AI Costs Have Token Stage

he promise of AI to skyrocket productivity and profitability has driven companies across industries to encourage employees to incorporate AI into their daily work. But are companies ready to face the cost?

Since the beginning of the year, organizations adopting AI at scale have already confronted shockingly high expenditures. For example, Uber employees burned through their entire annual Claude Code budget by April.¹ Another company reportedly incurred half a billion dollars in Claude usage fees in a single month.² Now, some companies are scaling back their AI initiatives and reprioritizing human productivity in an effort to be more cost-effective.³

To mitigate rising AI costs, AI providers such as Anthropic are moving toward usage-based cost models, where companies pay based on how many AI “tokens” employees use. For example, asking ChatGPT to perform a simple arithmetic calculation may use up to 500 tokens. However, more complex tasks, such as generating a chunk of code, may use more than 10,000 tokens.⁴ Basically, the deeper AI has to “think,” the higher the cost. Most people would agree this is a fair pricing model, but large organizations with complex data and countless models can see costs add up quickly.

Excessive AI costs aren’t the only emerging problem. Chipotle’s publicly available web chatbot was recently hijacked by hackers who used its tokens to generate Python code.⁵ Although a seemingly harmless scenario, this signifies an emerging risk of third-party token misuse, as organizations must now consider how excessive AI costs may arise from parties other than their own employees. Additionally, this highlights the importance of AI governance and how critical it is to examine vulnerabilities that AI may introduce throughout an organization’s digital ecosystem.

Currently, token consumption hasn’t been explicitly addressed in cyber insurance policies. Insurers have the opportunity to set the stage and specifically define, endorse, or exclude token misuse coverage. Pricing AI insurance remains top of mind for the cyber insurance industry, and perhaps there is potential to use tokens as an exposure base. A hypothetical calculation is shown below:

Step 1: Identify exposure amount
Estimated usage: 50 million tokens per month for 12 months

Total exposure = 600 million tokens per 12 months

Step 2: Develop base rate
Selected base rate = $1.25 per million tokens

Considerations include expected loss from AI errors, legal and regulatory risk, and operational risk.

Step 3: Apply risk factors
Public chatbot = +40% (higher risk than internal-use model)

Higher-reasoning AI model = –25% (less error)

Strong risk controls = –20% (strong governance)

Healthcare industry = +30% (higher liability)

Step 4: Premium calculation
Base premium = 600M tokens × $1.25 / 1M = $750

Net factor = 1.40 × 0.75 × 0.80 × 1.30 = 1.092

Final premium:
$750 × 1.092 = $819/year