Sapiom Raises $35M As AI Agent Costs Face First Hard Audit

Sapiom raised $35 million to lower the operating cost of AI agents, TNW reported on August 5, 2026, giving enterprise buyers a clearer sign that agent deployment is moving from model access to spending control.

The San Francisco startup sits between an agent and the models, tools or services it can call.

Its Router is designed to choose the cheapest capable option before a task runs, while the platform enforces a budget rather than leaving each agent action to default to a more expensive frontier model.

Routing Layer Targets Token Spend

Sapiom's routing layer has handled more than 270 million transactions in its first six months, and A Semafor case in which customer Polsia reduced a monthly Anthropic token bill from $1.2 million to about $100,000 after Sapiom ran evaluations.

Polsia's projected revenue had risen from $100,000 to $10 million in a year, making inference cost a constraint on the agent-based business model rather than a back-office line item.

Founder Ilan Zerbib framed the cost curve as unsustainable for startups that want to deploy agents at scale.

The contrast is commercially important because Anthropic also backed Sapiom, joining Okta Ventures, Menlo Ventures and Array Ventures in the Series A after Dragonfly led the round.

That investor mix makes the deal less like a simple AI tooling announcement and more like a hedge around model economics.

A model provider can still benefit if cheaper routing lets customers run more agent workflows, but the customer decision shifts from choosing the most capable model by default to deciding which model is sufficient for each step.

Agent Projects Meet Budget Controls

The funding arrives as corporate AI budgets face closer scrutiny.

Gartner forecasts that the end of 2027 could bring cancellations for more than 40% of agentic AI projects, with escalating costs among the leading reasons.

That figure gives Sapiom's product a specific market problem: agents may work technically but still fail commercially if every action carries frontier-model pricing.

Sapiom has raised $50 million in total, including a $15 million seed round six months earlier that Accel led.

The Series A therefore tests whether cost routing can become a control layer in the agent stack, not just an optimization feature for early adopters.

Enterprises evaluating AI agents now need to measure task routing, model sufficiency and spending limits alongside accuracy and workflow coverage.

If agent adoption depends on reducing waste before work begins, the orchestration layer becomes part of deployment governance rather than an optional procurement add-on.