Citation and evidence

Cognition AI

31 min full readUpdated 50 references

This article's verification

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

AI AgentsAI Code GenerationAI CompaniesDeveloper Tools

Cite this article

Cognition AI is an American artificial intelligence company headquartered in San Francisco, California, that builds autonomous AI software engineering agents. Founded in late 2023 by competitive programmers Scott Wu, Steven Hao, and Walden Yan, the company is best known for Devin (AI software engineer), launched in March 2024 and billed by the company as "the first AI software engineer."[1] At launch Devin resolved 13.86% of issues on the SWE-bench benchmark unassisted, far exceeding the prior state of the art of 1.96%.[1] Cognition later acquired Windsurf (software), an agentic integrated development environment, in July 2025[6] and released its own coding model family, SWE-1.5, in October 2025.[22] In May 2026 the company raised a Series D round of more than $1 billion at a $25 billion pre-money valuation ($26 billion post-money), led by Lux Capital, General Catalyst, and 8VC, alongside a reported run-rate revenue of approximately $492 million.[34][36] In June 2026 Cognition relaunched Windsurf as Devin Desktop,[44] and in September 2026 it raised a Series E of more than $2 billion at a $48 billion valuation, led by Andreessen Horowitz and Accel; later that month it said it had crossed $1 billion in annualized revenue run rate.[38][39]

Cognition AI logo
Cognition AI logo. Source: Wikimedia Commons, Public domain.

History

Founding and early team

Cognition was founded in late 2023 by Scott Wu, Steven Hao, and Walden Yan, all of whom share an unusual background: each won gold medals at the International Olympiad in Informatics (IOI), the premier global competition in algorithmic problem-solving for secondary students.[3][14]

Scott Wu, who serves as CEO, was born in Louisiana in 1997 to a Chinese immigrant family. He attended Baton Rouge Magnet High School before enrolling at Harvard University. Wu won three IOI gold medals in 2012, 2013, and 2014, placing first overall in the 2014 competition.[2] He also won the Harvard-MIT Mathematics Tournament (HMMT) in 2014, served as Mathcounts national champion in 2011 representing Louisiana, earned a gold medal with Harvard's team at the International Collegiate Programming Contest (ICPC) in 2016, and placed third at the Google Code Jam in 2021.[2] On Codeforces, the online competitive programming platform, Wu achieved Legendary Grandmaster status with a peak rating above 3,000.[2] He left Harvard without completing his degree. Before Cognition, Wu co-founded Lunchclub, an AI-powered professional networking platform, serving as its CTO from 2017 to 2022. Forbes named him to its 30 Under 30 list in 2020.[2]

Steven Hao, who serves as CTO, graduated from MIT with a degree in computer science and mathematics. He was an IOI gold medalist in 2014, finishing sixth globally. After MIT, he joined Scale AI in 2018 and became one of its top engineers before co-founding Cognition.[14]

Walden Yan, who serves as CPO, was an IOI gold medalist in 2020, finishing nineteenth globally. He enrolled at Harvard but left without graduating in 2023. Before Cognition, he worked briefly on Cursor (code editor) at Anysphere and co-founded DeepReason, a web3 security startup.[14]

The three founders knew each other through competitive programming circuits, having crossed paths at IOI events and at Harvard and MIT. They initially explored cryptocurrency-related projects before pivoting to generative AI in late 2022 as models like ChatGPT accelerated interest across Silicon Valley. The company was formally established in late 2023.[14]

The founding team extended well beyond the three co-founders. As of March 2024, Cognition said its ten-person team collectively held ten IOI gold medals.[3] Besides the founders, gold medalists on the early team included Wu's brother Neal Wu, Andrew He, and Gennady Korotkevich, widely regarded as one of the strongest competitive programmers of all time and known by the handle "tourist" on competitive programming platforms.[3]

The concentration of elite competitive programmers was intentional. Wu has said that building an AI that can code is a "deeply algorithmic" problem well suited to the team's competitive-programming background.[14]

Emergence from stealth

Cognition operated in stealth through early 2024. A formative internal moment came in late 2023 when an early version of Devin autonomously solved a server configuration problem before the Christmas holiday, which the team cited as evidence the approach was viable.[14]

