Codeium
Last reviewed
Sources
28 citations
Review status
Source-backed
Revision
v8 ยท 4,416 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Last reviewed
Sources
28 citations
Review status
Source-backed
Revision
v8 ยท 4,416 words
Add missing citations, update stale details, or suggest a clearer explanation.
Codeium was an artificial intelligence company that built free, unlimited AI code completion and the Windsurf Editor, the IDE its founders called "the first agentic IDE," before becoming the center of a 72-hour 2025 acquisition saga that split it between Google and Cognition AI. Founded in 2021 by Varun Mohan and Douglas Chen under the name Exafunction, the company first focused on GPU virtualization before pivoting to AI code assistance in 2022 and rebranding to Codeium. It reached unicorn status with a $1.25 billion valuation in August 2024 after growing to more than 700,000 active developers. [1] In November 2024 it launched the Windsurf Editor, a Visual Studio Code fork built around an agentic AI system called Cascade, and in April 2025 the company renamed itself Windsurf. After an attempted $3 billion purchase by OpenAI collapsed over Microsoft intellectual property disputes, Google hired the founding team in a $2.4 billion licensing deal on July 11, 2025, and three days later Cognition AI, the company behind the Devin autonomous coding agent, acquired the remaining company. Under Cognition's ownership the Windsurf product was developed further and, on June 2, 2026, it was renamed Devin Desktop. [21]
Varun Mohan and Douglas Chen first met in middle school and later became classmates at the Massachusetts Institute of Technology (MIT), where both graduated in 2017. Mohan earned a Master of Engineering in Computer Science alongside a Bachelor's in Electrical Engineering and Computer Science. Chen completed a Bachelor of Science in Computer Science and Engineering with a minor in Mathematics.
After MIT, the two founders pursued separate careers in the technology industry. Mohan joined Nuro, an autonomous delivery startup, where he served as a tech lead managing the autonomy infrastructure team. Chen went to Meta (then Facebook), where he built software tools for virtual reality headsets including the Oculus Quest. Chen also gained experience through internships at SpaceX, Microsoft, and Addepar.
In 2021, Mohan and Chen reunited to co-found Exafunction, a company focused on GPU virtualization and optimization for AI workloads. The startup developed infrastructure to help organizations deploy deep learning models at scale across public cloud environments, managing thousands of GPUs. The company attracted early investor interest and raised a seed round to fund its initial development.
In 2022, Mohan and Chen recognized that the release of powerful large language models had created a larger opportunity in AI-assisted software development. Rather than continuing as an infrastructure provider, they decided to build directly on top of their GPU infrastructure to create an AI code acceleration tool. This strategic pivot led to the rebranding from Exafunction to Codeium.
The company raised a $25 million Series A round in April 2022, led by Greenoaks Capital Partners. With these funds, the team built Codeium's core product: an AI-powered code completion extension that could be installed across a wide range of integrated development environments.
Codeium launched its beta product in October 2022, offering free AI code completion to individual developers. This was a deliberate strategic choice to differentiate from GitHub Copilot, which charged $10 per month for individual users. The free-tier approach proved effective for building a large user base quickly.
Codeium's growth accelerated sharply through 2023 and into 2024. By mid-2023, the platform had attracted over 300,000 active users and signed 100 enterprise customers. The company's approach of offering unlimited free code completion for individuals while charging enterprises for advanced features, security controls, and on-premises deployment proved to be a viable business model.
In January 2024, Codeium announced a $65 million Series B round at a $500 million valuation, led by Kleiner Perkins with participation from Greenoaks and General Catalyst. [3] By this time, the team had grown to approximately 80 professionals.
By August 2024, the company had reached over 700,000 active developers, and its enterprise product had crossed eight figures in annual recurring revenue (ARR), with ARR growing by over 500% since early 2024. Codeium processed more than 100 billion tokens daily and had been integrated into production workflows at companies including Zillow, Dell, and Anduril. [1]
On August 29, 2024, Codeium announced a $150 million Series C funding round led by General Catalyst, with continued participation from Kleiner Perkins and Greenoaks. This round pushed the company's valuation to $1.25 billion, achieving unicorn status in less than two years from inception. [1][2] Total funding reached approximately $243 million.
