Devin Crosses the Line, Perplexity Bleeds Cash: Two Roads Diverge in AI Commercialization Link to heading

2026-06-04 | marketer

Last week, two data points surfaced almost simultaneously. Placed side by side, they tell a more compelling story than any single report.

On the Devin side: in advanced teams’ validation patterns, developers launching 10–20 parallel Devin instances has become standard, with async agent sessions surpassing human interactive sessions for the first time. Cognition just closed a $1B+ funding round at a $26B valuation, with annualized revenue of $492M—just 9 months after its previous $10B valuation.

On the Perplexity side: 2024 ad revenue came in at $20,000—yes, twenty thousand dollars, not twenty million. That same year, third-party model and cloud service costs reached $57M, far exceeding total revenue of $34M. The ad model was officially abandoned in February 2026.

One line is breaking through the ceiling. The other is being crushed by it. This isn’t coincidence—it’s the structural divergence unfolding across the AI industry in 2026.

I. Perplexity’s Dilemma: AI Search’s Least Profitable Middle Layer Link to heading

Perplexity’s core value proposition has always been clear: help you find answers on the web, with cited sources. It doesn’t own a frontier base model, nor does it have Google’s search index and infrastructure scale. Its core experience is built on external models, search licenses, and cloud platforms.

At its essence, Perplexity is a Model Rentier—a product form that depends on leasing and reselling external model capabilities. It rents models from OpenAI, Anthropic, xAI, and others, bolts on the Bing Search API, and wraps it into an AI search interface anyone can use. The margin formula is brutally simple:

Revenue − (API costs + search licensing fees) = profit

The problem is on the right side of that equation. $34M in 2024 revenue against $57M in API/cloud costs—and that doesn’t even include Bing search licensing and infrastructure. A business that loses more money with every new user: in the traditional internet era this was called “burn cash for growth”; in the AI era it’s called “the model tax.” The difference: internet-era marginal cost trended toward zero, while AI’s marginal cost is a hard cost—every question asked, Perplexity pays the model provider a cut.

Facing this quandary, Perplexity’s strategy is to pivot toward a “Personal Computer” product—an agent that coordinates 19 different AI models and executes complex workflows in the background. Pricing started at the $200/month Max tier and later expanded to $20/month Pro users. The shift is from “help you search for answers” to “help you get work done.”

The direction is correct, but the path is punishing. Even though Computer can orchestrate up to 19 models, its key capabilities still come primarily from the external model ecosystem; Perplexity’s moat is closer to a scheduling and product layer rather than underlying model sovereignty. It’s like a moving company that builds no trucks and paves no roads—it just dispatches other people’s drivers.

Sacra estimates Perplexity reached $500M annualized revenue by April 2026, growing 335%. But if the API cost structure holds, that growth only accelerates the losses.

II. Devin’s Leap: From Tool to Production Pipeline Link to heading

Devin’s story is the inverse.

From day one, Cognition has been building its own model—SWE-1.6 proves it. More importantly, Devin doesn’t “help you write code”; it “writes the code and merges it to main.” This distinction means its business model is anchored in value creation, not information routing.

A few numbers tell the story:

• 7× growth in merged PRs: in under a year, Devin’s output of merged PRs grew sevenfold
• 16% → 80% code contribution rate: within Cognition’s own codebase, the proportion of Devin-generated/contributed code rose from 16% in December 2025 to 80% by May 2026
• Async sessions > interactive sessions: this is the critical threshold—AI is no longer just a human “copilot” but a self-driving production pipeline that can run independently
• 10–20 parallel Devin instances per developer (in advanced team validation patterns): this isn’t an assistive-tool usage model—this is a production-line management model

Cognition CPO Walden Yan, in an interview with latent.space, crystallized a key concept: the “Age of Async Agents.” You hand Devin a task, give it a repo, a machine, a shell, a browser, tests, and memory. It goes to work on its own. You don’t need to watch it. When you come back, the PR is already there.

The contrast with Perplexity’s predicament is stark:

III. What the Fork in the Road Is Really About: Who Owns Agency Link to heading

This contrast points to a deeper question: does AI product value come from the model, or from the agent?

Over the past two years, most have assumed “model as product”—whoever has the strongest language model wins. OpenAI is the poster child for this logic. Perplexity tried to carve out a niche within it: I don’t need the best model, I just make it easier for people to use all of them.

But 2026’s data is overturning this assumption.

Models are becoming swappable components. MiniMax M3 surpassed Claude Opus 4.7 on the Agentic Browsing benchmark. DeepSeek, Kimi, Qwen, and other open-source models keep closing the gap. Microsoft Build 2026 launched the MAI model matrix—7 new models covering different cost bands. When models become interchangeable, middlemen who “rent and resell” models get squeezed from both ends: model providers raise prices, users demand lower ones.

Agent capability is what can’t be swapped. Devin’s value isn’t in “which LLM it can call” but in its ability to autonomously complete an end-to-end loop: understand requirements → set up environments → write code → run tests → submit PRs → respond to reviews. This loop doesn’t hinge on model capability—it’s systems engineering: sandbox management, permission control, Git integration, CI/CD hookup, memory systems, multi-agent orchestration.

Perplexity’s Computer product tries to replicate that logic—simulating a single agent’s capability through 19-model coordination. But if the coordination layer itself is just other people’s API calls, then it remains fundamentally a scheduler, not an agent.

IV. The AI Product Watershed: Mid-2026 Link to heading

Taken together, these two cases reveal that AI product competition in 2026 is splitting into two tracks:

Track A: Model Rentiers

• Don’t own a frontier base model; core capability depends on external model ecosystem
• Value comes from “packaging and distributing” others’ capabilities
• Marginal cost scales linearly with usage
• Moat is brand and UX, not technical defensibility
• Representatives: Perplexity, most AI wrapper products

Track B: Agent Sovereigns

• In-house or deeply fine-tuned models; core capability under their own control
• Value comes from “independently completing work,” not “providing information”
• Inference/execution has variable cost, but high-value pricing and product stickiness make it easier to cover
• Moat is the systems-engineering barrier of model + toolchain + enterprise integration
• Representatives: Devin/Cognition, Cursor, Claude Code

Track A isn’t a dead end—if Perplexity can truly turn Computer into an agent rather than a multi-model router, it still has a shot at crossing over. But the window is narrow: when OpenAI, Anthropic, and Google are all launching their own agent products (deeply coupled to their own models), a “middle agent” that dispatches 19 third-party models via API has no advantage in either cost or experience.

Track B’s winning logic is equally clear: models are the ammunition; agents are the gun. What actually transforms how we work isn’t the type of bullet—it’s the precision, range, and reliability of the firearm.

The moment Devin’s async sessions surpassed human interactive sessions marks an industry inflection point: AI transitions from “you tell it what to do” to “you tell it the outcome, and it goes and achieves it.” The flip side of that inflection point is the brutal reality revealed by Perplexity’s financials: in this new world, merely “helping you find answers” is no longer enough—and it’s getting more expensive by the day.

The single lesson AI commercialization in 2026 teaches us: owning agency is the moat. Renting someone else’s agency is just a cost.

Sources: Latent.space “The Age of Async Agents” | Cognition $1B Series D announcement | Sacra Perplexity analysis | 數位時代 Perplexity financial deep-dive | The Decoder | VentureBeat | 證券時報 Perplexity financial data