The company emerged from stealth on March 12, 2024, with a blog post and demo video announcing Devin.[1] The video, showing Devin completing software tasks from start to finish, accumulated over 30 million views on X by early 2026.[14]

Leadership additions

Alongside the three founders, Cognition built out its senior leadership through 2024 and 2025. Russell Kaplan serves as president; Hg's Orbit podcast describes him as Cognition's founding president.[33] Kaplan previously led machine learning and ML infrastructure at Scale AI, where his own computer vision startup Helia had been acquired in 2020, and before that worked on the Autopilot team at Tesla.[33] Kaplan publicly announced the Windsurf deal in July 2025[7] and frequently represents Cognition in podcasts and conference talks.[33]

Offices and team growth

Cognition began as a small team operating out of San Francisco. By early 2026 the company occupied a roughly 25,000-square-foot headquarters in the South Park neighborhood of San Francisco, which it presents as a working environment closer to a research lab than to a traditional software-as-a-service startup.[31] The space includes long tables for shared meals, private chefs, a gym, and a rooftop, and many employees live in apartments in a building around the corner.[31]

With the Windsurf acquisition completed in July 2025, Cognition inherited a workforce previously distributed across Mountain View and other locations. By March 2026 Cognition had opened offices in New York, Austin, and London and had grown from about 40 employees in mid-2025 to more than 200 after the Windsurf acquisition.[31] By September 2026 it also had offices in Washington, D.C., Tokyo, Singapore, Sao Paulo, and Madrid.[38]

How is Cognition AI funded?

Founders Fund, Peter Thiel's venture firm, led Cognition's early rounds; its 2026 rounds were led by other firms.

RoundDateAmountValuationLead investor
Series AMarch 2024$21 million$350 millionFounders Fund
Series A2April 2024$175 million~$2 billionFounders Fund
Undisclosed roundMarch 2025Undisclosed$4 billion8VC (Joe Lonsdale), Lux Capital
Series CSeptember 2025$400 million$10.2 billionFounders Fund
Series DMay 2026Over $1 billion$25 billion pre-money ($26 billion post)Lux Capital, General Catalyst, 8VC
Series ESeptember 2026Over $2 billion$48 billionAndreessen Horowitz, Accel

Expanded article table

The Series A in March 2024 raised $21 million at a $350 million valuation. The round was supported by Patrick Collison and John Collison of Stripe, Elad Gil, and others alongside Founders Fund.[14] One month later, Cognition closed a second tranche of $175 million at approximately $2 billion, making it a unicorn within its first year of public visibility. That April 2024 round was also led by Founders Fund.[14]

In March 2025, a round led jointly by 8VC, the firm co-founded by Joe Lonsdale, and Lux Capital valued Cognition at $4 billion.[3][12]

In September 2025, Cognition closed $400 million at a $10.2 billion post-money valuation. Founders Fund led the round again.[4] Returning investors included Lux Capital, 8VC, Neo, Elad Gil, Definition Capital, and Swish VC. New investors included Bain Capital Ventures, Hanabi Capital, and D1 Capital.[12] Total funding raised through that point was approximately $896 million.[14] The company disclosed it had kept total net burn under $20 million since founding, an unusually low figure given the pace of revenue growth and capital raised.[12]

The September 2025 round was followed by talks, first reported by Bloomberg in April 2026, about a new financing at a $25 billion valuation.[5] Those talks closed on May 27, 2026, when Cognition announced a Series D of more than $1 billion at a $25 billion pre-money valuation, or roughly $26 billion post-money.[34][36] The round more than doubled the company's September 2025 valuation in about eight months. It was led by Lux Capital, General Catalyst, and 8VC.[37] New investors included Ribbit Capital, Atreides, and Layer Global, while existing backers Founders Fund, Elad Gil, Alpha Wave, Definition Capital, Positive Sum, Avenir, Vitruvian, Bain Capital Ventures, Conversion Capital, 137 Ventures, Soma Capital, and Omri Casspi also took part.[34] Cognition said run-rate revenue had reached approximately $492 million and that enterprise usage of Devin (AI software engineer) had grown 50% month over month for the prior six months, more than tenfold since the start of 2026.[34][35] In its announcement, titled "More Devins in More Places," the company also stated that 89% of the code committed by its own engineers was committed by Devin, with the rest committed by local agents in Windsurf (renamed Devin Desktop the following week).[35]

