The Efficiency Turn: How Enterprises Killed Tokenmaxxing and Repriced the Frontier

For most of the last two years, the dominant verb in enterprise AI was spend. The strategy that won internal arguments was simple: route every request to the most capable model available, give developers an unmetered key, and treat the bill as the cost of not being left behind. Context windows grew, so teams stuffed them. Reasoning models charged for thinking tokens, so teams let them think. The implicit thesis was that intelligence was scarce and money was cheap, and the worst outcome was being too cautious while a competitor shipped faster. There was even a name for the behavior, half-joking and half-confession: tokenmaxxing — maximizing token consumption as if throughput itself were the product.

In the last weeks of June 2026, that thesis broke in public. CNBC reported that OpenAI and Anthropic are facing a new commercial reality as their largest customers shift from tokenmaxxing to efficiency. The anecdotes underneath the trend were specific and uncomfortable: Uber reportedly burned through its entire annual AI budget in roughly four months, and the chief executive of the AI startup Lindy moved one hundred percent of the company traffic off Anthropic Claude models onto DeepSeek to cut costs. At the same time, the FinOps Foundation made AI token economics the center of its 2026 conference, recasting the AI invoice as something to be governed rather than simply paid. None of this is a story about enterprises abandoning AI. It is a story about enterprises finally pricing it.

This is the efficiency turn. The competitive surface of AI is moving — not the one everyone watches, where labs race to ship the next frontier model, but the quieter one underneath it, where buyers decide how much intelligence a given task actually requires and what they are willing to pay per unit of it. That shift reprices the entire market from the demand side, and it has direct consequences for the frontier labs whose revenue curves were drawn on the assumption that token consumption only ever goes up.

What Tokenmaxxing Actually Was

It helps to be precise about the behavior that is ending, because the word makes it sound like a fad and it was not. Tokenmaxxing was a rational response to a specific set of incentives that held for about two years.

The first incentive was capability uncertainty. When the gap between a flagship model and a cheaper one was the difference between a feature working and not working, paying for the flagship on every call was defensible. Nobody got fired for routing a customer-facing agent to the best model on the menu. The second incentive was organizational. AI budgets in 2024 and 2025 were frequently carved out of innovation funds or strategic initiatives, not operating budgets, which meant they were measured by what they enabled rather than what they cost. A team that shipped an AI feature was a success; the line item it generated was somebody else problem, often nobody problem. The third incentive was the pricing structure itself. Per-token billing made the marginal cost of one more call feel like nothing, and a thousand calls that each feel like nothing add up to a number that does not feel like nothing at all when the invoice arrives.

Put those incentives together and you get the behavior the industry is now unwinding: developers encouraged to pump as many tokens as possible into the most expensive model, with quality of output as the only metric anyone tracked and cost as an afterthought reconciled quarterly, if at all.

The problem with tokenmaxxing was never that it was wrong in 2024. It was that the conditions that justified it quietly expired. Capable models got cheap, the budgets got scrutinized, and the invoices got large enough to attract the attention of people whose entire job is to ask whether a dollar of spend produced a dollar of value. All three of those changes happened at once, and the result is the turn we are now living through.

The Week the Logic Flipped

What made late June feel like an inflection rather than a slow drift was the concentration of signals. The behavioral report, the model economics, and the governance response all surfaced inside the same window.

The Lindy detail is the one worth sitting with. Moving all of a company traffic from a flagship Western model to a cheaper alternative is not a procurement tweak; it is a statement that, for that company workload, the quality difference no longer justifies the price difference. That judgment was unthinkable to make loudly in 2024, when the flagship was visibly better at almost everything. It is sayable now because the budget-tier models got good enough that the comparison is no longer obvious — which is precisely the condition that lets price, rather than capability, become the deciding variable.

This is the demand-side mirror of a supply-side story I covered when OpenAI and Qualcomm moved the chip war off the training floor in the inference-silicon turn. The hardware industry is rebuilding itself around the fact that the expensive, recurring part of AI is inference. The efficiency turn is what that same fact looks like from the buyer side of the invoice: if inference is the recurring cost, then optimizing inference consumption is the recurring saving.

The Unit Economics That Made It Possible

The efficiency turn did not happen because buyers found discipline. It happened because the price of capability collapsed far enough that discipline finally paid off. When the flagship is the only model that works, frugality saves you nothing. When a model a fraction of the price clears the bar for most of your traffic, frugality saves you most of your bill.

