The 72-Hour Deal

On a Friday evening in July 2025, with OpenAI's $3 billion acquisition dead and Google's $2.4 billion talent raid still fresh, a startup called Cognition made first contact with Windsurf. By Monday morning, they had a deal. The price was $250 million — less than a tenth of what the first bidder offered. It bought them an IDE with 350 enterprise customers, $82 million in annual recurring revenue, and the wreckage of the most chaotic acquisition weekend in AI history.¹

The Weekend

The sequence started months earlier. OpenAI had negotiated a $3 billion acquisition of Windsurf, the AI-native code editor formerly known as Codeium. The deal collapsed over conflicts with Microsoft's partnership agreements — Microsoft had its own AI coding tool in Copilot, and the competitive overlap was too direct for the partnership to absorb.²

Google moved next. Rather than buying the company, Google paid approximately $2.4 billion to hire Windsurf's CEO Varun Mohan, co-founder Douglas Chen, and around 40 employees. Google also licensed some of Windsurf's technology. It was an acqui-hire structured to avoid the regulatory scrutiny of a full acquisition — you don't need antitrust approval to make job offers.³

That left the rest of Windsurf — the product, the brand, the IP, the customer base, and the employees who weren't moving to Mountain View — in a kind of corporate limbo. A company with real revenue and real users, suddenly missing its leadership.

Cognition's founders reached out that Friday evening. Scott Wu, Steven Hao, and Walden Yan — three International Olympiad in Informatics gold medalists who had founded Cognition in August 2023 — saw the opening. They were building Devin, the autonomous coding agent that had made headlines the previous year. What they didn't have was an IDE. What Windsurf didn't have was a CEO.¹

The negotiation took the weekend. On Monday, July 14, Cognition announced the acquisition of Windsurf's remaining assets — intellectual property, product, brand, and customer relationships — for approximately $250 million. Every Windsurf employee not hired by Google was included. Cognition waived all vesting cliffs and gave 100% of the acquired staff financial participation in the combined company.¹

Three bids on one company in the span of a few months. OpenAI offered $3B and got nothing. Google paid $2.4B and got the people. Cognition paid $250M and got the product.

What Cognition Bought

The numbers Cognition inherited were not trivial. Windsurf had $82 million in annual recurring revenue. Its enterprise business was doubling quarter over quarter. It had more than 350 enterprise customers and hundreds of thousands of daily active users.¹

Cognition itself was well-capitalized for the purchase. In September 2025, the company had raised a $400 million round at a $10.2 billion valuation.⁴ The Windsurf acquisition cost roughly 60% of that single fundraise. For a company valued at ten billion dollars, $250 million to acquire an IDE with $82 million ARR was, by the standards of 2025 AI dealmaking, a bargain.

The strategic logic was straightforward. Cognition had been building Devin as a standalone autonomous agent — a system that could take a task description, plan an approach, write code, debug it, and submit the result. Devin operated in a cloud sandbox, separate from any developer's local environment. It was powerful in concept and limited in practice. Windsurf gave Cognition a place to put Devin where developers already were: inside the editor.

Devin's Track Record

When Cognition launched Devin in March 2024, the benchmarks were genuinely impressive. Devin scored 13.86% on SWE-bench, a standard evaluation for automated software engineering. The previous best autonomous result was 1.96%. Devin was not incrementally better; it was categorically different from anything else available.⁵

The initial pricing reflected the ambition: $500 per month. Cognition positioned Devin as an autonomous engineer, not a coding assistant. The framing was deliberate — this was not autocomplete, this was a colleague.

Then independent testers got access. Of 20 tasks assigned to Devin in one widely circulated evaluation, it completed 3. A 15% success rate.⁶ The gap between the benchmark performance and real-world utility was significant enough to recalibrate expectations across the industry about what "autonomous coding agent" actually meant in production.

Cognition adapted. Devin 2.0 launched in April 2025 with the price cut from $500 to $20 per month — a 96% reduction. The repositioning was implicit: Devin was no longer priced as a replacement engineer. It was priced as a tool.⁶

The revenue growth told a different story than the completion rates. Devin went from $1 million ARR in September 2024 to $73 million ARR by June 2025.⁴ Cognition's internal model, SWE-1.5, pushed to 40.08% on SWE-Bench Pro, a harder variant of the benchmark.⁵ As of April 2026, Bloomberg reported Cognition was raising at a $25 billion valuation.⁴

The math here is worth pausing on. An agent that independent testers found completes 15% of assigned tasks is the core product of a company raising at $25 billion. The valuation is not based on what Devin does today. It is based on what the combination of Devin and Windsurf might do tomorrow.

