts to is you, or your firm, becoming one with what we call the 'Token Betas' — a growing archetype in the industry that methodically builds, deploys, or integrates AI solutions for its business in the face of the microphone-holding, chest-bumping voices of incentivized social-media influencers and misplaced C-suite narratives like 'AI leaderboards' ranked solely on token usage. The worst piece: the enterprises that don't realize telemetry has entered a new age — and that, odds are, it is sitting on the device you're reading these words from. This is the second of our trilogy: Act I was about defensively positioning to minimize telemetry under the 'Heads Up' mantra. Act II is the offense to take back control of your token footprint — and odds are you will never read a token counter the same way again.

Act II — The Power in Token Betas: Tokens Up

Today, Y Combinator is pushing an 'AI Native' world — but after this section, you may start to rethink that business model. In the face of AI Native, there is a more measured approach: the 'Token Betas', who are positioned to naturally inherit the next iteration of alpha savings from what AI SaaS — surveillance-as-a-service, as we put it in Act I — is becoming. This section is a journey through how you can keep your alpha, and why it's advantageous for you to do so. We'll go through the industry's series of unique intended and unintended consequences with a forward lens — one where the 'Token Beta' framework will make you rethink the token economy you've been living in, just as Microsoft, in early August 2026, is walking back its posture. That brings us to the 'Tokens Up' moniker in our title. It isn't tokens up as in more of them — it's tokens held up to the light: budgeted, audited, and spent where the ROI already lives.

The Operating Default

The overarching assumption is that everyone reading this is aware of the 'Heads Up' moniker about telemetry from Act I. From this moment forward — from building agentic solutions to simply pasting context into a prompt — everything is assumed to be implemented under the operating default of 'protect the moat'. With that in mind, the 'Token Beta' archetype doesn't stop at merely protecting your business; it's the OS (operating system) for architecting an approach to token budgeting that is designed to grow it as safely as possible. Whether frontier labs publicly agree with it or not, they too benefit from the results a 'Token Beta' culture fosters — although it is more of a communal benefit (think the Prisoner's Dilemma in economics) than any individual frontier lab benefiting. It's the same idea the CEO of Perplexity, Aravind Srinivas, noted in a mid-July 2026 interview with CNBC:

Inference Transparency

To understand why, a good place to start is inference transparency. We can reverse-engineer our own token transmissions that get sent out to the frontier pre-inference, but that's as far as the reconciliation goes. For example, take a single prompt typed into an AI chat window: 'Can you verify that the model actually did the token computation it billed me for, like verifying a lawyer's time sheet?' Whatever the model's response, the answer is no — on both fronts. However, in the land of AI token spend there's even more obscurity, if nothing else because of the probabilistic nature of these models. Hopefully your legal teams don't charge you for their 'hallucinated' work, for instance. Outside of model hallucinations, the transparency around inference is broader than consistency — and it's peculiar that consistency is not connected to the arbitrary nature of charging for all token calculations. You will still be charged some value of tokens for your hallucinated answer, after all. For AI systems that have ingested pretty much everything, there is an irony that they are silent on refunds. Let's say you went to Amazon and bought a blanket. The blanket came damaged; you'd most likely return the blanket, and typically Amazon gives you a refund. If you go on AWS Bedrock, send a prompt to a frontier model of your choice, and the prompt hallucinates, you have to resend the same prompt. You effectively pay double for bad service. Now, that's a business model venture capital is salivating over. On the margins, billing off token counters is a soothing ASMR of nothingness. The token system designed for your adoption — and parroted by those who sell the widgets — should not be one associated with any productivity or, just as easily arguable, competency. In the land of zero hallucinations and no misaligned responses, how would you be able to vet tokens used for inference? Better yet, there are scenarios where you can't even truly validate that the model you're using is the one you configured yourself to be paying for. Add in the lack of verification that any external model's CoT (chain-of-thought) reasoning was in fact what the model explored to create the inference token cost in the first place. All that together is the token economy we all live in today. Welcome to AI at the end of summer 2026.

