# Sapiom raises $35 million to cut the cost of running AI agents > Source: > Published: 2026-08-05 19:05:50+00:00 [Ilan Zerbib (@i_zerbib)](https://x.com/i_zerbib) said Wednesday that [Sapiom](https://www.sapiom.ai/) raised a $35 million Series A led by [Dragonfly](https://www.dragonfly.xyz/), bringing the San Francisco AI infrastructure startup's total funding to $50 million less than a year after its founding. [https://x.com/i_zerbib/status/2085064722060398872](https://x.com/i_zerbib/status/2085064722060398872) The financing, detailed in [Zerbib's announcement on X](https://x.com/i_zerbib/status/2085064722060398872) and a [company post](https://www.sapiom.ai/resources/blog/the-next-trillion-agents/), includes participation from [Accel](https://www.accel.com/), [Gradient Ventures](https://gradient.com/), [Coinbase Ventures](https://www.coinbase.com/ventures), [Operator Collective](https://www.operatorcollective.com/), [Formus Capital](https://formuscap.com/) and VanEck Ventures. Existing backers Okta Ventures, Menlo Ventures, Anthropic and Array Ventures also participated. Dragonfly managing partner Haseeb Qureshi is joining Sapiom's board. Sapiom did not publish a valuation. ([sapiom.ai](https://www.sapiom.ai/resources/blog/the-next-trillion-agents/)) The round follows a $15.75 million seed announced on February 6th and led by Accel, with participation from Anthropic, Vercel Ventures, Okta Ventures and other investors. Sapiom's compressed fundraising timeline reflects the scope of Zerbib's bet: the startup began as financial infrastructure that lets software agents pay for APIs and other services, then expanded into the execution system that governs how those agents select models, use tools and recover from failures. ([sapiom.ai](https://www.sapiom.ai/)) Zerbib previously co-founded automated-refund service Earny, which was acquired in 2021, before spending nearly five years at Shopify. He led payments engineering there and worked on Shop Pay and Shop Cash. That background shaped Sapiom's original focus on giving autonomous software controlled access to paid services without handing an agent unrestricted credentials or forcing a person to approve every transaction. ([okta.com](https://www.okta.com/en-ca/blog/customers-and-partners/founders-in-focus-ilan-zerbib-of-sapiom/)) ### From agent payments to agent operations Sapiom launched three products alongside the round. [Router](https://www.sapiom.ai/products/router/) provides an OpenAI-compatible endpoint that evaluates each request against cost, quality, latency, availability and company policy, then selects an allowed model. [Agent Studio](https://www.sapiom.ai/products/studio/) turns a stated business objective into inspectable TypeScript and lets developers move among a visual canvas, individual steps and source code. [Runtime](https://www.sapiom.ai/products/runtime/) manages production agents, including retries, schedules, pauses, signals and step-level execution records. ([sapiom.ai](https://www.sapiom.ai/products/router/)) Router gives Sapiom its clearest entry into existing AI budgets. A company can replace its model provider's base URL and API key rather than rebuilding an agent, allowing Sapiom to sit between the application and each inference request. Zerbib said Sapiom chooses less expensive models for routine work while reserving frontier systems for tasks that require them. That puts Sapiom in direct competition with OpenRouter and other model-routing services. Sapiom's distinction is infrastructure ownership: [Semafor reported](https://www.semafor.com/article/08/05/2026/startup-sapiom-routes-clients-ai-to-lowest-cost-tokens) that Sapiom serves open-weight models from its own hardware in a San Jose data center. Sapiom can therefore bill for the underlying compute rather than adding a premium to another provider's model price. ([semafor.com](https://www.semafor.com/article/08/05/2026/startup-sapiom-routes-clients-ai-to-lowest-cost-tokens)) The arrangement also creates a broader commercial position. Sapiom can meter the models, searches, browser sessions, messages and other services an agent uses, while tying each charge to a specific run and trace. Sapiom becomes the billing and policy layer across an agent's work rather than a vendor for one isolated capability. ### Cost is the opening Zerbib said Sapiom has processed more than 270 million transactions and handles over 100,000 agent runs a day, figures Sapiom reported for the six months since launch. The startup says its routing has cut inference costs by as much as 75%, although it has not released the evaluations or customer-level data needed to assess that result across workloads. ([sapiom.ai](https://www.sapiom.ai/resources/blog/the-next-trillion-agents/)) One disclosed case shows the size of the bills Sapiom is targeting. Semafor reported that Polsia, which operates swarms of agents, had reached about $1.2 million in monthly Anthropic spending. Zerbib said Sapiom's evaluations and routing reduced that expense to roughly $100,000 a month. The before-and-after figures come from Sapiom and Polsia rather than audited financial reporting. ([semafor.com](https://www.semafor.com/article/08/05/2026/startup-sapiom-routes-clients-ai-to-lowest-cost-tokens)) Sapiom's [published pricing](https://www.sapiom.ai/pricing/) starts with a free developer plan. Accounts created by August 31st receive 50 runs per day, with additional runs priced at $1. A $200 monthly plan includes 2,000 runs and charges 50 cents for each additional run. Production capabilities, including model access and web services, are metered separately when used. ([sapiom.ai](https://www.sapiom.ai/pricing/)) The financing gives Zerbib capital to buy infrastructure and push Sapiom deeper into the agent stack while enterprises scrutinize AI bills. The company is betting that developers will accept another control layer in exchange for lower inference costs, one account for external services and a record connecting every agent action to its price. If that exchange holds, Sapiom will collect revenue each time an agent reasons, acts or spends.