The headline price table tells the first half of the story. A budget-tier frontier-class model now serves at roughly fifteen cents per million input tokens and under thirty cents per million output tokens — an order of magnitude beneath where flagship pricing sat two years ago, and meaningfully below the mini and nano tiers from the larger Western labs. When the cheapest credible option is that far under the default option, the question is no longer whether to optimize but how much of your traffic you were overpaying for.

The second half of the story is structural, and it is where the FinOps framing earns its place. Treating the token invoice as the whole AI cost is itself a tokenmaxxing-era habit. The token line is only one of several cost buckets — it sits alongside the compute to host any self-managed inference, the data egress and retrieval costs of the systems feeding the model, the human review the outputs require, and the engineering time spent building and maintaining the whole apparatus. A team that drives its token bill to zero by self-hosting a cheap model and then spends three engineers babysitting the cluster has not saved money; it has moved the cost to a bucket nobody is counting. The discipline the efficiency turn demands is not minimizing one number. It is governing the total, which is exactly the muscle FinOps teams already have for cloud.

The reason this matters now and not a year ago is that capability stopped being the bottleneck. When the cheap models were visibly worse, no amount of FinOps rigor could safely push traffic down a tier. The collapse in the quality gap is what converted cost governance from a way to degrade your product into a way to keep your product the same while paying far less for it. I argued a related point about why raw capability has become a misleading axis to optimize in the benchmark illusion; the efficiency turn is the commercial consequence of taking that argument seriously.

Why The Quality Gap Closed

Every part of the efficiency turn rests on one technical fact: the cheap models got good. If they had not, none of the discipline would matter, because no FinOps review can safely route traffic to a model that breaks your product. Understanding why the gap narrowed is what lets you judge whether it will stay narrow.

The first driver is distillation. The cheapest credible models are increasingly trained on the outputs of the most expensive ones, compressing a large share of the frontier model behavior into a far smaller network that costs a fraction to run. Distillation does not capture everything — the hardest reasoning and the longest-horizon tasks still favor the large model — but it captures the broad middle of everyday requests remarkably well, and the broad middle is where most production traffic lives. The second driver is post-training. The gains of the last year came less from raw scale and more from better instruction tuning, preference optimization, and reasoning training applied to modest base models. Those techniques transfer down the size curve, so a small model in 2026 behaves more like a large model from a year earlier than its parameter count would suggest. The third driver is architectural efficiency. Advances in how models are served — including the move toward inference-optimized designs I covered in the diffusion-LLM turn — keep pushing the cost of a given quality level down independent of the model itself.

The practical consequence is a moving bar. The threshold of quality that a budget tier can clear keeps rising, which means the share of your traffic that can safely move down a tier keeps growing. A workload that justified the flagship on sixty percent of calls a year ago might justify it on twenty percent today, with the rest handled acceptably by a model costing a tenth as much. The efficiency turn is not a one-time reallocation; it is a ratchet that tightens every time the cheap models improve, which is every few months.

The risk to the thesis is a re-widening at the frontier. If the next generation of flagship models opens a large new capability gap — a genuine step change in reasoning or reliability that the budget tier cannot follow — then for the workloads that need that capability, pricing power swings back to the flagship and frugality once again saves nothing. This has happened before, briefly, with each reasoning-model generation. The reason it has not held is that the gap closes faster than it opens: distillation and post-training propagate the new capability down the cost curve within months. Betting against that propagation is betting that the labs will stop being able to make small models good, and nothing in the last two years supports that bet.

The Anatomy Of An AI Bill

The abstract case for governing the total cost rather than the token line becomes concrete the moment you trace a single workload end to end. Consider a customer-support agent handling a high volume of tickets — the canonical enterprise AI deployment, and one where the tokenmaxxing reflex did real damage.

In the spend-first configuration, every incoming ticket is sent to the flagship model with the full conversation history, the entire relevant knowledge-base article, and a long system prompt that has accreted instructions over months. Each ticket is therefore a large input on an expensive model, repeated across every ticket in the queue, with the same system prompt and the same boilerplate re-sent and re-billed on every call. The token invoice for this configuration is enormous, and it was the only number anyone watched.

Now trace the same workload through the efficiency lens. The majority of tickets are routine — password resets, order status, simple policy questions — and a budget-tier model answers them indistinguishably from the flagship. A cheap router reads each ticket and sends the routine majority down a tier, escalating only the genuinely ambiguous or high-stakes cases. The stable system prompt and the knowledge-base content become cached prefixes, so the repeated boilerplate is billed at a steep discount rather than at full rate on every call. Retrieval returns the two relevant passages rather than the whole article, cutting the input size. The token invoice falls by a large multiple — and crucially, the quality the customer experiences does not move, because the routing only downshifts the requests that never needed the flagship.