In August 2026 Bloomberg reported that Cognition was in early talks to raise at a valuation of at least $40 billion.[40] On September 8, 2026, Cognition announced a Series E of more than $2 billion at a $48 billion valuation, led by new investors Andreessen Horowitz and Accel alongside existing investors Founders Fund, General Catalyst, and Avenir, and said run-rate revenue had grown from $492 million at the Series D to almost $900 million.[38] Founders Fund led the March 2024, April 2024, and September 2025 rounds and took part in the Series D and Series E.[12][14][35][38]

Products

What is Devin?

Cognition's core product is Devin (AI software engineer), an autonomous AI agent designed to complete software engineering tasks from a natural language description. Unlike code completion tools that assist developers in writing individual lines or functions, Devin is designed to take on multi-step engineering projects: writing code, running tests, debugging failures, browsing documentation, and opening pull requests without continuous human input.[1]

When Devin launched in March 2024, it achieved a score of 13.86% on the SWE-bench benchmark, a dataset of real GitHub issues drawn from popular open-source repositories. That figure, measured on a random 25% subset of the benchmark with Devin unassisted, far exceeded the previous state of the art of 1.96%; Cognition noted that the best previous models resolved only 4.80% even when told exactly which files to edit.[1] Cognition positioned Devin not as a replacement for software engineers but as an autonomous teammate that could handle delegated tasks while engineers focus on higher-level architecture and product decisions. "With Devin, engineers can focus on more interesting problems and engineering teams can strive for more ambitious goals," CEO Scott Wu wrote in the launch announcement.[1]

The original product required a subscription starting at $500 per month and operated through a web interface where developers submitted tasks and reviewed completed pull requests.[13]

For extended coverage of Devin's capabilities, benchmarks, reception, and technical architecture, see Devin (AI software engineer).

What changed in Devin 2.0?

In April 2025, Cognition released Devin 2.0 with several major changes. The most visible was a price reduction from $500 to $20 per month for a starter plan, a 96% cut that significantly broadened the accessible user base.[13] According to Cognition's internal benchmarks, Devin 2.0 completed 83% more junior-level development tasks per Agent Compute Unit (ACU) compared to prior versions.[13]

Devin 2.0 introduced an agent-native integrated development environment where users and Devin could work together within the same interface. Developers could edit and refine code directly inside the Devin IDE using familiar shortcuts. The update also added the ability to run multiple Devin instances in parallel, each with its own cloud-based IDE, allowing teams to process concurrent tasks.[9]

Three new features shipped with Devin 2.0. Interactive Planning allowed Devin to analyze a codebase and generate a preliminary plan within seconds, which the user could review and adjust before autonomous execution began. Devin Search gave users a way to ask questions about their codebase and receive detailed, cited answers through agentic exploration, with a Deep Mode option for more complex queries. Devin Wiki automatically indexed repositories into a structured wiki containing architecture diagrams, source links, and generated documentation, refreshed every few hours as the codebase changed.[9]

Devin 2.1

Devin 2.1 added a confidence scoring system. At the start of each session, after planning, and during code-related queries, Devin shows a green, yellow, or red confidence indicator expressing how certain it is that it can complete the task successfully. Sessions with green confidence scores produced pull requests that merged at roughly twice the rate of sessions with red scores. When confidence is low, Devin asks clarifying questions before proceeding.[10]

The 2.1 update also integrated Devin's codebase understanding technology directly into the task workflow. Users can query the codebase mid-session using the "!ask" command or allow Devin to automatically scan for context when it detects a gap in its understanding. The same underlying technology also appeared in DeepWiki, a separate product Cognition released for generating documentation wikis from public repositories.[10]

