Deep Read, Reasoning

The Costly Bet on AI That Thinks Harder

Reasoning-optimized models now consume half of all AI token usage despite their computational expense. Developers are choosing slower, pricier systems because the quality gains justify the cost.

The Curio Team · 5 min readWritten with AI from the sources below and checked against them before publishing.

Something shifted in the past year that the efficiency-first crowd did not expect. Reasoning-optimized AI models, the kind that think out loud through long chains of intermediate steps before committing to an answer, now account for half of all tokens processed across major AI platforms. These systems are slower, more expensive to run, and by any straightforward cost-per-token calculation they should be losing to their leaner competitors. They're not.²

Why Reasoning Models Won Despite the Cost Math

Large Reasoning Models, or LRMs, extend what standard Large Language Models can do by generating explicit intermediate thinking tokens before producing a final answer. That process isn't free. Those tokens cost compute, time, and money, and they scale with problem complexity. A model reasoning through a hard coding problem might generate thousands of tokens the user never sees, just to arrive at a solution a cheaper model might have attempted in a fraction of the steps. By the raw efficiency measure, LRMs look wasteful.¹

And yet developers keep routing workloads through them. The share of total tokens handled by reasoning-optimized models has climbed to half of all usage. That's not a niche preference. That's a market verdict. Our read is straightforward: when the output quality gap is wide enough, cost-per-token stops being the decision variable. Capability becomes the decision variable.²

Half of all AI token usage now flows through reasoning-optimized models, even though they generate far more tokens per task than standard alternatives.

The o1 Moment That Changed Developer Calculus

The inflection point has a name and a date. OpenAI's o1 series, documented by Jaech and colleagues and released in late 2024, marked what researchers now call a significant breakthrough in AI reasoning. The o1 architecture integrated explicit reasoning steps into generation in a way that moved benchmark performance on hard mathematical and scientific problems well beyond what prior models could achieve. Developers who tested it on tasks that had previously required human expert review started seeing outputs that cleared the bar without intervention.¹

That experience reset expectations. Once a team sees a reasoning model handle a problem a standard model fumbles, the calculus changes. The question stops being 'can we afford the extra tokens?' and starts being 'can we afford to ship the wrong answer?' For code generation, legal document review, scientific analysis, or multi-step planning, the answer to that second question is almost always no. The o1 release didn't just introduce a better model. It introduced a new reference point for what good looks like.¹

• LRMs generate intermediate thinking tokens before producing final answers, making each task more computationally expensive than a standard model call.
• The o1 series from OpenAI, released in late 2024, demonstrated what reasoning-optimized systems could do on hard benchmarks.
• Despite the overhead, reasoning models now handle half of all token traffic across major platforms.
• The efficiency penalty is real and documented, but developers are absorbing it because quality gains justify the cost.

The Scaling Problem That Makes Efficiency Urgent

Reasoning models are winning on quality, but their token consumption doesn't scale linearly with task difficulty. A slightly harder problem can require dramatically more intermediate steps. That means the compute bill for an LRM-heavy product can spike in ways that are hard to predict and harder to budget. The inefficiency that looks manageable at low query volumes can become painful at production scale.¹

This isn't theoretical. The research literature on LRMs explicitly flags that these systems are significantly less efficient than standard LLMs because of the intermediate token requirement. At half of all token usage already, and with adoption still growing, the aggregate compute cost the industry is absorbing is substantial. The open question is whether that cost curve bends before it breaks anyone's infrastructure budget. So far the market keeps paying, but that tolerance isn't infinite.^{1 2}

When Better Beats Cheaper

We've seen this pattern before. The product that wins on raw performance often beats the product that wins on efficiency, right up until the efficiency gap closes enough to matter. Reasoning models are in the performance-wins phase right now. The 50% token-share figure is the evidence. Developers aren't choosing these systems because they misread the pricing. They're choosing them because the outputs justify the overhead.²

What this means for AI economics is that the standard assumption, that cheaper inference always wins over time, needs revision. The reasoning model category has demonstrated that a sufficiently large quality gap can sustain a premium product even when the cost differential is obvious and measurable. That reshapes how we should think about competition between model providers, about where infrastructure investment flows, and about which benchmark numbers actually predict market outcomes. Accuracy on hard tasks, not cost per token, is the metric that moved half the market.^{1 2}

The o1 series reset developer expectations in late 2024. Once teams saw what reasoning models could do on hard problems, cost-per-token stopped being the primary decision variable.

Our take is that this is a durable shift, not a temporary enthusiasm. The use cases driving reasoning model adoption, code that has to be correct, analysis that has to be defensible, plans that have to hold together across many steps, are not going away. Neither is the quality gap between a model that reasons carefully and one that doesn't. The industry will work on the efficiency problem, but the destination is a world where reasoning capability is the baseline expectation, not a premium feature. We're already halfway there, literally.