The lesson generalizes. Almost every production AI workload has a tail of hard requests that genuinely need the best model and a body of routine requests that do not, and the tokenmaxxing era paid flagship prices for both because nobody was measuring the split. The efficiency turn is, at bottom, the organizational decision to measure that split and pay accordingly. It is unglamorous, it is the kind of work cloud FinOps teams have done for a decade, and it is about to become one of the highest-return engineering activities a company with an AI bill can invest in.

What This Does to the Labs

The uncomfortable part of the efficiency turn is what it implies for the companies selling tokens. Their revenue narratives were built on consumption that compounds — more users, more calls, more tokens per call, all multiplying together. Efficiency attacks the third term directly and the second one indirectly.

The labs are not in trouble; they are growing at rates most companies never see. The question the efficiency turn raises is subtler than survival. It is whether the steepest part of the curve is already behind them. Analysts looking at the shift note that current growth rates are likely the fastest these companies will ever post, precisely because a meaningful share of that growth came from customers who were not optimizing — and who are now starting to. When your biggest accounts move from unmetered keys to governed budgets, you do not lose them, but you lose the part of their growth that came from waste.

Both major Western labs filed confidentially for public offerings in early June, which sharpens the stakes considerably. A consumption-growth story is wonderful in a private round, where the narrative is the product. It is more fragile in public markets, where the durability of the growth gets priced. I wrote about how the prospectus itself would make the bubble debate falsifiable in the Anthropic S-1 bubble test; the efficiency turn is one of the specific mechanisms that could bend the revenue line, because it directly targets the assumption that token volume per customer keeps climbing.

There is a defensive move available, and the labs are already making it: ship your own cheap tier so the downshift happens inside your product rather than to a competitor. Microsoft building first-party MAI models, the budget-tier offerings from every major provider, and the aggressive sub-flagship pricing are all the same play — capture the efficiency-seeking customer on your own cheaper SKU instead of losing them to DeepSeek. It works, but it changes the economics: a customer who moves from your flagship to your budget tier is still a saved account and a much smaller invoice. Efficiency cannibalizes margin even when it does not cost you the logo.

The Builder's Playbook

For the people actually shipping AI features, the efficiency turn is less a threat than a set of techniques that finally have organizational permission. None of these are new; what is new is that someone above you now wants them done.

Put a router in front of the flagship. The single highest-leverage move is to stop sending every request to the best model. A cheap classifier model can read an incoming request, decide whether it is trivial or hard, and route accordingly — trivial requests to a budget tier costing a fraction of the flagship, hard requests escalated to the expensive model. For most real traffic mixes this removes a large share of the bill without users noticing any change in quality, because the easy requests never needed the expensive model in the first place.

Cache aggressively. Repeated and near-repeated prompts are the norm in production, not the exception. Prompt caching, automatic prefix caching, and context caching across the major providers make repeat traffic several times cheaper than the headline rate. A workload with a stable system prompt and high request volume can see its effective per-call cost fall sharply just by ensuring the cacheable prefix is structured to actually be cached.

Right-size the context. Tokenmaxxing made stuffing the context window a reflex. Most tasks do not need the entire document, the full history, and every retrieved chunk. Retrieval that returns the three relevant passages instead of thirty cuts input tokens directly, and frequently improves output quality by reducing distraction.

A caution on that chart: the savings are not additive. Once a router has moved the bulk of traffic to a budget tier, the incremental benefit of caching on the remaining flagship calls is smaller in absolute terms. The discipline is to sequence the moves — router first, because it touches the most volume, then caching, then context hygiene — and to measure the total bill after each, not to assume the percentages stack.

Instrument value, not just cost. The FinOps framing is right that the metric to chase is value per token, not cost per token. The cheapest possible bill is zero tokens and no product. The goal is the quality you actually need for each task at the lowest cost that still clears the bar — which means you have to define the bar per task, measure whether the cheaper tier clears it, and only then route down. Cost optimization without quality measurement is just quiet product degradation, and it shows up later as churn rather than savings.

This connects to a procurement-side shift I traced when the buying model itself started to wobble in the universal-credits collapse. Efficiency at the engineering layer and discipline at the procurement layer are the same instinct arriving at two different parts of the organization: the era of buying AI by the bucket and using it by the barrel is closing.

The Counterargument: Jevons Strikes Back

The strongest case against the efficiency turn being bad news for the labs is the oldest one in resource economics. The Jevons paradox holds that when a resource becomes more efficient to use, total consumption of it often rises rather than falls, because the efficiency unlocks uses that were previously uneconomical. Cheaper tokens could mean dramatically more tokens — agentic workflows that run continuously, AI features embedded in products that could never have afforded them, consumption that grows the market faster than per-task frugality shrinks it.