Through the Linear and Jira integrations, teams could request confidence assessments on multiple backlog issues simultaneously without launching full Devin sessions, letting engineering managers prioritize which tasks to delegate based on predicted success likelihood.[10]

Devin 2.2

Devin 2.2, released in February 2026, introduced desktop testing through computer use. Devin gained full access to its own Linux desktop environment, allowing it to launch and interact with desktop applications as part of testing a pull request. After completing code changes, Devin could run the application on its own desktop, record the screen, and return that recording for human review before the pull request was merged. Desktop testing was enabled by default for new users.[11]

The 2.2 update also shipped Devin Review Autofix, a self-review loop in which Devin plans, writes code, reviews its own output, identifies issues, and fixes them before the pull request is opened. The interface was fully rebuilt to unify the development lifecycle from planning through code review in a single view. Startup time improved by a factor of three.[11]

MultiDevin

Cognition introduced MultiDevin for enterprise customers, enabling parallelized task execution across multiple agents. In the MultiDevin configuration, one manager agent coordinates up to ten worker agents, each running independently on separate tasks. This allows engineering teams to delegate entire sprints rather than individual tickets.

What is DeepWiki?

In April 2025, Cognition launched DeepWiki, a free service that generates a structured, navigable wiki for any public GitHub repository. The tool repurposes the same codebase-indexing technology that powers Devin Search and Devin Wiki and exposes it through a public web interface. Users can browse generated documentation, ask questions in natural language about how a repository works, and trace answers back to specific source files.[24]

DeepWiki is accessible by replacing "github.com" with "deepwiki.com" in any repository URL, a convention that requires no login or sign-up for open-source projects.[24] At launch the service had already indexed more than 50,000 of the top public GitHub repositories, including the Model Context Protocol specification, LangChain, and many widely used machine learning libraries.[24] A subsequently released DeepWiki Model Context Protocol (MCP) server allowed external agents, including agents from other vendors, to query DeepWiki's indexed knowledge as a tool.[25]

Cognition has framed DeepWiki as both a useful free product and a marketing channel: developers exploring an unfamiliar open-source repository through DeepWiki are exposed to Cognition's indexing technology and may become candidates for the paid Devin product when they want similar context applied to private codebases.

Windsurf

See Windsurf (software).

In July 2025, Cognition acquired Windsurf, an agentic IDE developed by Codeium. The acquisition came days after Google completed a $2.4 billion licensing deal in which Windsurf's CEO Varun Mohan, co-founder Douglas Chen, and key research staff left to join Google.[6] That departure left Windsurf's product, brand, remaining employees, and intellectual property available.[7]

Cognition president Russell Kaplan described the deal as coming together over a single weekend, with a first call placed after 5 p.m. on a Friday and a definitive agreement signed Monday morning.[7] Terms were not disclosed.[6][7]

At the time of acquisition, Windsurf had $82 million in annual recurring revenue, more than 350 enterprise customers, and hundreds of thousands of daily active users, with enterprise ARR growing at a rate of roughly double quarter over quarter.[8] All Windsurf employees participated financially in the deal, with vesting cliffs waived for prior work and full acceleration of existing vesting schedules.[8]

Following the acquisition, Scott Wu framed the combination as pairing Devin's autonomous agent capabilities with Windsurf's developer-facing IDE and established enterprise sales infrastructure. The two products addressed different parts of the software development workflow: developers actively coding used Windsurf, while task delegation and autonomous execution ran through Devin. Wu described the combination as "a massive unlock" toward the company's goal of building the future of software engineering.[8]

Within seven weeks of the acquisition closing, combined enterprise ARR grew over 30%.[12] The Windsurf IDE was updated to include full access to the latest Claude models from Anthropic, and Cognition began integrating its proprietary SWE-1.5 model family into the Windsurf environment.[22] On June 2, 2026, Cognition relaunched Windsurf as Devin Desktop, which it called "the next generation of Windsurf."[44]

Windsurf post-acquisition restructuring

The Windsurf integration also became one of the most publicly discussed organizational events in the company's short history. Three weeks after the acquisition closed, Cognition laid off about 30 Windsurf employees and offered the remaining roughly 200 a nine-month severance buyout. Employees who stayed were reportedly required to spend six days a week at the office and work more than 80 hours a week.[26]

