01 · TWO PRICES
At home, AI looks like a buffet. At work, it is a taxi meter.
Open ChatGPT or Claude in a browser and the deal feels familiar: pay one monthly amount, then keep chatting. The provider quietly decides the limits, slows heavy users, routes easy questions to cheaper systems and absorbs whatever the subscription does not cover.
That is not the deal most software companies receive. Connect AI to a product, a document workflow or thousands of employees and the bill usually moves with usage. The provider counts small pieces of text called tokens. Reading costs money. Writing costs more. Searching, storing context and calling tools can add separate charges.
CONSUMER CHAT
One visible monthly price
- Usage is pooled across millions of people
- Limits and slower periods contain the heaviest users
- Investor money can subsidise adoption
ENTERPRISE API
A meter behind every workflow
- Every employee and customer can add consumption
- Long documents are read again across multiple turns
- Agents can keep working after a person looks away
02 · THE FRONTIER PREMIUM
There is no single price for “AI.” The smartest tier can cost hundreds of times more.
Model menus resemble airline cabins. A small model is cheap and often good enough. Move toward the frontier—the systems sold for the hardest research, coding and reasoning—and the output rate climbs sharply. Then a long-context surcharge or a faster service tier can lift it again.
Cost to generate one million tokens
The graph is not an argument to avoid frontier models. It is an argument to reserve them for work that needs them. Asking the most expensive system to format a routine email is like sending a partner at a law firm to buy printer paper.
PRIMARY SOURCEOpenAI: GPT-5.2 Pro output was listed at $168 per million tokens↗03 · EVERY TURN COSTS
A chatbot answers once. An agent can quietly buy the answer many times.
A useful agent does not simply reply. It plans, opens a file, searches, reads the result, corrects itself, calls another tool and finally writes the answer. Each step is another turn through the model. To remember what happened, the agent often sends much of the growing conversation back through the meter.
WORKED EXAMPLE · TRY THE CONTROLS
Reviewing one 80-page tender
Assume the first read is 55,000 tokens. Each turn adds notes and tool results. We use a frontier rate of $10 to read and $50 to write one million tokens. This is deliberately simple: no engineering jargon, just the meter.
ESTIMATED MODEL BILL
04 · THE PUBLIC RECEIPTS
The warning is no longer theoretical. Large companies are already reaching for the brake.
In 2026, “tokenmaxxing”—rewarding employees for using more AI—collided with finance. The pattern was remarkably consistent: encourage unrestricted use, watch consumption detach from useful output, then add budgets, caps or cheaper routing.
It later cut cost per session with caps, cheaper routing and visible running costs.
Axios · August 2026 ↗MMETAInternal token use was reportedly tracking toward billionsExecutives discussed treating tokens like any limited operating resource, with caps per engineer.
TechCrunch · July 2026 ↗AAMAZONScrapped a usage leaderboard after employees gamed itUnnecessary agent activity raised computing costs without creating useful work.
InfoWorld · May 2026 ↗MMICROSOFTCanceled most internal Claude Code licenses, partly over costEven a cloud giant chose to consolidate its developers around another tool.
Axios · May 2026 ↗AACCENTURECapped routine token use after one tool grew 113-foldThe professional-services group moved from maximizing consumption toward measuring value.
ITPro · September 2026 ↗PPALANTIRIts CEO said enterprise customers were tired of paying for little valueAlex Karp publicly attacked frontier labs for charging too much while customers looked for cheaper options.
Axios · July 2026 ↗FURTHER READINGAssociated Press: workplaces look for cheaper AI as tokenmaxxing fades↗Cheaper tokens did not create cheaper bills. They created permission to use more tokens.
THE ENTERPRISE PARADOX
05 · A CONTROLLED PIPELINE
The safest AI budget is money you have already decided to spend.
takes the stored-value route. Add Forever Credits once, use them across the models and document tools you need, and top up only when you choose. The balance goes down as work is completed; it does not turn into an unlimited month-end invoice.
Top up deliberately
Buy a defined amount of Forever Credits. They carry no monthly fee and do not expire.
Spend from one balance
Use one wallet across supported language models, document work and image generation.
Stop at zero
When the balance is too low, the workflow asks for a top-up instead of creating a surprise bill.
You choose the ceilingNo open meter
CONTROL IS A PRODUCT FEATURE
No API keys. No separate model subscriptions. No uncapped invoice.
A credit balance does not make expensive intelligence cheap. It makes the risk finite and visible—the two properties missing from most enterprise AI rollouts.
SOURCES
Sources and methodology
Editorial note: Public list prices change frequently. The chart records prices available in September 2026 and links to the providers. The worked example is an explanatory estimate, not a quote for any provider or a claim that every agent sends context in exactly the same way. Company examples are attributed to the publications that reported them.