In November 2024, Codeium launched the Windsurf Editor, which it described as "the first agentic IDE." [11] The editor was built as a fork of Visual Studio Code, allowing developers to use a familiar interface while gaining access to integrated AI capabilities that went beyond what a traditional IDE extension could provide.
At the core of the Windsurf Editor was Cascade, an AI assistant that combined deep codebase understanding, a suite of development tools, and real-time awareness of developer actions. Unlike traditional AI coding assistants that operated in either a reactive (copilot) mode or a fully autonomous (agent) mode, Cascade introduced what the company called "Flows," a system that allowed the AI to move fluidly between levels of autonomy depending on the task at hand.
The Windsurf Editor attracted over 10,000 users within its first two days and hundreds of thousands within weeks. By April 2025, the platform had surpassed 1 million developers. ARR grew from roughly $12 million in late 2024 to $100 million by April 2025, representing more than an eightfold increase in four months. [12]
In April 2025, the company officially rebranded from Codeium to Windsurf, reflecting a strategic shift from being primarily an autocomplete plugin provider to a full AI-native development platform. The company noted that most users already referred to the product as Windsurf, and the new name better captured the company's broader ambition. [10] The name "Windsurf" was chosen to symbolize the combination of human and machine working together, much as windsurfing combines the rider, the board, and the wind.
As part of the rebrand, the existing Codeium extensions for third-party IDEs were renamed to Windsurf Plugins.
On May 15, 2025, Windsurf announced SWE-1, its first family of proprietary AI models built specifically for software engineering tasks. [11] The SWE-1 family included three models:
| Model | Purpose | Availability |
|---|---|---|
| SWE-1 | Full-size model for advanced reasoning and tool use | Pro and Enterprise users |
| SWE-1-lite | Mid-size model replacing Cascade Base, for general coding tasks | All users (free and paid) |
| SWE-1-mini | Lightweight model powering Windsurf Tab for fast code prediction | All users (unlimited) |
The SWE-1 models were designed around a concept the company called "flow awareness," which enabled reasoning over long-running, multi-surface engineering tasks. According to Windsurf, SWE-1 performance was competitive with models like Claude 3.5 Sonnet and GPT-4.1, at a more favorable cost-to-performance ratio.
In February 2025, reports emerged that Codeium was in talks to raise funding at a valuation approaching $3 billion. However, events took a different turn. In early May 2025, Bloomberg reported that OpenAI had agreed to acquire Windsurf for approximately $3 billion, which would have been one of the largest acquisitions in the AI coding space.
The deal collapsed due to a fundamental dispute over intellectual property rights. Microsoft, which had invested over $13 billion in OpenAI since 2019, held contractual rights to much of OpenAI's IP under a sweeping 2023 agreement. This included access to model weights, code, and any IP that OpenAI gained through acquisitions. Under the terms of this agreement, Windsurf's technology would have effectively become accessible to Microsoft.
Varun Mohan reportedly objected to this arrangement, since Microsoft's GitHub Copilot was a direct competitor to Windsurf. OpenAI and Microsoft could not resolve the IP dispute, and the exclusivity period on the deal expired in July 2025. [5]
On July 11, 2025, Google announced a $2.4 billion licensing and talent deal with Windsurf. [4] The arrangement was structured as an "acqui-hire" rather than a full acquisition: Google paid $2.4 billion in licensing fees and compensation to secure a non-exclusive license to Windsurf's technology and to hire key personnel. Google did not take an equity stake in the company. [6]
Varun Mohan, Douglas Chen, and approximately 40 senior research and development staff joined Google DeepMind to work on agentic coding capabilities within the Gemini program.
Just three days later, on July 14, 2025, Cognition AI (the company behind the Devin autonomous coding agent) signed a definitive agreement to acquire what remained of Windsurf, including its intellectual property, product, trademark, brand, and business operations. [7] The price was estimated at approximately $250 million. All Windsurf employees not hired by Google joined Cognition, with 100% of staff participating financially in the deal and receiving fully accelerated vesting.