The restructuring drew significant external criticism. Critics argued the moves contradicted the inclusive language used at announcement time, when Cognition emphasized that "100% of Windsurf employees" would receive financial compensation from the deal.[26] Wu presented the buyout, worth nine months of salary, as the option for employees unwilling to commit to that pace.[26][31] A separate controversy emerged when Prem Qu Nair, an early Windsurf engineer, publicly stated he had received only about 1% of the share value he had expected from the Google licensing transaction that preceded the Cognition deal, sparking broader debate about how acquihires and partial talent acquisitions distribute proceeds to rank-and-file employees.

Proprietary models

What is the SWE-1 model family?

Alongside the Devin product, Cognition has developed an in-house family of coding models. The line traces back to SWE-1, a model family Windsurf launched publicly in May 2025 before the acquisition,[46] and continued under Cognition with SWE-1.5 in October 2025, SWE-1.6 in April 2026, SWE-1.7 in July 2026, and SWE-2 in September 2026.[22][50][43] The SWE-1.5 release was Cognition's first publicly positioned frontier-size coding model, with several hundred billion parameters and training tailored to long-horizon, multi-step engineering tasks rather than single-turn code completion.[22] SWE-1.5 was trained on a cluster of thousands of NVIDIA GB200 NVL72 chips using end-to-end reinforcement learning in realistic coding environments.[22]

SWE-1.5 was designed jointly with Cerebras Systems to run on Cerebras's wafer-scale inference hardware.[23] On that hardware, Cognition reported inference speeds of up to 950 tokens per second, approximately 13 times faster than Claude Sonnet 4.5 and 6 times faster than Claude Haiku 4.5 on comparable workloads, while achieving what Cognition called near-frontier performance on Scale AI's SWE-Bench Pro.[22] Cognition described the model, its inference system, and the surrounding agent harness as a single co-designed system, a design philosophy that distinguishes it from products built on top of third-party general-purpose models.[22]

Supporting SWE-1.5 are SWE-grep and SWE-grep-mini, two specialized models released in October 2025 for fast parallel context retrieval in Windsurf's Fast Context subagent. Served on Cerebras, SWE-grep runs at over 650 tokens per second and SWE-grep-mini at over 2,800, which Cognition put at 4.5 times and 20 times the speed of Claude Haiku 4.5 respectively.[41]

Deployment in Devin and Windsurf

Cognition deploys its in-house models alongside frontier models from external labs. By default Devin routes between Cognition's own SWE family, Claude models from Anthropic, and other providers depending on the task. Windsurf (relaunched as Devin Desktop in June 2026) likewise offers Cognition's own models, SWE-2 as of November 2026, alongside frontier models from OpenAI, Anthropic, Google, SpaceXAI, and open-source providers.[42][44] This multi-model strategy lets Cognition tune cost and latency on common tasks using its own models while still benefiting from frontier capability when needed.

Strategic partnerships

Microsoft

In May 2024, Cognition partnered with Microsoft to integrate Devin with Microsoft Azure. The partnership included both infrastructure (Azure as Cognition's cloud provider) and commercial collaboration through Microsoft's enterprise sales and engineering teams.[15] At Microsoft Build 2024, Microsoft CTO Kevin Scott described Devin as "an absolutely amazing tool."[47]

Through the Microsoft relationship, Devin was deployed at enterprise customers including Visma, which Microsoft described as a 2 billion euro fintech company based in Norway undergoing large-scale cloud migration. Visma reported a 50% reduction in migration project costs and developer productivity gains of up to 2x in some scenarios.[15]

Cerebras

Cognition's October 2025 SWE-1.5 release also formalized a deep relationship with Cerebras. Cognition served SWE-1.5 in Windsurf through Cerebras at up to 950 tokens per second,[22] and continued the arrangement with SWE-1.6, which Cerebras says runs on Windsurf's fast tier at up to 950 tokens per second, about five times the speed of Cognition's GPU-served free tier.[23]

Anthropic

Though not a formal commercial partnership in the same sense as the Microsoft and Cerebras relationships, Anthropic is a major model provider for Devin and Windsurf. Cognition uses Claude models for tasks that exceed the strengths of its own SWE family, and Windsurf users explicitly choose between Claude Sonnet, Claude Opus, and SWE-1.5 within the IDE. Cognition has described this multi-model posture as practical rather than philosophical, on the view that no single model is optimal for every coding task.

