You Know What’s Cool? A Billion Tokens.

In The Social Network, Sean Parker famously tells Mark Zuckerberg:

“A million dollars isn’t cool. You know what’s cool? A billion dollars.”

I’ve started thinking about AI token spend the same way.

A million tokens sounds like a lot until you begin using AI as an operating system for your work. Not merely as a chatbot you visit occasionally, but as a researcher, analyst, developer, editor, project manager, and increasingly capable digital colleague.

Then a million tokens becomes background noise.

Next thing you know you’re spending a billion tokens every month.

And lately, that is roughly what I’m using every month (in Codex alone).

But before we jump into details we need to address the elephant in the room: The current AI zeitgeist feels a little like a tokenpocalypse. GitHub Copilot has moved its plans toward usage-based billing, with credits consumed according to the API cost of input, output, and cached tokens. Companies that spent the past year encouraging employees to use as much AI as possible are starting to impose budgets, limits, and controls. Some reportedly blew through their annual token allocations before the year was half over.

Even Fable 5, Anthropic’s most capable publicly available model, was temporarily moved outside normal subscription access, remaining available through usage-based options while Anthropic managed capacity. The company has said it intends to restore broader subscription access when capacity allows, but the signal was hard to miss: even the companies building these models are still working out how to price our appetite for them.

Put it all together and it feels as though the era of abundant, flat-rate AI might be ending just as people are discovering what serious agentic work can do.

The fear is that flat-rate subscriptions were a temporary subsidy and that serious AI users will eventually face something resembling a second cloud bill.

But is the tokenpocalypse really a crisis or just the moment AI economics begin growing up?

Over the past five months, I’ve used approximately a billion tokens per month and these aren’t API calls buried inside a software product serving thousands of customers. They represent one person using AI to do real work.

I use agents to investigate unfamiliar codebases, review plans, research decisions, create documents, test software, challenge assumptions, and execute long-running projects. Some tasks run for hours. Others spawn specialized subagents that work in parallel.

This is important because most discussions about enterprise AI still assume a fairly shallow usage pattern:

1. An employee writes a prompt.

2. The model produces a response.

3. The employee copies the response into a document.

4. Everyone congratulates themselves on “adopting AI.”

That is not where this is going. The more interesting model is persistent, context-rich collaboration. An agent may read hundreds of files, inspect previous decisions, use tools, conduct research, revise its work, run tests, and preserve enough context to continue the project later.

That creates much more value than a single chat response.

It also consumes dramatically more tokens.

A token is not exactly a word. It is a small unit into which a model breaks text and other information. But the precise definition matters less than the emerging pattern: as AI becomes more useful, token consumption expands to meet it.

The question for business owners is no longer, “Can we afford a few prompts?” It is:

What happens when every productive knowledge worker begins consuming hundreds of millions, or billions, of tokens?

At first glance, the answer looks expensive.

Then you inspect the bill more closely.

The naive calculation is easy:

One billion tokens × frontier-model pricing = an alarming invoice.

But it is also wrong.

AI providers charge different rates for input, cached input, and output. My usage is unusually input-heavy: approximately 95.2% cached input, 4.4% new input, and 0.4% output. For a representative one-billion-token month, that works out to:

• Cached input: 952 million tokens

• Uncached input: 44 million tokens

• Output, including reasoning: 4 million tokens

This is not an apples-to-apples performance comparison. It is a comparison of what the same token volume would cost. Whether the cheaper model can do the same work must be tested task by task.

• GPT-5.5 — Frontier work: approximately $810 per billion tokens

• Gemini 3.1 Flash-Lite — Routine, high-volume work: approximately $20, plus cache storage

• GLM-5.2 API — Open-weight agentic work: approximately $326

These estimates use my observed mix of cached input, uncached input, and output. They exclude tool charges and other infrastructure costs. Long-context surcharges, data-residency requirements, and cache-storage policies can also move the final bill.

The range from roughly $20 to $810 shows why architecture matters. A company can consume the same nominal volume of intelligence and face radically different economics depending on which model receives the work, how much context is cached, and how often expensive reasoning is invoked.

GPT-5.5 is intended for complex professional and coding work. At current prices, my representative billion-token month would cost about $810. That is the expensive case and a reminder that sending every task to the strongest available model is like hiring your most expensive executive to rename files and update the CRM. OpenAI GPT-5.5 pricing

Use frontier models where difficult judgment, ambiguity, or the cost of a mistake justifies them. Using them indiscriminately is not quality control. It is expensive routing.