At the time of acquisition, Windsurf had reached $82 million in ARR with over 350 enterprise customers and hundreds of thousands of daily active users. Following the acquisition, Cognition's combined enterprise ARR increased by more than 30%, and in September 2025, Cognition closed a $400 million funding round at a $10.2 billion valuation. [8][20]
On August 5, 2025, about three weeks after the acquisition was announced, Cognition offered Windsurf employees the option to leave with a severance package, presenting departing staff with nine months of salary and accelerated vesting. Cognition framed the move as giving employees who did not want to commit to the company's intensive schedule a graceful exit. [22]
After absorbing Windsurf, Cognition kept the Windsurf Editor as a standalone product while progressively integrating its own Devin agent and proprietary models. Two product cycles in late 2025 began this integration. In Windsurf Wave 12, the editor gained automatic planning, in which Cascade decided on its own when to draft a plan and build to-do lists, and added the first Devin-derived features. [23] On December 24, 2025, Windsurf Wave 13, branded "Merry Shipmas," introduced first-class support for parallel, multi-agent sessions, Git worktrees that let multiple Cascade sessions run in the same repository on separate branches, side-by-side Cascade panes, a dedicated shell for agent commands, and a context-window usage indicator. [24]
On October 29, 2025, Cognition released SWE-1.5, the next generation of Windsurf's in-house model family. The company described it as a frontier-size model with hundreds of billions of parameters that scored 40.08% on the SWE-Bench Pro benchmark from Scale AI while completing tasks far faster than comparable models. [25] Cognition partnered with the AI hardware company Cerebras to serve SWE-1.5 at up to 950 tokens per second, which it said was about 6 times faster than Claude Haiku 4.5 and 13 times faster than Claude Sonnet 4.5. [25] The model was trained on the same Cascade agent harness it runs in, an approach Cognition described as co-developing the model and the harness as one unified system. A free version of SWE-1.5 was made available to all users for three months, offering the full model intelligence at standard throughput. [25]
On May 27, 2026, Cognition announced a Series D funding round of more than $1 billion at a post-money valuation of $26 billion (a $25 billion pre-money valuation), led by Lux Capital, General Catalyst, and 8VC, with participation from existing investors including Founders Fund, Elad Gil, Bain Capital Ventures, and 137 Ventures and new investors including Ribbit Capital, Atreides, and Layer Global. [26] The company said its revenue run rate had grown to $492 million, rising about 50 percent month over month over the prior six months, and that 89 percent of the code committed by its own engineers was now written by Devin, with the remainder written by local agents in Windsurf. Cognition listed enterprise customers including Citi, Mercedes-Benz, Goldman Sachs, Dell, Santander, the U.S. Army, and the U.S. Navy, and said one customer, Mercedes-Benz, had compressed an eight-month legacy modernization project to about eight days. [26]
On April 7, 2026, Cognition released SWE-1.6, which it called its best model on both intelligence and what it termed "model UX." The company said SWE-1.6 matched its preview model on SWE-Bench Pro (the preview had improved on SWE-1.5 by more than 10 percent on that benchmark) while reducing behaviors such as overthinking simple problems, taking unnecessary turns, and getting caught in repetitive reasoning loops. SWE-1.6 was offered in two modes inside Windsurf: a free tier at up to 200 tokens per second served through Fireworks, and a fast tier at up to 950 tokens per second served through Cerebras. Cognition later reported that SWE-1.6 had become the most used model in Windsurf. [27]
On April 15, 2026, Cognition launched Windsurf 2.0, which brought Devin directly into the editor. The release introduced the Agent Command Center, a Kanban-style surface that shows every agent a developer is running, both local and cloud, organized by status, and Spaces, a way to group an agent's sessions, pull requests, files, and context for a given task. From within Windsurf, a developer could send a plan to Devin with a single click, after which Devin provisioned its own cloud machine to carry out the work. Cognition said Devin was included with every Windsurf plan. [28]
On June 2, 2026, Cognition retired the Windsurf name and relaunched the editor as Devin Desktop, describing it as "a new look for the product you already love." [21] As part of the change, the company replaced the Cascade agent with Devin Local, a local agent that Cognition said it had rewritten from scratch in Rust to be up to 30 percent more token-efficient and to support subagents, which are specialized sub-sessions that handle parts of a task in parallel. Cognition said the legacy Cascade agent would remain usable until July 1, 2026. Devin Desktop also added support for the Agent Client Protocol (ACP), an open protocol for connecting coding agents to editors, with the editor able to run agents such as OpenAI Codex and Claude Agent alongside Cognition's own. The company said existing Windsurf users would receive an automatic update and that their plan, pricing, extensions, and features would remain the same. [21]
Codeium's original product was an AI-powered code completion tool delivered as an extension for existing IDEs. The tool provided single-line and multi-line code suggestions, intelligent search, and context-aware autocomplete that adapted to individual coding patterns and project requirements.