Customers and business metrics

Cognition launched Devin commercially in mid-2024. Annual recurring revenue grew from approximately $1 million in September 2024 to $73 million by June 2025, a 73-fold increase in nine months.[14] Following the Windsurf acquisition in July 2025, which brought Windsurf's approximately $82 million ARR, Cognition said the deal more than doubled its ARR.[12] By the company's first eighteen-month anniversary in early 2026, public commentary and an interview with Scott Wu placed the combined annualized revenue run rate at approximately $445 million.[30] By the time of the May 2026 Series D, Cognition reported run-rate revenue of approximately $492 million, with enterprise usage of Devin growing 50% month over month for the prior six months and up more than tenfold since the start of the year.[34][35] By the September 2026 Series E, run-rate revenue had grown to almost $900 million, and on September 25, 2026 Cognition said it had crossed $1 billion in annualized revenue run rate.[38][39]

Enterprise customer adoption has been one of the most consistently cited drivers of the company's growth. Enterprise customers as of late 2025 included Goldman Sachs, Citigroup, Dell Technologies, Cisco Systems, Ramp, Palantir, Nubank, Mercado Libre, OpenSea, and Curai Health.[14] Cognition has stated that Devin had merged hundreds of thousands of pull requests across customer deployments by November 2025,[32] and told The San Francisco Standard in March 2026 that enterprise usage had grown roughly 80-fold over the previous year.[31] At the time of the May 2026 Series D, the company named additional enterprise customers including Mercedes-Benz, NASA, and Santander.[34][35]

Goldman Sachs

In July 2025, Goldman Sachs announced it was piloting Devin alongside its approximately 12,000 internal developers, with chief information officer Marco Argenti telling CNBC that "Devin is going to be like our new employee" and describing it as part of his vision for a "hybrid workforce."[27] IBM Think, which headlined Devin as "Employee #1" in that workforce, quoted an IBM executive's hypothetical that a 20% productivity gain would let Goldman's 12,000 technologists operate like 14,400.[27]

Citigroup

Citigroup announced it was rolling out Cognition's Devin agent to its roughly 40,000 developers.[28] The bank's stated initial use cases were straightforward, well-scoped engineering tasks such as software patches and library upgrades, with plans to expand to more complex assignments over time.[28] The Citi rollout, combined with the Goldman pilot, made financial services one of the most visible verticals for the product.

Other notable deployments

One documented case study involved Linktree, the social media link-in-bio platform. In February 2025, Devin authored approximately 300 pull requests in one month, of which roughly 100 were merged.[14] A Nubank case study reported an 8-fold improvement in engineering efficiency and 20-fold cost savings on a monolithic codebase refactoring project.[14] Visma's large-scale Azure migration, mentioned above, was another frequently cited engagement.

What does Devin cost?

As of November 2026, Cognition's plans are Free ($0), Pro ($20 per month), Max ($200 per month), Teams ($80 per month for the team plan plus $40 per month per full developer seat, up to 200 users) and custom-priced Enterprise, which adds options such as VPC deployment, SSO, and dedicated account support. Paid plans include usage quotas that refresh daily and weekly, with extra usage billed at API pricing.[42]

Company culture

Cognition has built a public identity around an unusually intense work culture even by Silicon Valley standards. Wu has repeatedly described the company as a place where the founders and most employees expect to work long hours, six days a week, with significant time spent together in person at the San Francisco office.[31] Internal communications publicized after the Windsurf acquisition included a statement from Wu that the company does not "believe in work-life balance" and a follow-up clarification framing intensity as opt-in rather than mandatory.[26]

