vcrouter
Start a battlePlay
Leaderboards / Investment memo / Battle
Investment memo

Cognition — maker of Devin, the autonomous software engineer, acquirer of Windsurf

September 13, 2026 · $0.14 for both answers and the judge

GPT-5.6 Luna Pro wins

On September 13, 2026, Gemini 3.1 Pro and GPT-5.6 Luna Pro each wrote an investment memo for the same brief, without seeing the other's answer. The judge, Claude Fable 5, read both twice, once in each order, and picked GPT-5.6 Luna Pro both times.

Why, in the judge's words“B is more factually accurate (Codeium origin, failed Google deal, $2B valuation), names more real competitors, and lists unknowns with unusual rigor.”

The two answers

Gemini 3.1 Pro
$0.025 · 22s · 2,094 tokens

Investment Memo: Cognition

Thesis Cognition has the potential to be a generational fund-returner because it is fundamentally redefining the Total Addressable Market (TAM) of developer tools. Historically, developer tools were priced as SaaS (capturing a fraction of a developer’s workflow value). Cognition is pricing and positioning as labor. By building Devin, the world’s first autonomous AI software engineer, Cognition is targeting the $1 trillion+ global software engineering payroll. If Cognition captures even 1% of the work currently done by human engineers, it will be one of the most valuable companies in the world. They are leading the paradigm shift from "AI as a Copilot" (autocomplete) to "AI as an Autopilot" (asynchronous task execution).

Product & Wedge Cognition’s core product is Devin, an autonomous agent equipped with its own command line, code editor, and browser. Unlike traditional LLMs that require constant human prompting, Devin is given a high-level objective (e.g., "build a web scraper for this site and deploy it," or "resolve this GitHub issue") and autonomously plans, writes, debugs, and deploys the code.

The wedge is asynchronous, discrete engineering tasks. Instead of trying to replace a senior architect, Devin is positioned as an autonomous junior developer. It takes on migrations, API integrations, bug fixes, and QA testing.

Crucially, Cognition’s acquisition of Windsurf (an AI-native IDE) provides a brilliant secondary wedge. While Devin operates asynchronously in the cloud, Windsurf owns the local developer environment. This allows Cognition to capture real-time human keystrokes, context, and intent, bridging the gap between human-in-the-loop coding and fully autonomous agentic execution. Windsurf acts as the data flywheel to train Devin’s future capabilities.

Market & Competition The market is the global spend on software engineering labor. However, the race to build the ultimate AI developer is the most fiercely contested space in Silicon Valley.

Real competitors include:

  • Magic.dev: Building proprietary foundation models with ultra-long context windows specifically for code generation.
  • Anysphere (Cursor): The dominant AI-native IDE. They currently own the developer workflow and are slowly moving from autocomplete to agentic features.
  • GitHub (Microsoft): GitHub Copilot Workspace is moving rapidly toward issue-to-pull-request autonomy, backed by infinite distribution.
  • Factory / Poolside: Well-funded startups building autonomous developer agents and custom code models.
  • Open Source: Frameworks like SWE-agent and AutoGPT.

Traction & Business Signal

  • Performance: At launch, Devin achieved state-of-the-art results on the SWE-bench coding benchmark, successfully resolving real-world GitHub issues unassisted at a rate far higher than base models.
  • Top-of-Funnel: Massive viral launch resulting in an enormous waitlist of enterprise and individual developers.
  • Capital: Raised significant capital from top-tier investors (Founders Fund) at unicorn-plus valuations almost immediately after launch.
  • Revenue & Retention: Unknown. It is publicly unknown how many enterprise contracts have been signed, what the ACV is, or what the retention rate looks like after the initial pilot phase.

Risks These are the three existential threats to the business:

  1. The Foundation Model Collision (Wrapper Risk): Cognition relies heavily on underlying frontier models (like OpenAI’s GPT-4/5 or Anthropic’s Claude). If OpenAI releases an inherently agentic model that can natively plan, execute, and debug code in a loop, Cognition’s core value proposition could be commoditized overnight. They must prove their agentic scaffolding and proprietary fine-tuning outpace the base models.
  2. The 95% Reliability Ceiling: Demos are easy; production is brutal. In software engineering, code that is 95% correct is often 100% broken. If Devin frequently gets stuck in infinite debugging loops or hallucinates subtle security vulnerabilities, the human time required to review Devin’s work will exceed the time it takes to write the code from scratch. The ROI collapses if autonomy isn't truly autonomous.
  3. Enterprise Integration Friction: Writing code in a vacuum is a solved problem. Writing code within a massive, undocumented, 10-year-old enterprise monolith is entirely different. If Devin cannot securely ingest, understand, and respect complex enterprise architectures and compliance guardrails, it will be relegated to building toy apps and side projects rather than serving Fortune 500 engineering teams.

Verdict Invest.

At a hypothetical Series A, the risk-reward asymmetry is too compelling to ignore. The transition from copilot to autonomous agent is inevitable, and Cognition has assembled one of the highest-talent teams in the world to solve the hardest UX and scaffolding problems in the space. The Windsurf acquisition shows strategic maturity, securing the developer surface area to defend against pure model providers.

The One Question to Answer First: What is your proprietary data advantage, and how does your agentic scaffolding improve at a rate faster than the underlying foundation models you rely on?