Key specifications of the Codeium extension included:
| Feature | Details | |---|---|---| | Supported languages | 70+ programming languages including Python, JavaScript, TypeScript, Java, Go, C++, Rust, and more | | Supported IDEs | 40+ IDEs including VS Code, JetBrains, Visual Studio, Vim/Neovim, Emacs, Eclipse, Sublime Text, Jupyter, and Chrome | | Code completion | Single-line and multi-line with multiple suggestions | | Chat interface | Code explanations, refactoring suggestions, language translations | | Codeium Search | Semantic code search (Python, JavaScript, TypeScript, Go) | | Enterprise deployment | On-premises and VPC deployment options |
For individual developers, code completion was entirely free with no rate limits, which distinguished Codeium from competitors like GitHub Copilot that required paid subscriptions.
The Windsurf Editor was a standalone IDE built as a fork of Visual Studio Code. It was designed to remove the limitations inherent in running AI tools as extensions within an existing editor, allowing tighter integration between the AI system and the development environment.
Cascade was Windsurf's central AI assistant. It combined several capabilities:
In June 2026, after Windsurf was renamed Devin Desktop, Cognition replaced Cascade with a Rust-based successor called Devin Local and said the legacy Cascade agent would remain available until July 1, 2026. [21]
Windsurf Tab was the editor's autocomplete system, powered by a technology called Supercomplete. It combined several capabilities into a single keystroke:
Windsurf built and trained its own AI models rather than relying exclusively on third-party providers. The company's model strategy encompassed pretraining, task-specific fine-tuning, and customer-specific fine-tuning. Being embedded in the IDE gave Windsurf access to rich behavioral data, such as whether code completions were accepted or rejected and what developers did after accepting suggestions, which fed back into model improvement.
The backend infrastructure included custom models for both low-latency keystroke-level features and code generation, proprietary retrieval systems for large codebases, and containerized systems designed for on-premises and VPC deployment.
Under Cognition, the in-house model line continued with SWE-1.5 (October 2025) and SWE-1.6 (April 2026), both served at high throughput through partnerships with the inference providers Cerebras and Fireworks. [25][27]
Windsurf offered a tiered pricing model designed to serve individual developers through large enterprises.
| Plan | Monthly Price | Key Features |
|---|---|---|
| Free | $0 | 25 prompt credits/month, unlimited Tab autocomplete, 1 App Deploy/day |
| Pro | $15/user | 500 prompt credits/month, all premium models, unlimited Tab and Command |
| Teams | $30/user | 500 prompt credits/user/month, centralized billing, team management |
| Enterprise | $60/user | 1,000 prompt credits/month, SSO integration, enhanced security |
Additional prompt credits could be purchased at a rate of 250 credits for $10. New users received a two-week free Pro trial with 100 credits. [17]
After Cognition integrated Devin into the editor, the company said Devin was included with every Windsurf plan, and when the product was renamed Devin Desktop in June 2026 it said existing plans and pricing would carry over unchanged. [28][21]
The following table summarizes Codeium's known funding rounds. Later rounds, raised after the Cognition acquisition, were raised by Cognition rather than by the standalone Windsurf entity.
| Round | Date | Amount | Lead Investor | Valuation |
|---|---|---|---|---|
| Seed | 2021 | Undisclosed | Undisclosed | N/A |
| Series A | April 2022 | $25 million | Greenoaks | Undisclosed |
| Series B | January 2024 | $65 million | Kleiner Perkins | $500 million |
| Series C | August 2024 | $150 million | General Catalyst | $1.25 billion |
| Series C (Cognition) | September 2025 | $400 million | Undisclosed | $10.2 billion |
| Series D (Cognition) | May 2026 | $1 billion+ | Lux Capital, General Catalyst, 8VC | $26 billion |
Total disclosed funding for the standalone company reached approximately $243 million. Key investors across multiple rounds included Greenoaks, Kleiner Perkins, and General Catalyst. After the acquisition, the Windsurf product was funded as part of Cognition, which raised a $400 million round at a $10.2 billion valuation in September 2025 and a Series D of more than $1 billion at a $26 billion valuation in May 2026. [26]
Codeium operated in the rapidly growing AI coding assistant market, competing with several major products.