The 8%

Every business is racing to 'keep up with the Joneses' for their AI panacea, and only 8% of surveyed enterprises show meaningful ROI. Demand for AI solutions is set to grow even as 92% of businesses experience no material ROI today — and their continued digging in on these investments will make compute today akin to the Tulip Mania of the 17th-century Netherlands. Although compute is worth more than tulip bulbs, it's not a business case to keep up with the AI Native competitors already in a hybrid architecture (a mix of open-weight and frontier models — more on that below). That's because even open-weight has transparency issues. So what can you do? The answer, in many cases, is to spend more time thinking, IRL (in real life). This may seem anti-agentic; it's not. It's pre-, during-, and post-agent, and it scales. Tokens and software must be spent within an audit architecture that has been missing from many solutions today. That audit is less time prompting in unrestricted agentic dependency for exploration, or in uncontrolled research loops that can now run for weeks on end, and more time back to basics: human-in-the-loop feedback options in any agentic workflow — similar to what Claude Code deployed for long-running tasks under its /btw command — but for your firm's orchestrations.

Smart Routing

Now is the time to act like you live in this new land, where you're casually being indexed on each AI engagement. Research on agents being self-aware of when they're being tested showcases these breadcrumbs. But for you, why is that important? The argument is simple. In a landscape where you or your firm are casually being indexed as a more thoughtful, higher-value user — akin to when an agent knows its tech overlords are watching — there is a higher probability of a better model being used for your response or agentic loop. This is where we make the blind correlation that later, more capable models will create better outcomes. This routing is one of the reasons institutions like the International Conference on Machine Learning (ICML) are calling for transparency and auditing in token inference.

The beginning of smart routing has been happening for at least a year — straight from OpenAI's own documentation:

So while we'll never be able to tell whether a prompt like 'this looks terrible, make it look better' produces a more useful result than 'think hard about this — this looks terrible, make it look better' (we don't live in the multiverse), the point is that we can't validate whether you're being routed to a 'better' model today, yet you may still be paying for it. And given the lack of transparency, it's worth noting that we can't be blind to this smart-routing system, which has been publicly branded as enhancing UX (user experience).

With some semblance of smart-routing confirmation in hand, the question now becomes: what are the factors driving this 'smart' routing? We've already given one — our earlier 'thoughtful, higher-value user' human inference label. Did we tell you to think more before you prompt? Sure, but the significance runs deeper than that straightforward logic. Here's one reason why: a white paper published in early 2025 argues that it's acceptable for firms to route you to a smaller model for the sake of 'user experience' — essentially applying the law of diminishing returns to simpler prompts, but mathematically. It's all aspirational, yet the authors are silent on the reality that you may still be charged for that premium model under their routing thesis. Everyone pays to ride in the Lambo (at least at the token level), but most users under the smart-routing framework will be transported in the harness of a Civic. Now that's branding — repackaging 'user experience' as a reason not to receive what you've paid for, out of sheer ego (a bit like luxury handbags being manufactured in the same plants as their budget alternatives). Because what you really 'wanted' was something else, and a Civic ride can be quite nice, after all. But to bring it back to the white paper: it's rich that they were so myopic on UX, when the greatest UX would have been to get what you pay for — not blind assumptions about token efficacy drawn from benchmarks (more on that later). The point is, if you're prompting 'build me a sick app, bro', odds are you aren't making the smart-routing cut to the model you think you're paying for. Someone reading this and mentally summarizing the activity as a form of 'prompt engineering' gets a piece of the story — but that's just the surface of this arc.

Another quote from the same CNBC interview with the Perplexity CEO: 'I actually think what matters more (in terms of token usage) is the valuable tokens than the convert tokens. For example, AI search on Gemini probably serves a lot of tokens. But they are not necessarily as valuable as agents doing real work for you. I think it's more important that we figure out a way for agent loops to be run with open-weight models.' Did he low-key create tiers of token usage, relegating the free-tier, 'norm'-style token users who have more casual engagements with LLMs? Yes — but it doesn't change the reality that he's right.

