The harness: what makes a model usable

Here's the thing. You train a powerful model — then what? If all you hand people is an API, you're essentially waiting for everyone to rewrite their own applications just to plug you in. Most people won't.

What Anthropic got right was wrapping a layer around the model. Matt calls it the harness. It's not some abstract concept — it's four concrete things: Claude Code (which takes over the actual writing of code), MCP (a plug that pulls in your data from wherever it's scattered), Skills (customizable capabilities), and Cowork (a full team-collaboration layer). Together they turn the model from "a chat box" into a place where you can pipe in your data and actually get work done.

Matt says this isn't unique to AI. Go back to the SaaS wave: what separated companies wasn't "who can build a CRM" — anyone could do that. The gap came down to the last mile: whoever understood best how to get the product into customers' hands, and make it the easiest one to pick up. Same story on the consumer side — he points to Airbnb, built by people who leaned more design-first, because making something good enough to use gets the market growing that much faster.

If you don't give people a path to actually use the model — a way to use it — you're basically waiting for everyone to migrate every application they have, just to use this one API.Matt Murphy · Menlo Ventures

No revenue, no product — so why bet on it?

Menlo got in early. About three years ago, in a round that valued Anthropic at over $4 billion, with no revenue and no product yet shipped, they invested anyway. Matt says VCs almost never write a check at that price for a company with zero revenue. By the time Menlo led the Series D in 2024 — more than $500 million — Anthropic already had early revenue and had locked in both Google and Amazon. So what justified the money? He breaks the logic into three pieces.

Team, technology, and a mountain of capital — together they formed a wall for anyone trying to get in. Matt says some of the other players chasing foundation models back then got in early too, but they either faded out or fell behind.

Enterprises will run several models at once

One trend arrived earlier than most people expected: in 2026, enterprises and growth-stage AI startups are increasingly turning to open-source models to keep costs down. So will open source crowd out Claude?

Matt's take: open and closed source complement each other, they don't replace one another. Enterprises sit on huge amounts of private data, and training their own custom models — or using open source — is how they keep control of it. Meanwhile, they'll still hire Anthropic for other work.

How do you choose among all these models? That's exactly what gateways like OpenRouter do: routing different requests to different models by cost, latency, and reasoning strength — it'll even decide, within the Claude family, whether to use Sonnet or Opus. (OpenRouter is another of Matt's portfolio companies.)

But he pulls back a bit. The companies fine-tuning routing to this degree are the most sophisticated ones — the ones with researchers and the bandwidth to tinker. Most ordinary customers won't bother; they'll stay almost entirely reliant on something like Claude. In his view, a multi-model world actually pushes the incumbents to get stronger and better, which is good for everyone.

Why startups can grow this fast now

Going from zero to $10 million, then to nine figures — that pace used to make no sense, and now it does, and not just for one or two companies. Matt says in more than twenty-five years of investing, it's only in the last few years that he's watched companies go from zero to $100 million, zero to $300 million. He almost feels for founders now: it used to be that hitting $10 million in a year put you in the top 1%; now you look around and you're not even sure you register. So what's actually different about this wave compared to the last SaaS one?

The biggest difference: it's product-led, not sales-led.

That low barrier is exactly why the numbers stop making sense. He points to two companies in his own portfolio:

The legal market — in Matt's own words, "the worst enterprise software market in history" — used to send VCs running the moment they heard it. Now it's one of the hottest markets around, because the impact of this technology cuts deep enough that lawyers are diving in headfirst. Lovable, meanwhile, targets the other 99% who don't write code, putting the tools to build and run a business directly into their hands.

The upshot: fewer things stand between an idea and a product people actually want than ever before. If you have a good idea, the market is already there, ready to go — you're not waiting on a phone or a network to spread first.

Why safety became an advantage, not a drag

Matt says Anthropic bet on trust and safety from day one — that was part of Menlo's original investment thesis. It's paid off in the enterprise market, where that's exactly what customers care about most. But that path recently ran into a storm.

The following is background from media reporting, not the interview itself, included here to fill in the picture:

Some in the cybersecurity world think Anthropic's rollout of Mythos was more marketing than genuine protection: vulnerability scanners already existed, this one's just faster and better, but not to the point where it needed to be gated for the public's protection. Matt disagrees. In his view, releasing a capability that powerful with no restraint, and then watching a wave of hacks follow, would stain the whole industry far worse — and slow innovation down right as everyone's heading into round two or three of this boom. His phrase for it: "you have to go slow to go fast."

So how should this actually be governed? The interview brings up an idea floated by DeepMind's CEO: have the AI industry set up its own independent governance body, along the lines of FINRA in finance. Matt thinks that's not a bad idea — better to have people who actually understand the impact writing the rules than government officials, and better than the current patchwork of case-by-case federal decisions and states each setting their own rules, which is rough on startups.

Finally, Matt threw some cold water on himself. He admits the hype is still running ahead of reality — these tools aren't reliable enough yet. The next big market he's most excited about is on the consumer side: an AI that's actually on your side, one that negotiates a better price for you or fixes a mistake in your insurance claim, instead of the kind companies hand you to serve their own interests. Though he adds that this probably won't be the one he builds.

Handing people a bare API (an interface programs call into) is the same as asking everyone to rewrite their own applications just to use you — most won't bother. Anthropic wrapped a layer around the model, what Matt calls the harness: Claude Code, MCP (a plug that pulls in data scattered everywhere), Skills, and Cowork — four pieces that turn the model from a chat box into a place that can pull in data and actually get work done.

The lesson lands on the last mile: getting people in the door, connecting their data, making it easy to pick up — that's the real dividing line. Even back in the SaaS wave, the winners were the ones who understood best how to get the product into customers' hands.

Menlo got in early. About three years ago, in a round valuing Anthropic at over $4 billion with no revenue and no shipped product, they invested anyway — VCs rarely wrote checks at that price for a company with zero revenue. What justified it? Three pieces of logic.

Going from zero to eight figures, then to nine, at this pace used to make no sense — now it does, and not just for one company. The shift is that the product drives growth on its own: users and engineers just open it and start using it, with adoption barriers low enough to be everywhere. The last SaaS generation ran on stacking sales teams, capped by hiring and ramp time.

Even the legal world — what Matt calls "the worst enterprise software market in history" — has been swept in. He also expects open and closed source to coexist long-term: enterprises use open-source models (which they can self-host) to control their private data, while still hiring Anthropic for other work, routed through gateways like OpenRouter by cost and latency; companies without researchers to spare stay almost entirely on Claude. (Legora, Lovable, and OpenRouter are all in Matt's portfolio.)

Anthropic bet on trust and safety from the start — part of Menlo's original thesis, and exactly what enterprises value most. That path recently hit a storm; the following background comes from media reports, not the interview itself.

Some in the field think Anthropic's Mythos rollout leaned more on marketing than genuine protection. Matt's response: "you have to go slow to go fast" — releasing that much raw capability, and inviting a wave of attacks, would slow the whole industry down right as it enters round two or three of this boom. On governance, he backs DeepMind's CEO's idea of the AI industry building its own independent body, along the lines of FINRA (the US brokerage industry's self-regulator), with people who actually understand the field writing the rules.