Sapiom's $35M: your agent bill is a routing decision

Sapiom raised a $35 million Series A this week to do something unglamorous: pick a cheaper model for each call your AI agents make. Dragonfly led, with Accel, Gradient, Coinbase Ventures, Okta Ventures, Menlo Ventures and Anthropic participating. Total funding is $50 million, roughly a year after founding. The company is worth two minutes of your attention — not because you should buy it, but because the problem it sells against is the one that kills agent projects in month four.

What actually happened

Per Sapiom's own announcement, the platform sits between your agents and the models, choosing "the best model, compute, and tools for each task" at runtime instead of honoring whatever model name a developer hardcoded during a demo. It also enforces spend budgets before a run executes. Sapiom reports 270 million transactions processed and more than 100,000 agent runs a day since launching six months ago, and claims one customer cut inference costs by 75%.

Two details matter more than the round. First, TNW reports Sapiom runs open-weight models in its own San Jose data center and bills compute directly rather than marking up an upstream API — a different business than a pure broker. Second, TNW counts more than 80 companies doing routing, plus AWS Bedrock and Azure folding it in as a feature. Routing is commoditizing while it's being funded.

The pitch leans on Gartner's forecast that over 40% of agentic AI projects get cancelled by 2027, with cost as the leading reason. That's a vendor citing a research firm, so treat it as directional. The direction is right.

Why model routing matters for your business

Here's the failure we keep walking into. A team prototypes an agent on the best model available, it works, it ships, and nobody revisits the choice. Six months later the agent is classifying support tickets — a task a model costing 3% as much handles fine — and the invoice is a mystery nobody wants to open.

The fix is architectural and it's cheap: the model name is configuration, not code. One layer, one env var per workload, per-task budget caps, and logging of tokens and cost per run. Do that and you can swap providers in an afternoon when prices move — and they move constantly.

You may not need Sapiom. You do need the seam it plugs into. If your agent calls anthropic.messages.create from twelve files, you don't have a cost problem yet. You have a refactor you've been deferring.

Sources: Sapiom, TNW, FinSMEs.