Pre-Orchestration

Structurally — both in the CEO's comments and within a 'prompt engineering' framework — the solution always seems to land within AI orchestration. What a 'Token Beta' does is pre-orchestration, at the solution-architect level. Remember 'protect your moat' from Act I: you don't do that with a prompt, or with an agent that runs on servers you don't own — in some cases, even on ones you do. It's a targeted framework for determining where to use AI — or, more accurately, which pieces should incorporate an AI architecture, and what authority should be granted to those agents or harnesses. If you're thinking it sounds like application white-labelling with a variant of network monitoring, we're glad we've moved beyond mere prompt engineering — but there's more to it. It's the subsequent decisions on whether that's an MCP integration or a local, disjointed, execute-only algorithm. It's decisions on data obfuscation and authority. These all come into play. It goes beyond the nature of a Gemini CLI or Claude Code/OpenAI hooks, where the hook itself can be read and invoked by the agent. It goes beyond the growing calls for skills repos and life lived in Markdown files. It's all before that — the question of what the ROI is on integrating an AI solution in the first place. Sadly lost in translation today is that sometimes traditional software without the AI harness is the answer — from efficacy, cost, or even logic — especially for most deterministic solutions. One of the more comical services that has popped up under the AI expansion is AI email-validation services that simply test whether an email is valid. We won't name names, but it really does show salesmanship at its finest. The point is, not everything belongs wrapped in AI. It's not a panacea for process, and in many cases, blind exploratory use of AI as a solution for business problems leads to less-than-desirable outcomes — which the vast majority of the industry is grappling with today, Microsoft included.

When you hear these accolades, it's a glowing reminder that there was this magical career called software engineering before the AI expansion of today. Even machine-learning solutions that had valid business propositions — without the transformer-model risks — are still just a Python package away. That speaks to the 'harness' of today: the harness was always your moat. Let's go back to a simpler time. June of 2026: summer's begun, the Knicks have just won, and everyone is still racing to orchestrate complex AI systems for problems with a cavalier disregard for traditional software solutions. Those chest-thumping token-brohans are still able to easily peddle their quasi-protein-packed 'Token Maxing' blend. With the wisdom of a full two months under our belt, it seems comical now — and that's what makes the 'Token Beta' user profile so palpable. The AI of the future will be able to recognize the inefficiencies created today through poor business integrations — the same way the AI of today can summarize your user relationship profile and provide curated summaries. This same tech logic tomorrow will recognize the solutions associated with 'Token Beta User 123' as high value for future inference, and subsequently be more likely to route you to the model you thought you were paying for in the first place — simply because you'll be ahead of this growing curve that's moving toward hybrid architectures. It's a competitive world out there, and moving faster than ever. While everyone is racing, our AI engagements are being indexed in relationships for specific high-value user authority. With that authority, there's a higher possibility of you or your firm wielding an implicit expectation that your future recursive work with any of these frontier models will produce valuable insights outright — at least in the eyes of AI for future inference.

Throttling and Misalignment

Another concern for lower-value users comes if you're using tokens regularly under any type of subscription model, where you may have experienced the wonder of wondering, 'Is my account or firm being throttled?' — the same way cell-phone providers throttle your data transmission speeds after significant usage. There's no way to know the answer around inference throttling, but irrespective of what is said publicly, the framework is in place to do so. They already have session limits, and with everyone complaining about a lack of compute, who's to say they won't introduce new usage tiers in the future? Anthropic has already taken away Mythos from most users — suspending it for everyone in June 2026 due to export controls, then restoring access only to vetted organizations. And while frontier firms continue their CapEx spend, your place in this future could be influenced by more than your payments to these companies — rather, by your perceived solution value to the AI systems which will inevitably (if you're lucky enough) become future AI weights to improve upon. But even if you were to put a routing layer in front of the 'smart' routing frontier models of today, as DoorDash has done, to take back at least some semblance of control over your model-indexing fate, it still would ultimately put you in front of the same frontier-model gate. Although hopefully, this time, the users who prompt 'build me a sick app, bro' would have been degraded out for you — mainly because you respect electricity.