This is a real possibility and the labs are betting on it explicitly. Sub-flagship pricing is partly a Jevons play: make intelligence cheap enough that it shows up everywhere, and total volume swamps the per-call savings. If that bet pays, the efficiency turn is not a headwind at all but a flywheel, and the customers reining in waste today are funding the agentic workloads that consume far more tomorrow.

The reason I do not think Jevons fully neutralizes the turn is timing and attribution. Even if the market grows, the growth shifts from existing customers overspending to new workloads at deliberate, governed rates. That is a healthier business but a different one — lower revenue per token, more competitive, with price as a first-class axis rather than an afterthought. The labs whose valuations were set against the tokenmaxxing curve have to re-underwrite against the governed one. Even in the optimistic Jevons world, the per-customer overspend that flattered the early numbers does not come back. The classification of AI as permanent operating infrastructure, which I covered in the enterprise reclassification, is precisely what brings the cost discipline that ends the overspend — you do not let a permanent operating cost run unmetered.

The Second-Order Effects

The efficiency turn does not stop at the boundary between labs and their direct customers. It propagates outward into the whole application layer that was built on top of cheap-seeming tokens, and the second-order effects are where the most interesting consequences live.

Start with the AI-native startups. A large cohort of companies founded in 2024 and 2025 priced their products against token costs they assumed would keep falling and usage they assumed would keep climbing — and many of them resold flagship intelligence at a markup. Those companies now face the same repricing their customers do, from both sides. On the cost side, their own token bills are under the same scrutiny, which is why a company like Lindy moves its entire traffic to a cheaper model the moment the math works. On the revenue side, their customers are running the same efficiency analysis on every subscription, asking whether the wrapper around the model is worth its margin when the underlying model got an order of magnitude cheaper. The startups that survive the turn are the ones whose value was never really the token passthrough — the ones that own a workflow, a dataset, or a distribution channel the model alone cannot replace.

Then consider where the saved money goes, because it does not simply vanish from the market. A company that cuts its support-agent bill by a large multiple does not usually return the savings to shareholders; it redeploys them into the next AI workload that was previously uneconomical. This is the Jevons mechanism operating at the level of a single budget: efficiency on the existing workload funds expansion into new ones. The net effect on total spend is genuinely ambiguous and probably varies by company — but the composition of the spend shifts unmistakably from a few expensive workloads run wastefully toward many cheaper workloads run deliberately. That is a more durable foundation for the build-out than the tokenmaxxing curve ever was, precisely because it is grounded in measured value rather than in the fear of being left behind.

Finally, the turn changes what good engineering looks like inside these organizations. For two years, the prestige move was to wire up the most capable model and ship fast; cost was somebody else department. The efficiency turn makes cost-aware architecture a first-class engineering skill — routing, caching, retrieval hygiene, per-task quality measurement, and the discipline to choose the smallest model that clears the bar. These are not glamorous, but they are becoming the difference between an AI feature that is sustainable and one that gets killed in the next budget review. The engineers who internalize the turn early will spend the next two years being the people who can ship the same AI product their competitors ship while paying a fraction of the bill — which, in a market that has finally started counting, is the most valuable thing an engineer can be.

What To Watch

A handful of concrete signals will tell you whether the efficiency turn is a durable repricing or a quarter of belt-tightening that reverses the moment the next must-have model ships.

Watch the published price moves from the flagship labs. Deep, proactive cuts to flagship API pricing — not just cheaper new tiers, but reductions on the premium SKUs — would be the clearest admission that the demand side now sets the price. Watch the revenue-per-customer disclosures if and when the IPO documents become public; a widening gap between user growth and revenue growth is the efficiency turn showing up in the financials. Watch whether budget-tier models keep closing the quality gap, because the entire turn rests on cheap models being good enough, and a sudden re-widening of the gap at the frontier would hand pricing power back to the flagships. And watch where the FinOps discipline spreads next: the first wave is the obvious overspenders, but the technique generalizes to every company with a token bill, which is rapidly becoming every company.

The deeper signal is cultural. Tokenmaxxing was a belief as much as a behavior — the belief that more intelligence applied more liberally was always better, and that the cost of restraint exceeded the cost of waste. The efficiency turn is the moment that belief stops being free. It does not mean enterprises believe in AI less. It means they have started treating it the way they treat every other input that became essential: as something to be measured, governed, and bought no larger than the job requires. That is not the end of the AI build-out. It is the moment it grows up.