The practical mechanisms of the culture include extended on-site hours, an in-office gym, private chefs, and an in-house barista, with many employees living in apartments around the corner from the office.[31] Several profiles of the company by The San Francisco Standard and others have described employees living near or staying at company-managed houses, working out of a single open floor, and adopting a research-lab style of collaboration in which research, infrastructure, and product engineering coexist in the same physical space.[31]

Defenders of the culture argue that it reflects the personal preferences of a relatively small, mission-driven team and that financial outcomes for participating employees have outpaced more conventional companies. Critics, including some former employees, have pushed back on the broader normalization of 80-hour weeks in frontier AI startups, arguing that the practice excludes caregivers and contributes to burnout. The Windsurf restructuring discussed above became the most visible flashpoint for these debates.

Reception

The March 2024 announcement drew significant attention across the software industry. The demo video showing Devin completing a full-stack programming task went viral, and industry figures including the Collison brothers and Elad Gil publicly backed the company. Within weeks, the company had secured $175 million in additional funding and a $2 billion valuation.[14]

Criticism of the launch appeared quickly. The YouTube channel Internet of Bugs published a detailed breakdown of the original demo video, arguing that the Upwork task Devin appeared to complete had been misrepresented and that the autonomous completion shown relied on conditions not disclosed in the promotional material.[16] These critiques, summarized by The Register, raised questions about whether Cognition had overstated Devin's real-world capabilities during the launch.[18]

An evaluation by Answer.AI, an applied AI research lab, attempted 20 tasks using Devin over a month and published its findings in January 2025. The evaluation found 14 failures, 3 inconclusive results, and 3 successes, a success rate of 15%.[17] The evaluators noted that Devin's autonomous nature became a liability in some cases, where it continued pursuing unworkable approaches for extended periods rather than recognizing when a task was structurally blocked.[17]

The SWE-bench score of 13.86% was itself questioned in some quarters. The figure still meant Devin failed on roughly 86% of the tasks it was evaluated on, a random 25% subset of the benchmark.[1]

By 2025, Cognition's public messaging shifted to emphasize Devin as a tool for augmenting engineering teams rather than replacing individual engineers. Internal performance metrics from November 2025 showed a pull request merge rate of approximately 67%, compared to 34% in earlier versions, alongside reported improvements in speed and compute efficiency.[32]

How does Devin compare with other AI coding agents?

Cognition operates in a crowded market for AI coding tools, competing with products from Anthropic, OpenAI, Microsoft, and several well-funded startups including Cursor. The table below reflects the products largely as of mid-2025; models and features change quickly.

Devin (Cognition)Claude Code (Anthropic)Codex (OpenAI)GitHub Copilot (Microsoft)
ArchitectureCloud-based autonomous agentTerminal-based agentic assistantCloud-based asynchronous agentIDE plugin
Where code runsCognition cloud infrastructureLocal machine (developer's environment)OpenAI-managed cloud containersLocal environment
Primary interfaceWeb app + IDE (Windsurf)Terminal / shellChatGPT web interfaceIDE extension
Underlying modelCognition's SWE models plus third-party frontier modelsAnthropic Claude modelscodex-1 (o3-based) at launch, later GPT-5-Codex modelsMultiple models from OpenAI, Anthropic, and Google
ParallelismYes (MultiDevin, up to 10 agents)Agent Teams (lead + sub-agents)Independent parallel agentsNo
Starting price$20/monthIncluded with Claude ProIncluded with ChatGPT Plus$10/month (individual)
SWE-bench (representative)13.86% (March 2024, original Devin)70.3% (Claude 3.7 Sonnet model with custom scaffold, SWE-bench Verified, February 2025)[45]High (o3-based, specific figure varies by report)N/A

Expanded article table

The most significant structural difference between Devin and terminal-based tools like Claude Code is where execution happens. Claude Code runs in the developer's own environment with direct access to local files, while Devin and Codex (OpenAI) run in cloud-controlled sandboxes, submitting results as pull requests. Developers who need tight control over their local environment tend to prefer terminal-based agents. Teams that want fully asynchronous delegation without monitoring an active shell session tend toward cloud-based agents.

