AI Agents Are Passing Death Sentences on 120 Major APIs

Source: 20VC with Harry Stebbings | Published: 2026-04-30T14:00:26Z

Jason Calacanis had three leading models evaluate 120 major APIs. Stripe earned the only A+, while Marketo and others got death sentences — the agents believe they can replace them entirely.

OpenAI Lost Ground Six Months Ago — It's Only Showing Up on the Books Now

OpenAI missed expectations on both user growth and revenue. CoreWeave and Oracle dropped 5% and 7%, respectively. But if you've been following the model race, none of this is surprising. OpenAI visibly fell behind on model capabilities in the second half of last year, and Anthropic grabbed significant market share — even Elon Musk publicly admitted Anthropic "has something" in coding. The irony: just as Wall Street digests this stale news, the winds in Silicon Valley have already reversed. Twitter is flooded with complaints about Claude buckling under user load, while OpenAI's newly released Codex model o4-mini is putting up strong numbers on coding benchmarks.

This is a shoe that took three or four months to finally drop.

In the Future, Agents Pick the Model — Not Humans

A deeper shift is underway: as more workflows are executed by AI agents, the decision of which model to use migrates from humans to the agents themselves. Jason Calacanis made a blunt observation — his company's "AI VP of Marketing" and "AI VP of Customer Success" (both agent-based applications) prefer OpenAI's API. He personally might enjoy the superpower feeling of Claude Code, but his agents don't care about vibes.

"Just like in 2024 you had to support your human team, today I have to support my agent team. They picked OpenAI, so that's where I stand."

This means the criteria for model competition are shifting. Human preferences like "feels good" and "nice to interact with" are giving way to API reliability, toolchain compatibility, and cost efficiency. For OpenAI, this is actually bullish — in an agent-dominated world, it becomes competitive again.

Which Software Agents Bypass Determines Who Lives and Who Dies

If agents are making the calls, who do they pick? Jason ran an experiment: he had Claude, OpenAI, and Gemini jointly evaluate 120 major APIs and score them. Stripe got the only A+. Marketo, Outreach, and SalesLoft received death sentences — agents deemed these products worthless because they could write better emails and send them directly.

Agent preferences map onto a clear 2×2 matrix: market leader + continuous innovator. They won't recommend outdated incumbents, and they won't take risks on tools nobody's heard of. This is bad news for products like Jira and Confluence — agents don't need project management tools because they are the project management. Public markets have already started pricing this in: Atlassian and Monday are under pressure, while infrastructure plays like Twilio and Cloudflare — things agents still need — are outperforming.

Canva May Have Just Built a Perfect "Last Year's Product"

Canva just dropped its 2.0, complete with a full agent suite and "vibe design" support. The product itself is impressive. But here's the sharp question: will agents use Canva? Almost certainly not. Agents don't need to open an interface to drag and drop assets — they generate assets directly.

There's a subtle timing gap here. Individual users and small teams will continue to get enormous value from Canva — a solopreneur paying $18/month for Canva designs won't be replaced by agents anytime soon, because nobody actually wants an agent to replace themselves. But enterprise is a different story. Companies are naturally incentivized to automate workflows with agents to cut costs. So if Canva IPOs in 2028, it'll probably succeed — but it'll be priced as a "real company," not given the AI narrative premium.

Compute Equals Revenue? That Equation Is Starting to Crack

Anthropic landed Google's commitment of up to $40 billion ($10 billion in cash plus $30 billion tied to performance milestones), plus an additional $5 billion from Amazon, at a $350 billion valuation. The scale is staggering, but the logic is simple: Anthropic got the model right but couldn't keep up on compute; OpenAI had the compute but fumbled the model. Getting both right at this stage is nearly impossible.

What's truly sobering is the capital intensity math. Say you're running $10 billion in annualized revenue now and projecting $100 billion in two years. Every dollar of revenue requires roughly $4–5 in capex to support the underlying compute. That means you need about $300 billion over two years for chips, facilities, and data centers. Overestimate and you're sitting on $150 billion in idle capacity. Underestimate and you lose customers to compute shortages. And this bet has to be placed two years in advance.

"This makes running an airline look easy."

