How will AI tools be priced in a post-tokenmaxxing world?

With astronomical computing bills stacking up, a spendthrift era of “tokenmaxxing” is giving way to a more frugal mindset among companies building AI systems.

Taking a cue from this trend, workflow software platform Pegasystems will soon roll out a new pricing model that eliminates token-based rates in favor of a flat charge per resolved case. The company is one of a wave of software-as-a-service (SaaS) providers rethinking per-seat and other traditional pricing models in favor of charging for actual results.

The shift comes as businesses worry that incentivizing AI adoption at all costs hasn’t necessarily led to higher ROI. Instead, they end up using costly, sophisticated AI for tasks that could be better served by cheaper, simpler tools, one analyst noted to Morning Brew. Meanwhile, SaaS vendors are facing ongoing pressure as the rise of AI agents and coding threatens their own value propositions.

“We’ve been watching our clients struggle with and come to terms with the fact that leveraging some of this agent technology, especially in high volumes of production, can lead to pretty rapidly escalating and unpredictable costs,” Pega CTO Don Schuerman told us. “We felt we had a really strong approach in our architecture that allowed us to keep those costs much more sustainable.”

Think first

That approach centers on shifting the heavy reasoning workload to the design phase of the agentic workflows, Schuerman said. After agents develop the workflows and humans evaluate them, cheaper and lighter-weight AI models simply follow the script. Schuerman said this makes tasks both more efficient and more predictable for high-stakes applications.

“What I don’t want the agent doing is re-reasoning the whole workflow, which is where the token costs actually get really expensive,” Schuerman said. “So I want to use most of the reasoning at design time, and I want the runtime to be deterministic when it can be, and use agentic reasoning for very small, specific tasks.”

The full AI toolbox

Liz Miller, VP and principal analyst at Constellation Research, said she expects other companies to start to think more judiciously about which AI model or tool makes the most sense for a given task. In some cases, that might mean weaving in more traditional deterministic machine learning algorithms of the kind that companies have been using for years.

“We can keep using those for certain actions and certain activities, and what we can do is layer those agentic workflows on top of that to get those larger things done,” Miller told us. “We’re going to start seeing those AI ops skills start to come in, where we start to decide where and when and how we are applying these different AI models and AI functions.”

Small language models, or slimmer open-weights models fine-tuned on specialized tasks, can also help reduce costs when applied in the right places, Prakhar Mehrotra, PayPal’s SVP and global head of AI, said at a recent Bloomberg Intelligence event in Manhattan.

“I can maybe take an open-source model, fine-tune it, and make it part of this broader agentic [flow]. So you’re not consuming the full tokens of a frontier model,” he said. “That’s a trade-off, and it’s problem-specific, but that’s what you’re trying to optimize, and I believe the [small language models] are part of this equation.”

The end of tokenmaxxing?

This sort of rationing runs counter to the internal campaigns that companies have been running to push employees toward AI, which sometimes lead workers to tap agents for even menial tasks to boost leaderboard rankings. Miller said that tokenmaxxing was a mentality of no-holds-barred experimentation driven by a fear that AI wouldn’t be adopted quickly enough.

“I always have to exhale when I hear [tokenmaxxing], because it is probably one of the most detrimental things to actually finding success in AI that we could have gone through, and I hope we’re starting to come out of [it],” Miller said. “I hope it was a brief phase that sounded really cool at the time.”

SaaS switch-up

Pega is also one of several SaaS companies debuting new outcome-based pricing options as agentic offerings and coding tools from big AI labs have roiled stock prices in recent months (Pega’s own stock is down 46% year to date). Companies like Zendesk and Intercom have started offering outcome-based pricing, and Intuit CTO Alex Balazs told us the financial software giant is exploring alternatives to subscription models, too.

But Pega’s design-phase strategy comes with risks, according to Futurum Group VP and Research Director Keith Kirkpatrick. Namely, it puts more pressure on ensuring that governance of the design phase is airtight.

“Pega’s design-phase reasoning could improve reliability by reducing runtime variability, but it also concentrates risk in up-front configuration and governance,” Kirkpatrick wrote in a recent research note. “Enterprises will scrutinize whether Pega can deliver consistent agent performance across complex, regulated workflows without introducing new bottlenecks or governance gaps. If Pega succeeds, it could set a new industry standard, but if reliability or transparency falters, buyers may retreat to more traditional, if costlier, models.”