AI Token Prices Are Hitting New Record Lows

The cost of thinking machines keeps falling

The price of using artificial intelligence is collapsing. According to a CNBC report published Tuesday, the cost of AI “tokens” — the basic units of text that large language models read and generate — has fallen to new record lows, extending a trend that has defined the industry since generative AI went mainstream.

Tokens are the currency of the AI era. Every prompt a user types and every word a model writes is broken into these small chunks, and developers are billed per million of them. That pricing model has made the token the closest thing the industry has to a commodity benchmark: a single number that captures how expensive it is to put machine intelligence to work.

That number keeps shrinking.

Why prices keep dropping

Several forces are pushing in the same direction at once.

The first is competition. A handful of well-funded frontier labs, a growing field of challengers, and an increasingly capable open-weight ecosystem are all chasing the same developers. When rival models perform comparably on everyday tasks, price becomes the sharpest differentiator — and the fastest way to win an account is to undercut the incumbent.

The second is engineering. Model builders have gotten dramatically better at squeezing more output from the same silicon, through techniques such as quantization, distillation into smaller models, smarter batching of requests, and caching of repeated inputs. Newer generations of accelerators deliver more throughput per dollar and per watt. Each improvement lowers the floor on what a provider can charge and still make the economics work.

The third is strategy. For many providers, cheap tokens are a customer-acquisition tool. Getting developers to build on a platform now, at a loss if necessary, is a bet that usage will grow fast enough — and lock in deeply enough — to pay off later.

What it means for the industry

For businesses building on top of AI, falling token prices are unambiguously good news. Applications that were financially absurd a couple of years ago — summarizing every customer support ticket, reviewing every line of code, running agents that make hundreds of model calls to complete a single task — become plausible line items. Cheaper inference is arguably the single biggest unlock for the “agentic” software that labs have been promoting, because agents are token-hungry by design.

For the providers themselves, the picture is more complicated. Deflation in a core product is uncomfortable when the capital costs behind it are enormous. Training frontier models and building the data centers to serve them require spending on a scale usually associated with utilities or heavy industry. If revenue per token keeps sliding, the only path to healthy margins is volume — vastly more usage, or higher-priced premium tiers for the most capable models and the most demanding workloads.

That tension helps explain why the industry increasingly resembles a classic commodity market layered underneath a premium one. Routine, high-volume tasks are drifting toward near-zero marginal cost, while the newest reasoning-heavy models still command a premium from customers who need them.

The bigger question

Record-low token prices raise an awkward question for investors: is the value in AI accruing to the companies that make the models, or to the ones that use them? Historically, when an input becomes cheap and abundant, the profits migrate upward to whoever builds the best product on top of it.

If tokens keep getting cheaper — and every incentive in the market suggests they will — the winners of the next phase may be less the labs racing to the frontier than the businesses quietly turning near-free intelligence into something customers will pay for. Read More