Insights

Tokenomics: Why your AI bill keeps growing while AI keeps getting cheaper

Adnan Masood, PhD. Chief AI architect, UST

Two facts sit side by side in most technology budgets today. The unit price of AI computation has fallen steadily since 2023, in some cases by an order of magnitude. Over the same period, total AI inference spend has risen every quarter. Neither number is wrong, and reconciling them has quietly become one of the more consequential cost-management questions facing enterprise leadership in 2026.

The explanation is structural. Modern AI models "think" before they answer, and that thinking is billed at premium rates while remaining invisible in most product dashboards. AI agents make twenty model calls where an application once made one. Context windows have grown to accommodate entire document repositories, and every token in those windows is billed on every call. Unit prices fell; unit consumption exploded. This is a familiar pattern to anyone who lived through the cloud computing transition. The lesson from that era applies directly: the winners were the companies that treated consumption as a governance problem rather than a procurement problem.

I call this discipline tokenomics: the deliberate allocation of a fixed intelligence budget across a portfolio of workloads to maximize business outcomes per dollar. The word "portfolio" is chosen carefully, because the right mental model comes from finance rather than IT.

DIVIDER

Paying frontier prices for routine work

Every major AI provider now sells a hierarchy of models at dramatically different price points. Anthropic's lineup spans roughly $5 per million output tokens at the small tier to $50 at the frontier tier, a tenfold spread. Google's spread runs closer to thirtyfold. OpenAI's, across its full family, is wider still. These tiers exist because most enterprise work does not require frontier intelligence. Classifying a transaction, extracting fields from a form, and deflecting a routine support question are tasks a small model performs with accuracy indistinguishable from its expensive sibling, at a tenth to a thirtieth of the cost.

Here is the uncomfortable finding from every engagement I have led: the default enterprise behavior is to route everything to the most capable model, the way a company might send every package overnight because nobody built a mailroom policy. The price spread between tiers is the ceiling on your savings, and at 10x to 30x, that ceiling is one of the largest controllable variables in your technology cost structure.

DIVIDER

Three flavors of solution, in order

The market offers a menu of approaches, and the sequencing matters more than the selection.

The first flavor is disciplined assignment, and it captures most of the value. Map each workload to the cheapest model that passes a quality bar you define. Cache the repeated portions of prompts, which providers discount by up to 90 percent. Move overnight and back-office work to batch processing, a flat 50 percent discount across every major provider. Cap the hidden "reasoning" spend on simple tasks. None of this requires sophisticated technology; it requires a policy, an owner, and about six weeks. In our experience and in the published evidence, this alone reduces inference spend by 50 to 70 percent at held quality. One engineering organization documented cutting total AI spend by nearly 60 percent purely by restructuring prompts to exploit caching.

The second flavor is intelligent routing, and it is genuinely powerful but second-order. A router is software that inspects each incoming request and automatically selects the cheapest capable model. The peer-reviewed research is real: learned routers have demonstrated cost reductions above 80 percent on favorable benchmarks. But those headline numbers were measured on model pairs with price spreads no longer available, and on academic benchmarks that predict enterprise task success poorly. On realistic domain traffic, expect routing to add an incremental 10 to 25 points beyond disciplined assignment, and only where request complexity genuinely varies. Cloud providers now sell this natively; AWS and Microsoft both offer managed routers with published, and modest, savings figures.

The third flavor is permanent migration. At very high volume, stable, repetitive workloads can be moved entirely onto small, custom-trained models, eliminating variable API costs. This is the highest ceiling and the highest effort, and the economics are only roughly 50 million tokens per month on a stable task. Below that threshold, the project never pays back.

DIVIDER

The risk that matters is asymmetric

Leadership should internalize one asymmetry above all others. Sending a simple question to an expensive model quietly wastes pennies. Sending a hard question to a weak model yields an incorrect answer, triggering retries, escalations, compliance exposure, and eroded customer trust. The industry ran this experiment publicly: OpenAI's GPT-5 launch in August 2025 shipped a cost-saving router that sent too many complex queries to a weak model. The backlash forced Sam Altman to issue a public apology and a rollback within four months.

The governance implication is direct. The threshold that determines how aggressively to economize is a business decision about risk appetite and belongs to leadership rather than to individual application teams. Bias it toward overspending on ambiguous requests until production evidence supports tightening.

DIVIDER

The metric and the questions

Hold your organization to cost per successful task, never cost per token. A cheap wrong answer that requires three retries and a human review costs more than one expensive right answer, and per-token dashboards hide this completely.

Three questions for your next technology review. First: what share of our AI spend runs on frontier models, and what fraction of that work actually requires them? If nobody can answer, you have found your problem. Second: who owns the quality bar and the routing threshold for regulated workloads, and can we explain, after the fact, why a given decision went to a given model? Third: have we exhausted the boring levers, caching, batching, and tiered assignment, before funding anything sophisticated?

Framed properly, tokenomics is capital allocation applied to intelligence: cutting spend where quality is safely substitutable and protecting spend where failure is expensive. The organizations that build this discipline now will fund their next wave of AI ambition out of the waste in their current one. The ones that do not will keep wondering why the cheapest technology in history keeps getting more expensive.

Learn why token efficiency is becoming the next competitive advantage.