Short version: On August 5, 2026, a San Francisco startup called Sapiom — barely 11 months old — raised a $35M Series A led by Dragonfly, with Anthropic, Coinbase Ventures, Accel and VanEck Ventures on the cap table. It sells the plumbing to move an AI agent from a slick demo to something that survives production: a model Router it says cuts agent runtime cost by up to 10x, plus a build-and-test Agent Studio and a managed Runtime. The founder takeaway: "get my agent to production" is now a funded market someone will sell you — which turns build-vs-buy on routing, metering, and settlement into a real decision instead of an afterthought. If you've watched a demo agent look brilliant and then fall apart at scale, see also what it actually costs to run a coding agent right now.
What happened#
Sapiom announced a $35 million Series A led by Dragonfly, with participation from Accel, Gradient, Coinbase Ventures, Operator Collective, Formus Capital and VanEck Ventures, alongside existing backers Okta Ventures, Menlo Ventures, Anthropic and Array Ventures (Sapiom blog, Business Wire via Morningstar).
The round lands roughly 11 months after the company was founded in 2025 and about six months after a $15M seed, bringing total funding to $50 million. Founder and CEO Ilan Zerbib previously worked at Shopify and at Earny (acquired). The company frames its mission with a deliberately large number — "power the next trillion agents" — but the actual product is narrower and more useful: infrastructure for the unglamorous middle of the agent lifecycle.
The pitch: agents die at scale on three things#
Sapiom's whole thesis is the demo-to-production gap. An agent that dazzles in a scripted demo tends to break the moment real users hit it — and it breaks in three predictable ways: cost, reliability, and control. Sapiom sells a piece for each.
- Router — an OpenAI-compatible endpoint. You point your existing agent at it instead of at a single model, and for each call it picks the most efficient allowed model based on task shape, context size, cost, quality, latency, reliability, availability and your policy, then meters the call. Because most calls in an agent loop don't need your most expensive model, Sapiom says this can cut runtime cost up to tenfold (The Next Web).
- Agent Studio — a local environment to build, test, inspect and deploy agents with the context of your existing codebase.
- Runtime — managed production infrastructure where agents execute, with access controls, failure recovery, step-level visibility, and guardrails.
There's a fourth thread worth flagging: Sapiom describes giving agents "trusted access to the API economy" by abstracting identity, wallets, policy, risk and settlement into one integration — the rails for agents that actually transact, not just chat. That explains the crypto-flavored cap table (Dragonfly, Coinbase Ventures, VanEck). By its own numbers, the platform has processed 270M+ transactions and runs 100,000+ agent runs per day since launching about six months ago (PYMNTS).
Why the Anthropic check is the interesting part#
A frontier lab investing in agent operations is the tell. Anthropic makes money when you burn tokens; backing a company whose headline feature is sending fewer calls to expensive models only makes sense if Anthropic believes the binding constraint on agent adoption isn't model quality — it's whether teams can operate agents without them melting a budget or failing silently. That's a bet on operations, not intelligence, and it lines up with where 2026's money has generally moved: down the stack, toward the picks and shovels rather than another chatbot.
What it means for founders#
You are the customer this round is aimed at. A few concrete reads:
- The gap is real, and now it's priced. If your agent works in a demo and stalls in production, you are not doing it wrong — you're hitting the exact problem a $35M round just formed around. Treat cost, reliability, and control as first-class engineering, not launch-week cleanup.
- Routing is the cheapest win, and it's drop-in. The single biggest lever on agent cost is not calling your best model for every trivial step. Sapiom's Router is one way to buy that; rolling your own is another. If you'd rather build it, we walked through the pattern in how to build a model escalation ladder and how to cap per-user LLM cost before bill shock.
- But a router sees everything. An OpenAI-compatible proxy is trivial to try and hard to fully trust — it reads every prompt and response, and it becomes a dependency in your critical path. Pilot it on non-sensitive traffic, confirm the metering matches your own numbers, and keep the OpenAI-compatible shape so you can rip it out. Portability is the whole point of that interface; don't trade it away.
- "Agents that transact" is arriving. The wallet/identity/settlement layer signals that agent commerce — agents paying for APIs and services autonomously — is moving from thesis to product. If your roadmap includes an agent that spends money on a user's behalf, the questions of who authorizes it, what it's allowed to buy, and how it settles are about to become table-stakes design decisions, not edge cases.
The honest caveat#
The eye-catching numbers — up to 10x cost reduction, 270M+ transactions, 100,000+ runs/day — are the company's own, published to announce a raise. They're plausible for a routing-plus-runtime platform, but they're not independently audited, and "up to 10x" is a ceiling, not an average. The signal you can bank is the category, not the vendor: multiple serious investors, including a frontier lab, just agreed that operating agents in production is a distinct, expensive problem worth its own layer of the stack. Whether you buy that layer or build it, the demo-to-production gap is now the part of your roadmap that decides if your agent is a toy or a business.
