In a late-afternoon call I sat in on, someone quietly read the Twilio CEO’s line about ROI and tokenmaxxing aloud to the group.
That moment felt like a pivot: you could feel the mood shift from giddy experimentation to sober math, and I kept wondering whether the frenzy had actually died or merely changed its mask.
In another room, Shopify engineers were told bluntly to use only frontier models like OpenAI’s GPT-5.6 Sol and Anthropic’s Fable 5.
That edict from Farhan Thawar landed like doctrine — not cost-savings — and it forced me to ask what belief looks like when it wears a price tag.
What is tokenmaxxing?
At a coffee table demo I watched last month, someone used “tokenmaxxing” as a shorthand for throwing compute at every idea until one stuck.
That shorthand matters because the term captures two behaviors: indiscriminate volume and the emotional comfort of scale, and if you strip away the jargon you find a spending habit disguised as strategy.
Why are some companies choosing frontier models only?
On Shopify’s internal channels the rule is simple: if the model isn’t frontier, don’t bother.
That real-world policy reveals a wager — market speed and novelty are being treated as risk mitigants, and the logic is straightforward: use the most capable models to shorten feedback loops even if each token costs more.
In an interview the Twilio CEO framed the backlash differently by asking whether companies are netting ROI from AI use.
That framing nudges you to measure data and dollars rather than religious fervor, and when you combine his caution with Shopify’s faith, you see two competing playbooks: economize, or invest aggressively in perceived superiority.
At Olive, Bill Nguyen reported burning an eye-watering 774 billion tokens in a month, about $4.5 million (€4.2 million) in estimated costs.
That confession reads like a confession of faith in latency, fidelity and first-mover advantage, and it forces anyone who cares about unit economics to ask whether marginal gains on a frontier model justify marginal spend.
At a demo day I attended, founders defended frontier-only choices by pointing to faster iteration and less model hair-pulling.
That defense is seductive because time-to-market and competitive risk have real teeth, and for some companies the calculus favors paying more per token to dodge months of tuning and integration.
At a small startup I watched scramble, the engineering lead compared choosing cheaper models to running with one arm tied behind your back.
That metaphor hits because product-market fit is often a race, and in races some teams will bet their future on a perceived speed advantage.
In a hallway conversation I had with a former data-center exec, the word “waste” came up more than once.
That worry is practical: tokenmaxxing without measurable uplift looks, on spreadsheets, like burning cash rather than buying outcomes, and companies that lack clear metrics will pay for faith in models.
At a trade show I listened as a CTO argued that using only frontier models signals commitment to quality to customers and partners.
That stance creates social pressure — engineers and execs benchmark themselves against peers — and the result is tribal reinforcement where compromise feels like heresy.
At a dinner with product leaders I asked whether you can square frontier worship with fiscal discipline.
That question matters because you’re either optimizing for short-term market positioning or long-term efficient scaling, and mixing the two without measurements leads to strategic whiplash.
At one startup the CEO told me it was “frontier or nothing,” a line that sounded more like sermon than strategy.
That rhetoric produces disciples and drop-outs: some will double down, others will quietly build cost-aware paths around the altar, and I saw both happening in the wild.
At every table where I probed this behavior, two emotions surfaced: fear of being slow and fear of being wrong.
That dual fear fuels spending and defends it, and it explains why tokenmaxxing didn’t die — it simply mutated into a premium-only creed where the cost is justified by the promise of competitive immunity.
At this point you should ask whether your team treats model choice as a vendor decision, an engineering choice, or a strategic bellwether.
That practical self-audit is how I advise founders: measure the lift you get per dollar, pressure-test assumptions about time-to-market, and let data, not dogma, set your spending cadence.
At the end of several conversations I heard the same metaphor again and again, comparing frontier loyalty to a cathedral of belief.
That image explains why governance and incentives matter — when faith looks like architecture, policies follow and dissent becomes costly, and you must decide whether you’re joining a congregation or running a business.
At the closure of a board meeting I attended, someone suggested a middle path: pilot frontier where it moves the needle and use cheaper models for predictable workloads.
That compromise can work if you instrument outcomes tightly and set abort points, but few organizations have the discipline to do it without political fallout.
At every turn I kept asking myself: do you want to be the company that spent as much as it could on compute, or the one that proved returns on that spending?
That question is the test you owe your investors, your engineers and your customers — and the answer you choose will define whether your frontier-only credo is a smart bet or an expensive faith?