Token Prices Are Falling 97% While OpenAI's Revenue Doubles — the Cost-Per-Task Pivot Is What Connects Them

In late July, OpenAI cut the price of its cheapest GPT-5.6 model by 80%: Luna fell from $1.00 to $0.20 per million input tokens, and from $6.00 to $1.20 per million output tokens. Cutting your lowest-priced product by 80% on the eve of an expected IPO looks like a surrender of pricing power. Unless the thing being sold has changed.

Read the surface numbers and they contradict each other. Per-token prices have fallen about 97% from GPT-4 to GPT-5.4, a deflation rate that would gut a business pricing by the word. Yet OpenAI has driven its annualized revenue run rate to more than $40 billion — roughly double where it sat at the end of 2025, by Bloomberg's reporting — with a public listing expected. A company cannot lose 97% of its unit price and double its revenue unless the unit it actually charges for, and costs out, is no longer a word. The contradiction is a placeholder for the real story: the unit being sold is changing, not just the price.

What is being sold changed, not just what it costs

A token price meters an input: how much text goes in and how much comes out. For a chatbot that meter had some meaning. But the fast-growing product is the agent — a program that reasons, calls tools, retries, and loops until a task is done. Its buyer pays for the finished outcome — a support case closed, a change merged, a loan approved — not for the message count along the way.

OpenAI has been explicit about this re-basing. Its own guidance to enterprise customers is to track "cost per accepted outcome," not token usage, because the lowest token price doesn't produce the lowest total cost if a model fails, retries, or needs correction. The July 30 price cuts were announced in those same terms: GPT-5.6 Luna, OpenAI says, delivers frontier-class performance from a year ago at about six cents on the dollar per task and roughly nine times the speed. Per-token deflation is real, but the company is increasingly letting the market price finished tasks.

That framing flips the obvious fear. The worry has been that cheaper tokens make AI companies worse businesses. The decomposition suggests the live condition is the reverse: per-task prices can fall and still be profitable if two things move correctly — the number of tasks grows, and the cost to produce each one falls faster than its price.

The chip is the supply-side half of that bet

The second condition is where the chip expansion enters. In June, OpenAI and Broadcom unveiled Jalapeño, an inference chip built from the ground up for running large language models. Benchmarks OpenAI published in August put it ahead of Nvidia's Blackwell systems by a wide margin — about 1.5 to 1.9 times more AI work per watt, and 1.7 to 3.6 times lower end-to-end latency across the models tested. Latency matters disproportionately for agents, because an agentic task strings many sequential steps together; shaving each step compounds the saving.

What a chip does is cut the cost of each completed task — which is precisely what makes an 80% price cut survivable. Formally, it is an operating leverage story: if useful work grows faster than the cost to serve it, margins expand even as headline prices fall. The unit economics of "cheap, abundant" AI depend on the manufacturing side of OpenAI losing cost as fast as the pricing side does. Broadcom supplies the design, and OpenAI plans to deploy the chip at gigawatt scale alongside Microsoft, ramping to "full tilt" in the first half of 2028.

Finance is the demand-side half

The average retail investor cannot buy OpenAI directly until that IPO lands. But the finance expansion is where the model becomes legible as a template — because finance is a task market where the completed outcome carries a large, measurable dollar value.

The commercial case is Customers Bank. Under a multiyear deal, OpenAI engineers are embedded at the bank to automate lending and onboarding. Customers Bank targets cutting its efficiency ratio — operating costs as a share of revenue — from about 49% to the low 40s. It says AI-completed work has already saved roughly 28,000 hours, close to 15 employees' worth, and that commercial loan closings could drop from 30–45 days to about seven. A task carrying that dollar value is one where "cost per outcome," not token metering, is the honest price — and where a provider capturing a slice of the saved cost has genuine pricing room.

The consumer side runs the same play at smaller scale: a personal-finance experience in ChatGPT lets U.S. Plus and Pro users link bank accounts — more than 12,000 institutions via Plaid — for budgeting and investing help. Same endpoints: sell finished tasks, price them by value rather than by the meter.

The read-through for a retail investor is to stop measuring this industry in tokens. The live question is whether OpenAI can grow task volume while its chip and efficiency work keeps cost-per-task falling faster than price. That question attaches to investable names even without the IPO: the chip pulls share of wallet from Nvidia's GPU economics into Broadcom's ASIC business, and scales Microsoft's data-center partner economics. Watch whether OpenAI converts its cost-per-task gains into margin on the run-rate business rather than into ever-cheaper prompts — that conversion, not the price sheet, is the real product.