Eight hundred dollars per person per month would make most executives uncomfortable. It can also be understated for workloads with enormous contexts: OpenAI applies higher rates to GPT-5.5 prompts above 272,000 input tokens. The relevant question is whether the work is valuable enough to justify that premium.

Google designed Gemini 3.1 Flash-Lite for cost-sensitive, high-volume work. At my token mix, processing charges would be about $20 per billion tokens, although explicit cache storage adds cost based on the amount retained and its duration. Google Gemini API pricing

The dramatic price difference is useful only if the model performs reliably. Classification, retrieval, extraction, formatting, monitoring, and other reversible tasks are natural candidates but each workflow needs evaluation before the savings are treated as real.

These routine steps can represent most of an agent’s token consumption without representing most of its intellectual difficulty. Moving them to a smaller model preserves frontier capacity for the moments that require judgment while making the surrounding workflow dramatically less expensive.

GLM-5.2 is a large open-weight model designed for long-horizon coding and agentic work. Through Z.ai’s hosted API, the same token volume would cost about $326. Its downloadable weights also create options around privacy, control, customization, and infrastructure ownership. Official GLM-5.2 repository Z.ai GLM pricing

Open weights do not mean free inference. Self-hosting replaces a per-token bill with hardware, electricity, engineering, security, and capacity costs. The benefit is optionality—not an automatic discount.

That trade can become attractive at sufficient scale or when privacy and control matter more than the lowest headline price. For an individual user, however, buying or renting the infrastructure required to run a model this large will not automatically beat a hosted API. Open models change who can supply the intelligence; they do not eliminate the cost of producing it.

Here is an intentionally simplified routing scenario (not a recommended universal allocation). Instead of starting with arbitrary percentages, sort the work by consequence and reversibility:

• Routine and reversible: Use the smallest reliable model. If the result is wrong, it is inexpensive to detect and redo.

• Important but verifiable: Use a capable mid-tier or open model, then validate the result with tests, rules, or human review.

• Consequential and difficult: Escalate to frontier intelligence when ambiguity is high and mistakes are expensive.

These tiers are not permanent labels. A workflow should move down the cost curve as evaluation improves: once an important task becomes predictable and verifiable, it may no longer require frontier treatment. Conversely, new or unusually consequential work should escalate until the organization has evidence that a cheaper route is safe.

The sub-$100 future does not depend on frontier models suddenly becoming free. It depends on keeping frontier work a small, intentional share of the total and proving that cheaper models can reliably handle everything else.

Tokens are easy to measure, so we are tempted to treat them as the fundamental unit of AI cost. They aren’t.

The more useful unit is:

Cost per successfully completed unit of valuable work.

This is the crux of the entire AI ROI narrative today:

This does not mean businesses should ignore consumption. The operating discipline fits into three imperatives:

• Route: Use the least expensive model that reliably completes the task.

• Cache: Avoid paying repeatedly to process unchanged context.

• Measure: Track cost per successful outcome, not tokens consumed.

Budgets, context compaction, evaluations, and escalation rules all support one of those three principles. Model selection should be based on evidence, not brand hierarchy.

The answer is no longer “maybe someday.”

Under an aggressively routed workload, the raw token economics are already approaching that threshold.

Getting the same overall performance for less than $100 is the harder part. But three forces are moving in the right direction.

First, smaller models keep getting better. Work that required a frontier model a year ago can increasingly be handled by faster, cheaper systems.

Second, open-source models such as GLM-5.2 are narrowing the gap on coding, reasoning, and long-horizon agentic work. They also give businesses the ability to control their own infrastructure, customize deployments, and negotiate against proprietary providers.

Third, AI systems are becoming better at routing. A lightweight model can organize the work, retrieve context, perform routine operations, and escalate only the difficult decisions.

The future operating model may look like this:

• A small, inexpensive model handles the bulk of activity.

• An open-source model runs private or specialized workflows.

• A frontier model handles the consequential minority of decisions.

• Humans remain responsible for goals, judgment, and accountability.

Under that architecture, a billion-token month for less than $100 is not a prediction that every model becomes free.

It is a prediction that we stop using the most expensive intelligence for every step.

And that may be the bigger lesson for executives.

The future cost of AI will not be determined solely by how much model providers charge. It will be determined by how intelligently businesses architect their work.

The tokenpocalypse is real if every token is sent to the most expensive model and nobody measures what comes back.

But with caching, routing, and increasingly capable open models, a billion tokens may become less expensive than one mediocre software subscription.

A billion tokens is cool.

But you know what’s cooler?

A billion tokens producing measurable business value for less than $100 a month.