Happen Bank, formerly LendingClub, is being cautious with its use of AI and token spend by putting in token caps on employees. 

“We were concerned about token costs before it was cool to be concerned,” Scott Sanborn, chief executive, told FinAi News.  

“We encouraged everyone to use AI and told them that cost is not a factor out of the gate that we’re overly worried about, but let’s put in place the controls to make sure we don’t have some of the headlines of, ‘Oh my god, I spent my annual token budget and it’s only April.’”

To manage that rising cost, the digital bank “put in caps per person, per day and per month,” Sanborn said. 

The bank has a dashboard that can be reviewed by executives and team leaders to see how many tokens employees are burning, he said, adding that an employee might be asked to justify tokens spent. 

The San Francisco-based bank also is in the early phase of rolling out dashboards that allow employees to see their own token spend, Sanborn said. 

“‘We think it’s great you’re experimenting, but you spent $2,000 on Friday. What are you up to?’” he said. 

“We don’t view it the same way as somebody going to an expensive business dinner and buying an expensive bottle of wine, but we want to make sure we understand and can start to engage early on,” Sanborn said.

If a token spend is justifiable, then the employee can continue with experimentation or the task at hand, Sanborn said, adding that team leads can add more tokens to the employees account. 

Each department has its own token caps. The engineering department has the largest, he said, adding that certain tasks such as code conversion use a lot of tokens. 

The company did not disclose exact token caps or dollar caps for AI usage. 

Picking models 

Another way of taming runaway token costs is picking the right model for the right task, Sanborn said.  

With the caps and usage dashboards, employees will be more cognizant about picking the right model and ask themselves, “Do I really need Opus or Fable to do this silly research question or could I go a few models back and spend one-fifth [of the tokens] and get the same result,” Sanborn said. 

Picking the right model for the right task is a practice that other FIs including Temenos, JP Morgan, RBC and Alkami already are doing, according to FinAi News’ prior reporting.  

Temenos, for example, has “a good handle on token utilization, which can get pretty expensive,” Sai Rangachari, chief product officer at the core banking provider, told FinAi News. “We have controls [on token usage] that we can put in place.” 

Temenos educates its employees on the best models to use for specific tasks, Rangachari said, adding, that the company also uses some open-source models to reduce token costs. 

Capping use 

Capping token usage tackles only part of the exposure, Alexander Polyakov, chief product officer at derivatives marketplace for AI token costs IFX Exchange, told FinAi News. 

Volume caps help finance teams forecast spend, but the per-token price moves whenever a lab reprices — and that repricing risk sits unhedged on the balance sheet, he told FinAi News. 

LLM labs reprice tokens mainly because their own costs and competitive position keep shifting. Newer, more efficient versions (better chips, optimized inference) let them lower prices. When a rival lab undercuts them on price or releases a stronger model, the company can decide to match the price, whether it’s higher or lower, Polyakov said.  

Pricing is also a lever for steering usage — labs cut prices to win developers and market share, or raise them once a model proves valuable enough that customers are locked in and less price-sensitive. On top of that, underlying compute costs (chip supply, energy, data center capacity) aren’t fixed, so the cost of serving a token today isn’t guaranteed to be the cost tomorrow, Polyakov said. 

“If we have the certain limit of usage and we use all the tokens that we forecasted, the repricing can significantly affect the P&L,” Polyakov said. 

The IFX LLM Price Index illustrates how far model costs have spread out.  

As of early July, the index priced Anthropic output tokens at $22.09 per million and OpenAI output tokens at $18.75 per million. At the same time, DeepSeek‘s output tokens cost $0.37 per million. 

If the cost of close-source models continues to rise or even remains flat, a vast majority of FIs will consider deploying open-source models to clamp down on token costs, Rangachari said. 

Register here for the FinAi Lending Summit, set for Oct. 7-8 in Las Vegas.