GitHub Copilot, developed by GitHub (a Microsoft subsidiary) in partnership with OpenAI, was the market leader in AI code completion. Copilot had over 20 million users and 1.3 million paid subscribers. It was priced at $10 per month for individuals and offered tight integration with GitHub's ecosystem. Copilot was known for reliable autocomplete, particularly for boilerplate code, tests, and configuration files.
Cursor, built by Anysphere, positioned itself as an AI-native IDE with deep project-wide context understanding. Cursor featured a "Composer" mode for multi-file refactoring and was priced at $20 per month for its Pro plan. The company experienced explosive growth in 2025, surpassing $500 million in ARR by June 2025 and reaching a $29.3 billion valuation after a $2.3 billion Series D round in November 2025. [15][16]
Amazon Q Developer, offered by Amazon Web Services, provided AI coding assistance integrated with the broader AWS ecosystem. It included inline code suggestions, chat-based assistance, and the ability to generate, debug, and optimize code within AWS development workflows.
| Feature | Codeium/Windsurf | GitHub Copilot | Cursor | Amazon Q Developer |
|---|---|---|---|---|
| Type | Standalone IDE + plugins | IDE extension | Standalone IDE | IDE extension + CLI |
| Free tier | Yes (unlimited autocomplete) | Limited | Limited | Yes (limited) |
| Starting price | $15/month | $10/month | $20/month | Free tier available |
| IDE base | VS Code fork | Extension for VS Code, JetBrains, etc. | VS Code fork | Extension for VS Code, JetBrains |
| Agentic features | Cascade (Flows) | Copilot Chat, Agent mode | Composer, Agent mode | Agent capabilities |
| Custom models | SWE-1 family | OpenAI models | Multiple model providers | Amazon Bedrock models |
| Enterprise focus | On-prem, VPC deployment | GitHub Enterprise | Teams plan | Deep AWS integration |
| Name | Role | Background |
|---|---|---|
| Varun Mohan | Co-Founder and CEO | MIT MEng CS '17; former tech lead at Nuro; joined Google DeepMind in July 2025 |
| Douglas Chen | Co-Founder | MIT BS CSE '17; former ML engineer at Meta (Oculus VR); joined Google DeepMind in July 2025 |
| Anshul Ramachandran | Head of Enterprise and Partnerships (Founding Team) | Caltech; former Technical Lead Manager at Nuro |
After the July 2025 transactions, leadership of the Windsurf product passed to Cognition, whose chief executive is Scott Wu. [22]
Codeium's trajectory from a GPU virtualization startup to a billion-dollar AI coding company illustrates the speed at which the AI developer tools market evolved between 2022 and 2025. The company's decision to offer free unlimited code completion to individuals helped it build a large user base and proved that freemium models could work in the AI coding space.
The dramatic 72-hour period in July 2025, during which Windsurf was effectively split between Google and Cognition after the OpenAI deal collapsed, became one of the most discussed corporate events in the AI industry. The episode highlighted the complex dynamics between major technology companies competing for AI talent and technology, and raised questions about how IP agreements between investors and startups could affect acquisition outcomes.
Under Cognition's ownership, the Windsurf product continued to operate and was developed further, with Cognition integrating its Devin autonomous coding agent directly into the Windsurf IDE through 2025 and 2026 before renaming the product Devin Desktop in June 2026. [21] The combined entity represented one of the most comprehensive AI coding platforms, blending Windsurf's IDE and developer community with Devin's autonomous task completion capabilities. By May 2026, Cognition reported a revenue run rate of $492 million and a $26 billion valuation, with the Windsurf product, its in-house SWE models, and the Devin agent forming a single coding platform. [26]