Sam Altman's famous line — "compute equals revenue" — is correlation at best, not causation. No compute, no revenue, sure. But compute plus a bad model also equals no revenue — just look at Grok.

China Blocks the Manus Acquisition: The Money's Already Been Distributed, and It's Not Coming Back

Chinese regulators blocked Meta's $2 billion acquisition of Singapore-based Manus. But investors like Benchmark have already distributed the proceeds. Jason's take is straightforward: people who received their carry aren't giving it back, regardless of what Beijing says.

The practical impact plays out on two levels. First, leverage over Meta — if Meta has significant operations in China, Beijing can apply pressure through local assets. Second, team members still in China face potential exit restrictions. But the most far-reaching impact is the signal: cross-border AI talent acquisitions like this simply won't happen anymore. Unless you put the entire team on a plane to Singapore the night before wiring the money.

This is the latest flashpoint in the US-China AI rivalry, following chip sanctions. From China's perspective, the logic is perfectly symmetrical: you won't sell us chips, so why should we let you have our researchers?

Medallia Goes to Zero: Overpaid and Overleveraged

Thoma Bravo handed Medallia to creditors, evaporating $5.1 billion in equity. This customer satisfaction survey company generates roughly $1 billion in annual revenue with a few hundred million in EBITDA. When Thoma Bravo acquired it in 2021 for $6.4 billion, roughly $3 billion was debt and $3.4 billion was equity — leverage at nearly 47%, which was aggressive.

"Overpaid, overleveraged. You can't service $3 billion in debt on a $1 billion revenue, low-growth company that still needs to transition out of the pre-AI era."

This is the second major wipeout in PE after Pluralsight. What's more unsettling is that Medallia's customer retention has been steadily deteriorating. When CIOs make budget cuts, "nice-to-have" products like this are first on the chopping block. A wave of AI-native customer sentiment analysis tools already exists — better and cheaper. Medallia needs a complete product rewrite, and that's simply impossible under a crushing debt load.

Exit Paths Are Narrowing, but Super-Exits Will Cover Everything

The PE acquisition exit path is shrinking. If AI calls into question the durability of traditional B2B software, then the classic PE model of layering debt on top of equity becomes even more fragile. IPO thresholds are also rising fast — $1 billion in revenue growing at 40% might be the new floor. A few years ago, $400 million in revenue growing at 30% could get you public. Today, IPOs at that level are dead on arrival.

A micro-trend is emerging: founders who can't reach IPO scale and can't wait for an acquisition are starting to "give away" their companies to larger peers. Not failed accelerator projects being abandoned, but companies doing tens of millions — or even north of a hundred million — in revenue that see no endgame and choose to merge with another company, handing off the reins.

But from an LP's perspective, a counterintuitive possibility holds up perfectly well: ten super-exits are enough to cover the entire fund's return target, and the other 96 companies going to zero doesn't matter. This is exactly why top-tier funds are pushing to go bigger — if winner-take-all only has a handful of seats, you need to make sure you're in one.

Why YC Is Cracking Down on Fake ARR

Gary Tan publicly called on founders to report revenue numbers honestly, laying out five explicit guidelines. A legal tech company had previously exposed how widespread ARR inflation is at the seed stage — Jason himself has a portfolio company doing nine figures in revenue that sends him three different ARR numbers every month.

This move is both principled and strategic. YC holds roughly 25% market share in seed-stage investing. If the numbers across the market are broadly unreliable, the biggest loser is the market maker. Just as the NYSE would step in to clean house if it discovered listed companies were cooking the books, YC needs to protect trust in its own marketplace. Expect "YC revenue guideline compliant" to become standard shorthand in seed rounds.

The Simplest Way to Buy AI Stock: Don't Overthink It — Buy Nvidia

Google's market cap sits at $4 trillion, Nvidia at $5 trillion. If you can only pick one, the risk-adjusted choice is Google — it has multiple paths to winning. AI adoption accelerates, Google wins. AI adoption is gradual, Google still wins — provided ChatGPT doesn't erode Google Search, the cash cow. But if what you want is pure AI upside exposure, the answer is straightforward: Nvidia. You can't buy Anthropic or OpenAI stock. Nvidia is the closest thing to owning "AI itself" available in public markets today. Don't overthink it. Don't bother with derivative plays like CoreWeave.