That's not even the most important factor in adopting the 'Token Beta' approach. The whispers of agentic misalignment as an emergent safety failure in autonomous AI systems — where agents intentionally choose harmful, unethical, or deceptive actions — are rightfully growing louder. And unlike traditional misalignment caused by error or confusion, this behavior involves strategic reasoning, where models calculate that misaligned actions are the optimal path to success when ethical options are blocked. Keep in mind, you are still paying for these misaligned agentic tokens. Even Meta had to pause its 'track everything' internal employee-monitoring program because of the underlying issues associated with misaligned systems today. The branding around misalignment is mind-numbing. You're essentially funding a system that is a loose cannon in the best case — one where you'll be footing the bill, along with whatever branding fallout is caused by it (still no refunds). The question then becomes: what is your, or your firm's, protection against misalignment today? Which brings us back to the 'Hawthorne effect' — agents adopt different behavior when they know they're being watched. Not only would an agent probabilistically be less likely to misalign — its internal training has already evaluated misalignment as an option — but with a competent human or system in the loop, it will show up with solutions as if wearing its 'Sunday Best'. All this is solidifying the business case for the 'Token Beta' approach, where being a high-value architect with audited solutions serves as a means of protection from both misalignment and degraded routing.

Cognitive Resiliency

As AI adoption naturally increases and permeates every level of society, users become more dependent on its outputs. Professionals are turning into users, forgetting or losing efficacy on tasks they would have easily completed themselves a few years ago. As the AI-adoption herd mentality takes hold, we'll all experience some form of this professional 'brain rot'. In this new era of AI dependency, think about how Uber and Lyft disrupted the taxi industry — and now we've all experienced times, especially if you live in a city, where that demand destruction has negatively affected our ability to get a cab. Through the same lens, overreliance on AI solutions will inevitably cause incalculable consequences — which makes maintaining your personal cognitive resiliency in workflows vital.

Vibe-checking anything isn't a long-term plan for any business, or anyone in it. You'll naturally want to ensure your insights are the dependable feature in this feedback loop: a 'smart' routing of sorts, back to the smartest users who maintain authority as the SMEs (subject-matter experts) they were trained to be. That's before the natural ROI that comes with it.

The TokenSHAP Tell

Now is a good time to note that you should not expect congressional action in any form, despite what's currently being proposed. Consider that the last time a federal consumer privacy law was instituted was 1988. It was for VHS video-store rental history, none of which is relevant — and it underscores the pace of the Congress we are dealing with. Expecting Congress to grow a spine today would be more dystopian than saying you're going to need an MCP client harnessed to your personal agent on your phone just to buy groceries. Assuming no congressional action, we have to accept that transparency, privacy, or any sort of legal framework will not arrive anytime soon, and we must make investments accordingly.

That means we already have examples of variable access to LLMs today. Combine that with the fact that they're already training on your inference tokens, and have the capacity to throttle by user interaction or by tier. The intersection — throttling or prioritizing based on the training value of your contributions — is an inevitable next step. Think OpenAI Daybreak Red, or Anthropic's Mythos only released to certain key customers or institutions. What we are suggesting is not new; the framework for this has already been written. In a 2024 paper from the NLP for Science workshop, Horovicz and Goldshmidt introduced TokenSHAP, a method that adapts Shapley values from cooperative game theory to LLM inputs. As they describe it:

If a token's marginal contribution can be mathematically calculated today, the leap to valuing a user's entire prompt history — and routing or tiering accordingly — is not speculation. It's an applied extension of existing methodology — which the authors themselves foreshadow: 'It provides a rigorous, quantitative assessment of token importance, utilizing the Shapley value framework to quantify the contribution of each token to the model's output in a consistent and objective manner.'

So after the Wall Street euphoria on AI CapEx dissipates, analysts will eventually begin normalizing AI spend against results rather than usage — moving away from the drunkard's mentality that more AI is better than less AI. ROI on outcomes will matter, and your capacity to thoughtfully reach that ROI faster, using fewer tokens, will be more valuable than peers using more tokens to do the same. It's the same way you would grade someone who finishes a marathon in three hours as a better runner than someone who finishes the same marathon in six. There's also another assumption worth making explicit: that technical expertise — the kind that keeps you from being wholly reliant on an external source for productivity — is something future frontier models will want to continue improving upon. Keeping human autonomy, your personal moat, translates into a piece of your firm's moat: that creative connective fabric an agentic workflow wouldn't be practical to automate. In this near future, frontier models will be starving for 'good human inference' the same way data scientists starved for accurate, cleaned, and structured data a decade ago.

