Exclusive / Startup Sapiom routes clients’ AI to lowest-cost tokens

Reed Albergotti
Reed Albergotti
Tech Editor, Semafor
Aug 5, 2026, 12:19pm EDT
Technology
Sapiom founder Ilan Zerbib
Sapiom
PostEmailWhatsapp
Title icon

The Scoop

Quintessential San Francisco AI startup Polsia employs nobody and operates swarms of AI agents to help run other businesses. Its projected annual revenue skyrocketed almost overnight this year from $100,000 to $10 million, but so did its token spending, reaching a debilitating $1.2 million a month on Anthropic.

Its founder hired Sapiom, another San Francisco startup that specializes in streamlining token costs to run a series of evaluations, after which Sapiom reduced Polsia’s token-consumption tenfold, to about $100,000 per month, said Ilan Zerbib, Sapiom’s founder and CEO.

“It’s just unsustainable,” he said, adding that startups like Polsia can’t deploy their product, even if there’s lots of demand, at the costs that frontier-model providers like OpenAI and Anthropic are charging.

Sapiom is betting it can lower those token costs, and is raising a $35 million series A led by venture-fund Dragonfly’s Haseeb Qureshi. The company raised $15 million earlier this year in a seed round led by Accel. The company is also launching a new, single API key — a model router that automatically sends requests to the most efficient AI model for the job. The product puts it squarely in competition with OpenRouter, the biggest name in that field.

AD

Sapiom, which has a few thousand customers, is different from other routing firms because it serves open weight models from its own server racks, located in a data center in San Jose. Most competitors don’t have their own infrastructure and act more as a middleman between other providers.

Sapiom also doesn’t charge a premium over models it provides to customers. Instead, because it’s running the models on its own infrastructure, it can charge customers directly for the compute.

Title icon

Know More

The company’s fundraising comes as corporate AI budgets are getting their first real audit. Forrester predicted enterprises will postpone a quarter of planned AI spending to 2027, and a KPMG survey of more than 2,100 senior executives in June found only 7% of them could point to established returns stemming from using AI. Uber’s chief operating officer said in May that AI costs were proving “harder to justify” than expected.

AD

“In 95% of cases, it doesn’t make sense to go to a very expensive frontier model,” Zerbib said.

At the same time, Zerbib believes lower costs help frontier model companies in the long run, as lower costs will allow companies to build more agents, and some of that token spend will require the most powerful AI models.

Sapiom also helps customers connect their autonomous agents to the outside world and economy. For instance, it offers payment processing and SMS services.

While the biggest token customers, software developers, still prefer the most powerful AI models from companies like Anthropic and OpenAI, even bigger growth will come when AI agents that do simple tasks proliferate throughout the economy, said Zerbib: “There’s maybe 10 million developers — maybe 50 million,” he said. “If you think about agents, we’re talking about trillions of agents that will operate in the economy in the next three years.”

AD
Title icon

Reed’s view

Today, when companies tighten their belts, it almost always means cutting real people, with families and mortgages. In the future, layoffs might look more like AI agents being put on ice.

In that sense, Sapiom is like an outsourced labor contractor. Lots of companies have pitched themselves this way, but they’re missing an important part of the equation: companies don’t want to pay for overpriced AI labor.

The job of an AI labor provider isn’t to build up AI “employees” that can do everything, but to analyze in real time what work needs to get done, and then find the most cost-efficient solution for that particular job.

Big companies will be able to do this themselves, or hire expensive consulting firms to do it for them. Startups and small-to-medium-sized companies won’t. And with inference becoming one of the largest line items in a startup’s budget, this cost efficiency means more runway.

Title icon

Room for Disagreement

Routing may already be a commodity. Amazon and Microsoft now bundle intelligent prompt routing into Bedrock and Azure, open-source routers are free, and OpenRouter, with 25 trillion tokens a week flowing through it, offers customers its Auto Router plus Fusion, which combines cheap models to rival frontier performance. One market tracker counts 80 active routing competitors.

AD
AD