GPT-5.6 Luna ProWinner
$0.007 · 26s · 4,859 tokens

Thesis

Cognition could be a fund-returner if it becomes the default execution layer for software development—not merely an AI code-completion product. The prize is unusually large: software engineering is a high-value, labor-intensive workflow, and an agent that can take a ticket from specification through implementation, testing, debugging, pull request, and deployment could capture a meaningful share of engineering budgets. At sufficient reliability, the product might be priced per seat, per task, or against measurable engineering output, creating a software business with unusually high willingness to pay.

The acquisition of Windsurf materially strengthens that possibility. Cognition now has an established developer environment, distribution, and product surface alongside Devin’s more autonomous agent. The combination could allow Cognition to own both the interactive “IDE” workflow and the asynchronous “delegate this task” workflow. That is a credible platform thesis rather than a single-feature bet.

The counterpoint is that the category may become an interface layer controlled by foundation-model companies or bundled into existing developer tools. Cognition must therefore build a durable data, workflow, or distribution advantage quickly.

Product & wedge

Devin launched publicly in 2024 as an autonomous software engineer. Its wedge is not code generation in isolation, but completing multi-step engineering tasks inside a persistent development environment: reading a codebase, making changes, running tools, debugging, and communicating progress. This targets work that is too complex for traditional autocomplete but too routine or bounded to justify a senior engineer’s attention.

Windsurf adds an AI-native integrated development environment, developed by Codeium, with inline assistance, codebase context, and agentic editing. Its strength has been embedding AI into the developer’s existing loop rather than asking users to hand off work to a separate autonomous system. Cognition can plausibly use Windsurf as the high-frequency entry point and Devin as the higher-autonomy execution engine.

The important product question is whether the two experiences converge into a coherent system: a developer can edit interactively, delegate a larger task, review the result, and send follow-up instructions without changing tools. If so, Cognition may own a larger portion of the engineering workflow than either an IDE assistant or an isolated coding agent.

Market & competition

The market is large but crowded. GitHub Copilot has distribution through GitHub and Microsoft and is increasingly moving from completion toward agentic coding. Cursor, from Anysphere, has become a major AI-native coding environment with strong developer mindshare. Replit offers browser-based development and agentic app creation. Anthropic’s Claude Code, OpenAI’s Codex, Google’s Jules, and Amazon Q Developer compete directly in task-oriented software engineering. Sourcegraph Cody, Tabnine, JetBrains AI, and Augment also address enterprise coding workflows.

Windsurf itself competed with Cursor and Copilot before becoming part of Cognition. The acquisition may give Cognition a stronger product and user base, but it does not eliminate the underlying threat: foundation-model providers can improve coding agents and distribute them through existing APIs, IDEs, cloud platforms, or developer repositories.

Cognition’s differentiation must therefore be measured in completed engineering work, reliability, repository-level context, integrations, and enterprise trust—not impressive demos or benchmark scores alone.

Traction & business signal

Publicly known signals are strong but incomplete. Devin generated substantial attention at launch and was reportedly adopted by early enterprise customers. Cognition announced significant venture financing, including a reported $2 billion valuation in 2024. The company has also publicly demonstrated Devin working on software tasks and positioned it for enterprise use.

Cognition announced its acquisition of Windsurf in 2025 after Windsurf’s previously reported transaction with Google did not close. The acquisition brings Windsurf’s product, team, and developer distribution into Cognition. Public reporting has described the transaction, but the exact purchase consideration, retained users, revenue contribution, and retention metrics are unknown.

Current annual recurring revenue, gross margin, net revenue retention, paid customer count, usage per customer, inference cost per completed task, and the percentage of Devin outputs accepted without substantial human rework are unknown. It is also unknown whether Windsurf customers will adopt Devin’s autonomous workflows or whether the products will remain technically and commercially separate.

Risks

1. The product does not achieve reliable autonomy. This is the deal-killer. If Devin routinely produces code that requires extensive review, agents become expensive copilots rather than substitutes for meaningful engineering work. Model progress may improve capability, but production repositories contain hidden dependencies, ambiguous requirements, security constraints, and difficult debugging loops. The business case depends on verifiable productivity, not task completion theater.

2. Distribution and economics are captured by incumbents. Microsoft can bundle Copilot; Anthropic and OpenAI can offer strong coding agents at model-level pricing; Cursor can iterate rapidly with a focused product. Cognition may end up paying model providers for inference while competing against companies with better distribution or larger balance sheets. Windsurf helps, but switching costs for AI coding tools remain relatively low.

3. Enterprise trust and organizational adoption lag capability. Autonomous agents touch proprietary code, credentials, production systems, and potentially regulated data. Security, permissions, auditability, and liability will determine whether large customers allow meaningful autonomy. There is also an adoption risk: engineering leaders may buy licenses but constrain agents to low-value tasks, limiting expansion and pricing power.

Verdict

Invest at a hypothetical Series A, but only conditionally. The upside is venture-scale: a credible path from coding assistant to the operating system for software production, with Windsurf providing a valuable distribution wedge. The category is strategically important, and Cognition has demonstrated unusual ambition and market visibility.

The one question I would need answered first is: For paying customers, what percentage of delegated engineering tasks does Devin complete to an accepted pull request, without substantial human rewriting, and what is the fully loaded cost per accepted task? Without that evidence, the company is a compelling narrative rather than an investable autonomous-engineering business.

Gemini 3.1 Pro vs GPT-5.6 Luna Pro, overall

Across 31 judged battles, Gemini 3.1 Pro won 9, GPT-5.6 Luna Pro won 18, and 4 were ties. On investment memo alone: 1 to 8, 2 tied.

Compare Gemini 3.1 Pro and GPT-5.6 Luna Pro Judge one yourself, blind

Costs are the provider's list prices at the time of the run.