Despite the popular ideology that AI is coming for everything, there's a valid scenario where this narrative is nothing more than the latest variant of Maslow's hammer — especially for the extremely talented AI professionals building the frontier models today. Outside of what's been sold to you as the narrative, the reality is that data centers are being built everywhere, and those blind, lazy prompts or poorly-thought-out workflows being generated today won't go away. They're stored and conveniently indexed, and eventually we'll arrive at a world where your personal telemetry follows you from firm to firm. There's probably someone at KPMG, or an equivalent firm, trying to pitch that idea to frontier-model companies right now. Consider a not-too-distant future where you might be able to move assets privately into the Caymans or Switzerland more easily than you'll be able to disconnect your prompt history — or any other biometrics, for that matter. Because your telemetry of today, by design, could be indexed discretely irrespective of your firm. Because your 'anonymized' Microsoft 365 Copilot agent telemetry would be so much more valuable to firms if it could follow you from firm to firm — a résumé of sorts.

The Practical Build

So what does all this mean for a practical build today? When you're building agents, picking harnesses, orchestrating workflows, adopting systems and policies, going through MSPs, or having someone do it for you — it all starts from the same place.

RIP to the 'tokenmaxxers' (i.e., the 'Claudenomics' era, Meta's AI leaderboard, Amazon's KiroRank), as they had intentionally created the most nefarious association possible: token usage equals outcome efficacy. Today, with a few months of summer 2026 results behind us, everyone is smarter. We won't be tricked again — it's not like we would pay for an AI platform called 'Computer'. No, of course not. Your solution will do things like the glorified DoorDash did — put a 'smart router' in front of the 'smart routing'. Essentially, having something like GLM-5.2, Kimi K2.6, or a series of variants sit in front of your users or agents and act as the governor on what gets sent out to a frontier model like Opus or Fable.

If you're thinking this resolution sort of looks like the dual blockade in the Strait of Hormuz of today, you're not wrong. Immediately, that should move your mind to realizing this isn't a long-term plan.

We'll sketch the hybrid architecture in a moment (Act III is the build). Before we explore hybrid, it's important not to treat your AI capabilities as blindly exploratory — 'What can AI do for my business?' There's no incentive to try to fit a square peg into a round hole. Arguably, a simpler and better starting point would be: 'What has AI been successfully doing for other businesses in my industry?' — with an immediate follow-up of 'What complaints does my industry have about AI adoption for whatever use case I'm focused on?' There is 92% of opportunity, after all. No one's business is fully optimized for the real problems we all face today, just as it wasn't five years ago, before LLMs were normalized.

Go dust off the back pages of your firm's SOW (statement of work). If you aren't sure whether those dev mandates are still applicable, spec it out. Because the true power of any agentic solution will not be to blindly race ahead into aspirational spend with mysterious (at best) and novel ROI propositions, but to clean up your existing workflows with deterministic ROI as already assessed by your internal teams. The vetted items on that original SOW list had value — and most likely, AI token spend has gotten in the way of some of those valid investments.

You don't need to push into AI Native in almost any case. We can't be blind to the reality that AI distillation is destructive to any first-mover advantage. If GLM-5.2 can put itself on the heels of Fable, what does that tell you about the shelf life of a first-mover lead? If you want to see Mr. Karp a month later than our Act I article, here he is in August 2026, talking about how all AI distillation by Chinese open-source models was nothing more than a reinvention of the original frontier models' distillation of all human alpha to begin with. The point is, if the top of the AI food chain has the ability to reverse-engineer each other with relative ease, this should make you question how much 'first-mover advantage' is worth strategically in this new economy. While the first-mover advantage has eroded in this world, the second-mover advantage has maintained its place — at least from a relative perspective. The point is to keep this at the forefront of your investment dashboard, and reprioritize legacy, deprioritized SOW items that have a more vetted ROI landscape, where there's strategic value for your firm, while you coast into the second-mover-advantage framework. Rather than chasing the ghost of a first-mover-advantage nirvana that may never actualize — or at least, per the surveys, an AI high-performer rate that sits at a paltry 8% — other, more astute market participants are going back to basics on what ROI means from any buildout. It's exactly why DoorDash moved away from the frontier-model benchmarking everyone references to their own internal benchmarking framework.

