The Era of Unlimited Tokens Is Over: Cloud AI Costs Are Out of Control, How Can Regular Users Afford AI?

📖 Glossary

AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.

Abstract: Citigroup has banned Claude Opus and GPT-5.5 flagship models, Adobe has cancelled Claude's unlimited usage agreement, and Meta's ~6,000 employees consumed 73.7 trillion tokens in 30 days—enterprise cloud AI costs are spiraling out of control. When flagship models consume far more tokens per interaction than standard models, usage-based pricing is hitting everyone hard. Products like Kaihe AIBOX, a local AI agent computer, offer a new approach: hardware you buy once, with pay-as-you-go cloud API calls you control—no more panicking over monthly token bills.

On July 2, news broke across the tech world: an internal Citigroup email revealed the bank had completely shut off employee access to Claude Opus 4.6, 4.7, and GPT-5.5 flagship AI models. The reason wasn't security—it was cost.

This isn't an isolated case. Adobe cancelled Claude's unlimited usage agreement, and Meta's roughly 6,000 employees consumed 73.7 trillion tokens in 30 days, with 2026 internal AI spending projected to reach billions of dollars.

The era of unlimited tokens is over.

What Happened: Enterprise AI Bills Are Out of Control

Let's lay out the key facts.

Citigroup: Internal emails show that flagship models consume far more AI credits per interaction than standard models, driving the surge in usage. Citi disabled Claude Opus 4.6, 4.7, and GPT-5.5 on June 24, originally planning to restore access on July 1—but hasn't.

Adobe: Cancelled Claude's unlimited usage agreement, switching to usage-based limits. Employees can no longer freely invoke flagship models for tasks.

Meta: ~6,000 employees consumed 73.7 trillion tokens in 30 days, up from 60.2 trillion. One employee built an internal leaderboard to compare AI usage—the "champion" burned through about $1.4 million in 30 days. CTO Bosworth wrote in an April memo: "No one should spend more on AI usage than their own salary."

Enterprise AI cost失控 trend analysis

These cases all point to the same problem: cloud AI's usage-based pricing model is forcing enterprises to recalculate.

Why It's Out of Control: Three Traps of the Token Economy

Many people think using AI is just chatting—how expensive can it be? It's more complicated than that.

Trap 1: Flagship and budget model token prices differ by 25x. Claude Opus 4.6 output costs $75 per million tokens; Gemini 3 Flash output costs $3 per million tokens. Same "ask a question, get an answer" interaction, wildly different costs. But employees don't care which model they're using—they just want the best answer.

Trap 2: Platforms like GitHub switched from fixed annual fees to usage-based pricing. Enterprises used to buy annual plans with unlimited access. Now they pay per actual call. More calls, bigger bills. Citi's direct reason for banning flagship models was GitHub's pricing change.

Trap 3: No usage controls. Meta's case is the most blatant—among 6,000 employees, someone built a leaderboard to compete on token usage. No quotas, no real-time monitoring, costs spiraling out of control. One company's OpenAI bill went from $2,000 to $15,000 per month, only to discover an intern had been running model fine-tuning experiments with the company's API key for three weeks.

The Regular User's Dilemma: Token Bills Aren't Just an Enterprise Problem

Enterprise cost overruns seem distant from everyday users, but the same logic is trickling down to consumers.

Cloud AI tools are tightening free tiers and raising paywalls. What used to be free GPT-4 access now requires a Plus subscription; what used to be unlimited Claude conversations now limits free users to a few rounds per day. Flagship model token costs ultimately get passed to regular users through subscriptions, per-call fees, and tier restrictions.

The harder truth: regular users don't even know how many tokens they've consumed. You chat with AI all afternoon, write a few articles, create several proposals—it seems free, but token consumption is silently accumulating on the server side. When the free quota runs out, you either pay or get downgraded to a dumber model.

Token cost transmission from enterprises to individual users

This is why products like Kaihe AIBOX are gaining attention. It's a local agent computer that runs 24/7 at home, with OpenClaw agent management system pre-installed. Its approach differs from cloud AI: you buy the hardware once, agents run locally, and daily task scheduling and data processing happen on-device. Cloud API calls only happen when you need large model capabilities—and you control the usage.

That doesn't mean local AI devices have zero API costs—the edge-cloud collaboration architecture means cloud model calls still cost money. But at least you choose which model to use, when to use it, and how much—rather than being locked into a monthly subscription or held hostage by token bills.

Calculating the Cost: Three Ways to Use AI

For regular users, there are currently three main approaches:

Option 1: Cloud AI subscriptions. A monthly fee of tens to hundreds of dollars for flagship model access. Pros: latest, smartest models. Cons: limited free tiers, paywalls beyond quota. Your conversation data lives on someone else's servers.

Option 2: Direct API calls. Pay per token—use what you pay for. Good for developers, bad for regular users who don't know how many tokens a conversation costs or what the monthly bill will be. Citi and Meta employees both experienced cost blowups this way.

Option 3: Local deployment. Like Kaihe AIBOX—hardware you buy once, agents running locally, cloud API calls only when needed. Direct AI through WeChat voice messages, no computer required. You control which model, how much, and when—no "mystery bill" surprises.

No single approach is universally cheapest—it depends on usage volume. Light users can stick with cloud subscriptions; heavy users and enterprises need to do the math. But the trend is clear: as AI usage becomes more frequent, usage-based pricing will eventually make everyone start watching "how many tokens I burned today."

Trend Analysis: AI Cost Anxiety Is Just Beginning

The Citi, Adobe, and Meta cases are just the start. As more enterprises bake AI into workflows and more people use AI daily, token cost issues will only become more widespread.

In response, the industry is seeing several shifts:

Enterprise side: Companies are deploying AI gateways—middleware between applications and large model APIs—for authentication, rate limiting, cost allocation, and audit logging. Who used what, which project consumed the most—fully transparent.

Consumer side: More people are seeking "cost-controllable" AI usage methods. Not chasing the smartest model, but finding "good enough and affordable" solutions. Local AI devices, lightweight models, on-demand calls are becoming new options.

Model side: Big tech is also pushing down token prices. Gemini 3 Flash at $3 per million tokens is already cheap, and Chinese models are even cheaper. But flagship models, given their high training costs, are unlikely to see major price cuts anytime soon.

The era of unlimited tokens is indeed over, but that's not necessarily bad. When users start paying attention to costs, the market will be forced to produce more efficient solutions—more precise model selection, more reasonable quota mechanisms, more transparent pricing. For regular users, controllable costs are the only sustainable costs.

The Era of Unlimited Tokens Is Over: Cloud AI Costs Are Out of Control, How Can Regular Users Afford AI?