“The cost of serving Grok Bot was unsustainably high relative to revenue.” With that explanation, Elon Musk’s SpaceX throttled usage limits on its AI agent over Halloween weekend, so that some users woke up to find their weekly allowance used up. The same week, Y Combinator CEO Garry Tan said he spends “about maybe $100,000 to $200,000 a year on tokens” and called it “a crazy privilege”. Gartner said open-weight models can handle 90% of agent workflow steps at as little as a twentieth of the cost of running a frontier model on every step. Google’s Mandiant reported criminals running up more than $100,000 a day on stolen AI keys. The thesis is that the agent era has a token bill, and it is being paid by providers, power users and, increasingly, by victims. The businesses that do well will route most work to cheaper models they control, price AI features to cover their cost, and guard their keys like cash.
Go deeper: who profits from the AI build-out, in The AI Gold Rush Pays the Sellers of Picks and Shovels, and the case for open weights, in ‘AI Is Going to Be Like Electricity’.
Providers can’t absorb it
Grok Bot users reported weekly usage jumping from as low as 50% to 100% overnight, TechCrunch reports. Musk confirmed it wasn’t a glitch. The cut comes even though SpaceX’s AI division turned a $431 million profit, a rarity in the industry. Agents are far more expensive to serve than chat, because each task involves planning, tool calls and retries. When providers price that below cost to win users, they eventually ration it, often suddenly.
Power users show the real cost
Tan’s example shows what heavy agent use costs. His agent has “read all of my email and all of my Dropbox… even my therapy notes going back 20 years”, and Y Combinator runs an internal agent with “the total context of every founder we’ve ever backed”, he said. He thinks personal agents will go mainstream once one problem is solved: “It’s too expensive right now.” He also worries that free agents from Big Tech would bring their own control problems.
Most steps don’t need the frontier
Gartner’s analysts argue the economics are already shifting. “Nine in 10 agent workflow steps are routine, and open-weight models handle them at parity for a quarter of the price of the cheapest closed model,” Mark Petty and Raymond Paquet wrote, The Register reports. Compared with running a frontier model on every step, the best case is a twentieth of the cost. Because open models can be more verbose and retry more, the realistic saving per task is “closer to five or six times”. They predict that by 2030, 90% of enterprise operational agent steps will run on open-weight models, with frontier models kept for research and hard reasoning. They also note that “every routine step sent to a frontier API moves a slice” of a company’s workflow history “outside your perimeter”, and the harness that routes work between models is where the value lies.
Stolen keys, stolen budgets
Expensive tokens attract thieves. Mandiant says “LLMJacking” charges can exceed $100,000 a day, with some organisations reaching $1 million before revoking access, The Register reports. One group it tracks as UNC6993 used a misconfigured service-account key to create more API keys and enable services, running up $100,000. Another intruder made more than 500,000 API calls in under eight hours to power “Jet Night”, a cut-price token service selling “unlimited” frontier access for about $60 a month that Google believes runs on stolen keys. Google expects criminals to “increasingly prioritize LLMJacking” over cryptojacking.
A token-cost playbook
- Route by difficulty. Use frontier models for planning and hard reasoning, and open-weight or smaller models for classification, extraction, tool calls and routine generation. Put a gateway in front that enforces this.
- Measure cost per outcome, such as per resolved ticket, completed order or merged change, rather than tokens consumed.
- Set hard caps at the provider and alerts on unusual spikes. A stolen key with no cap is an open cheque.
- Keep keys out of code and environment files, use a secrets manager, rotate regularly, and give service accounts only the permissions they need.
- If you sell AI features, price them by usage. Flat plans let a few heavy users, or one attacker, wipe out your margin. Prepaid credits align your revenue with your token bill.
Pricing AI features with VBWD
The last point is where many SaaS companies get caught: their AI bill is metered, but their pricing isn’t. VBWD, the self-hosted subscription and commerce platform we build, handles this with token bundles. Any feature can be sold as prepaid credit alongside a subscription, usage is tracked per customer, and access can be tied to plan entitlements, so expensive AI tools are only available on the plans that pay for them (per-feature licensing). Because it is self-hosted, your usage data and customer records stay in your own database. A worked example is in How to Build an AI Assistant SaaS With VBWD, and VBWD offers a free migration assessment.
The read
Agents are expensive to run, and someone always pays. This week it was Grok’s users, who were rationed, Garry Tan, who pays a fortune, and LLMJacking victims, who pay without knowing. The way out is to stop treating tokens as a fixed cost of doing AI and to manage them like any other resource: route most work to cheaper models, measure results, cap spending, protect the keys, and charge customers in a way that covers the bill.
Written for Red Robot with AI assistance and human editing. VBWD is our own product. Based on reporting by TechCrunch, TechCrunch, The Register and The Register.
One thought on “The Agent Era Has a Token Bill: Grok’s Throttle, Garry Tan’s $200,000 and the $100,000-a-Day Key Thieves”