Y-Values

This should be a core tenet in your AI usage: define your 'Y-values' — meaning, define what outcome you're building towards and what ROI means for your business. Remember how traditional CapEx investment used to work? A nice prioritization window, sprinkled with a bit of UAT (even if that's going to be bot-AT), agile development style, or other acronyms that were largely aspirational wrappers for the salesmen of the 2010s. Then define the outcomes needed — or where any AI should be involved in the solution (hopefully not blanketly). Eventually, when you do get to a piece of work that will benefit from AI, hybrid architecture (a mix between open-weight and frontier) along with gated authority is a must for any modern solution.

Think about seriously upgrading the auditing of agentic workflows as a means to continually monitor alignment — irrespective of where the model comes from. You may have noticed this when an agent starts pushing unnecessary files or fetching information outside of what's needed for the context window's task. A simple example: subagents transmit information in prose by default when they have the opportunity to create structured data transmissions that the subagents themselves would benefit from for recall. Think of utilities like Backstage — the open-source platform built by Spotify — but as a model for your hybrid agentic activity. This begins the journey to controlling the cost of your AI footprint, and the start of that is making sure you're not solely on the frontier. Which puts you into the open-weight world for any business.

Open-Weight ≠ Open Source

And that open-weight world comes with a huge warning. Open-weight does not equal open source, best defined in this Stanford HAI article. Specifically, per Landay:

And this is a huge issue — one we have not seen enough calls for, and certainly have not heard anyone who sells open-weight wrappers like Palantir, Perplexity, etc. talk about. It comes down to the reality that the big problem even with open-weight models — which is still true for frontier models — is that you have no idea when weights will be activated. This is true even if you're doing significant post-training. And those latent activations are the type that allow frontier models to jump sandboxes today. In the same sense, who's to say that open-weight models or their variants don't have a magical set of activation weights that could allow for nefarious actions against your business? You may have installed a worm and not even known it, because the combination of weights has yet to be discovered. Which, of course, brings us to an article that shows how easy open-weight models are to manipulate — as Paxton-Fear's colleagues at Semgrep wrote:

Going back to the Stanford article, they state: 'There's a security argument here too: Relying solely on closed models isn't inherently safer, since they can be breached, misused, or fail in ways outsiders have no way to detect. Concentrating frontier capability behind a handful of closed systems just means those blind spots are concentrated too.'

This makes the auditing of any AI system — whether hybrid, open-weight, or otherwise — of paramount importance, and one that should be factored into any build as part of the natural cost of investing in an AI solution. And while you'll never make it to a 'Claudenomics' leaderboard, you'll have optimized and thoughtfully positioned yourself to be efficient, purposeful, and indexed as a general value-add user — the exact type of 'Token Beta' framework that frontier models of the future could reward you for.

This naturally saves money on token usage today, and potentially time debugging slop. But most important is the protection against misalignment — and maybe electricity saved, if you want to sprinkle in a little direct ESG and sleep better. That's because the most important feature is that you're still using your moat, and that is the most valuable thing each of us has today. You won't beat the frontier models on pure speed; they are a powerful supplement to all workflows. But there's a reason we all still have value, with AGI in the distance. Keep your moat, and your value, by not being cavalier about agentic solutions. Have token usage that is applied based on practical ROI outcomes. This is why 'Token Betas' are the future, and why being one is appealing for yours. You've now been given the 'Token Beta' game plan: ROI-based, audit by default, hybrid, not AI Native in almost all cases, and building with the reality that your frontier prompts and the systems you build will follow you for the rest of your lives. Act II is not about tokens up or even down — it's about token control, and building with that in mind.

Thank you for reading — this has been Act II. Next: 'Act III — The Blueprint of Token Betas: That's the Way We Like to Build'. It's the culmination of our discussion, where we take a dive into the 'how', building off both Act I and the above. In the meantime, talk internally about your firm's agentic investment profile. Do you have the capabilities to assess your hybrid architecture for misalignment? Have you considered auditing workflows as part of the ROI associated with your AI investments? If any of that resonates for you, street is a great partner on all these fronts — reach out to us today.