Windsurf 2.0

The first major product of the merger shipped on April 15, 2026 — nine months after the acquisition. Windsurf 2.0 introduced three features that made the Cognition integration explicit.⁷

The first is Agent Command Center, a kanban-style interface for managing multiple agent sessions simultaneously. Instead of running one agent conversation in a single editor pane, developers can now spin up parallel sessions, monitor their progress, and intervene when an agent stalls or drifts. It treats agents less like chat partners and more like processes to be supervised.

The second is Spaces, a system for bundling context — files, documentation, configuration — into portable packages that can be attached to any agent session. The problem Spaces addresses is one that every AI coding tool faces: agents perform poorly when they lack context about the codebase, but gathering that context manually for each session is tedious enough that developers skip it. Spaces makes the context reusable.

The third is the Devin handoff. A developer can plan a task locally using Windsurf's Cascade agent, get the approach right in a familiar environment with full code visibility, and then hand the execution off to Devin running in the cloud with a single click. Devin picks up where Cascade left off, executes the plan autonomously, and returns the result.⁷

The handoff model is the real product thesis. It concedes that autonomous agents are not reliable enough to work unsupervised from the start. But it proposes that an agent supervised through the planning phase and then released for execution might hit a higher success rate than one that operates autonomously from the first prompt. The IDE becomes the supervision layer. The human does the hard part — scoping the task, choosing the approach — and the agent does the repetitive part.

The Price of the IDE

Windsurf's pricing changed significantly in the months between the acquisition and the 2.0 launch. On March 19, 2026, Windsurf transitioned from a credit-based system to a quota model. Pro went from $15 to $20 per month. Teams went from $30 to $40. A new Max tier appeared at $200 per month.⁸

The free tier was hit hardest. Users reported being locked out of core features that had previously been available without payment. The transition from credits to quotas meant that the old mental model — spend credits on the models you want — was replaced with a tiered access model where the most capable models required the most expensive plans.⁹

This is the same pattern we have documented across the AI coding tool market. Subsidize to acquire users, then adjust pricing to reflect actual inference costs. Windsurf's version was accelerated by the acquisition — new ownership, new cost structure, new pricing. The $82 million in ARR that Cognition acquired was generated under the old pricing. The new pricing is designed to fund the combined company's ambitions.

The Consolidation

Step back and look at the AI coding tool market as of May 2026. The total market hit $12.8 billion.¹⁰ Three consolidation moves have reshaped the competitive landscape in under a year.

Microsoft has Copilot, integrated into GitHub and VS Code, backed by OpenAI's models, installed on 20 million enterprise seats and now transitioning to usage-based billing that will test how many of those seats survive the price adjustment.

Cursor, the other major VS Code fork, reached $2 billion ARR and entered talks for a deal — potentially with SpaceX and xAI — at a reported $60 billion valuation.¹⁰ If that deal closes, Cursor gets access to xAI's Grok models and the operational infrastructure of companies that build rockets and run data centers.

Cognition now has Windsurf plus Devin — the IDE plus the agent, combined under one roof, raising at $25 billion.

The major independent left standing is Claude Code, Anthropic's terminal-based coding agent. It overtook Copilot as the most-used AI coding tool on small engineering teams, with 75% adoption among teams under 50 engineers.¹⁰ Claude Code does not have an IDE. It runs in the terminal. Its independence from the IDE wars is both its limitation and its insulation from the consolidation happening around it.

The developer response to all of this consolidation has been to hedge. Surveys show 70% of engineers now use between two and four AI coding tools simultaneously.¹¹ The average is 2.3 tools per developer. Loyalty to a single tool is low. Switching costs, for now, are measured in hours, not months. Every vendor in this market knows that their users are one pricing change away from testing an alternative.

The Bet

Cognition's theory of the case comes down to a single integration. Devin alone was an agent that worked in a sandbox, disconnected from the developer's actual workflow, completing a minority of assigned tasks. Windsurf alone was a capable IDE without a differentiated autonomous agent. Together, the pitch is that the IDE provides the supervision layer that makes the agent viable — and the agent provides the autonomous execution capability that makes the IDE more than an editor.

The 72-hour deal was not just fast. It was the kind of deal that only makes sense if you believe that the IDE is the control plane for AI agents in software development. Google paid ten times more and got engineers. OpenAI offered twelve times more and got nothing. Cognition paid the least and got the surface area — the product that sits open on a developer's screen for eight hours a day, the subscription that auto-renews monthly, the enterprise contracts that take six months to renegotiate.

Whether that surface area is enough depends on the agent getting better. Devin at 15% task completion is a demo. Devin at 50% is a tool. Devin at 80% is a shift in how software gets built. Cognition is betting that the path from 15% to 80% runs through the IDE — that agents improve faster when they have access to the developer's full context, supervision during planning, and a feedback loop that starts in the editor rather than in a cloud sandbox.

The $25 billion valuation prices in that improvement happening. The 72-hour deal bought the place where it would need to happen. The rest is execution — the kind that, for now, still mostly requires a human at the keyboard.