AI models are getting more capable and, in many cases, cheaper per token. VentureBeat reports that customers are still running into a less friendly arithmetic problem: agentic workloads can multiply token use so quickly that a small request becomes a recurring software bill with teeth.
The shift is easy to miss if the benchmark is a chatbot answering a standalone question. A short exchange may use a few hundred to a couple thousand tokens, according to the report. A query that requires web research may reach several thousand. The expensive part begins when companies ask an AI system to act as an agent across several tools, such as spreadsheets, a billing system and a CRM, to produce a report or update a dashboard.
That kind of task is not one model call. VentureBeat describes a chain of steps: reading files, checking records, searching for context, doing intermediate calculations and assembling the result. Each step consumes tokens. If the agent needs retries or handles messy data, the meter keeps running.
Token amplification is the trap
The mechanism is called token amplification. The model does not keep human-style memory between turns. To continue a conversation or complete a staged task, it is often given the prior messages, its earlier outputs and uploaded material again as context. As the context grows, later calls cost more than earlier ones.
That means a task that looks like one plain-language request can require hundreds of thousands, or even millions, of tokens across the full run. VentureBeat’s example puts one moderately sized job at about 280,000 tokens, roughly $1 at current prices. Scheduled every 15 minutes for a dashboard, the same job reaches $96 a day, or $2,880 a month. If a more complex report costs $10 per run, that rises to $28,800 a month for one departmental workflow.
Those figures are estimates, and they depend on the model, pricing, data quality and implementation. The direction of the problem is less ambiguous: more steps, more context and more retries mean more tokens.
Billing moved to follow usage
The report says Anthropic, Microsoft, OpenAI and many other AI providers moved major products toward usage-based billing in April and tightened token allowances on flat-rate plans. Tom’s Hardware has reported sticker shock among GitHub Copilot users after Microsoft’s pricing changes, including customer reports of increases as high as 100-fold.
For companies running many agents, token spending has become an engineering issue and a finance issue. VentureBeat says Uber, Microsoft, Amazon and Walmart are among firms that have moved to curb AI costs. The blunt lesson is that prompt design, context management and agent-loop boundaries now affect margins, not just model quality.
AI vendors and customers have technical options. VentureBeat points to prompt caching, routing work to cheaper models, batch processing, semantic caching and tighter context-window management as ways to reduce token consumption. Those techniques can cut meaningful percentages from a bill.
The harder problem is demand. As models improve, companies ask them to do longer and more complicated work. DeepSeek’s V4 Pro and V4 Flash have reportedly undercut comparable Western models by as much as 17 times and 25 times, respectively, according to VentureBeat, but cheaper tokens do not automatically fix workflows that expand by orders of magnitude. The AI economy still has a meter, and agents are very good at leaving it running.
This story draws on original reporting from Tom's Hardware.