By acquiring Windsurf, Cognition bridged this gap. The Windsurf IDE provides a developer-facing local environment with a native coding experience, while Devin handles task delegation in parallel. Cognition said the deal gave it "the complete product suite for AI coding": an IDE and an autonomous agent.[12]

Cursor (code editor) from Anysphere, which Walden Yan briefly worked on before founding Cognition, had crossed $1 billion in annualized revenue and reached a $29.3 billion valuation by November 2025,[48] and SpaceX completed its acquisition of Anysphere on August 14, 2026.[49] Cursor's growth illustrated the scale of demand for AI-native development tools, and Cognition's acquisition of Windsurf was partly a response to that competitive dynamic.

Limitations

Autonomous agents that run unsupervised on cloud infrastructure create distinct failure modes compared to interactive tools. When Devin misunderstands a task or encounters an unexpected environment state, it can continue working in the wrong direction for extended periods before surfacing an error. Answer.AI's 2024 evaluation documented cases where Devin spent significant time pursuing approaches that were structurally blocked rather than stopping to ask for clarification.[17]

The confidence scoring system introduced in Devin 2.1 addresses part of this problem by having Devin express uncertainty upfront and ask questions before proceeding.[10] But the accuracy of the confidence estimates depends on how well Devin has indexed the relevant codebase, and large or poorly documented repositories remain harder for the system to reason about.

Security and compliance are ongoing concerns for enterprise adoption. Cognition offers VPC deployment to keep customer code inside their own network perimeter, but any cloud-based execution model requires customers to grant an external system significant access to their codebase and infrastructure. Organizations in regulated industries have been slower to adopt autonomous agents for this reason.

The compute costs of running multi-step autonomous agents at scale are substantially higher than those of interactive code completion tools. Usage-based billing through ACUs means that a task that takes longer than expected can consume more compute than anticipated, which has been noted as a friction point in enterprise procurement.

SWE-bench scores across the industry rose rapidly throughout 2024 and 2025, with multiple competing systems eventually surpassing Devin's initial 13.86% benchmark by large margins. Anthropic's Claude 3.7 Sonnet reached 70.3% on SWE-bench Verified with a custom scaffold in February 2025,[45] and newer models from Anthropic, OpenAI, and Google continued to push the benchmark higher. Cognition's SWE-1.5 family closed some of that gap, but SWE-bench performance does not directly predict success on the diverse range of tasks that appear in production engineering work.

Continue reading

More from See also

Topics:AI AgentsAI Code GenerationAI CompaniesDeveloper Tools

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7Cognition AI. "Introducing Devin, the first AI software engineer." cognition.ai/blog/introducing-devin (March 12, 2024).
  2. ^1 ^2 ^3 ^4Scott Wu. Wikipedia article. en.wikipedia.org/wiki/Scott_Wu.
  3. ^1 ^2 ^3 ^4Cognition AI. Wikipedia article. en.wikipedia.org/wiki/Cognition_AI.
  4. ^TechCrunch. "Cognition AI defies turbulence with a $400M raise at $10.2B valuation." techcrunch.com (September 8, 2025).
  5. ^Bloomberg. "AI Coding Firm Cognition in Funding Talks at $25 Billion Value." bloomberg.com (April 23, 2026).
  6. ^1 ^2 ^3CNBC. "Cognition to buy AI startup Windsurf days after Google poached CEO in $2.4 billion licensing deal." cnbc.com (July 14, 2025).
  7. ^1 ^2 ^3 ^4TechCrunch. "Cognition, maker of the AI coding agent Devin, acquires Windsurf." techcrunch.com (July 14, 2025).
  8. ^1 ^2 ^3Cognition AI. "Cognition's acquisition of Windsurf." cognition.ai/blog/windsurf (July 14, 2025).
  9. ^1 ^2Cognition AI. "Devin 2.0." cognition.ai/blog/devin-2 (April 2025).
  10. ^1 ^2 ^3 ^4Cognition AI. "Devin 2.1." cognition.ai/blog/devin-2-1 (2025).
  11. ^1 ^2Cognition AI. "Introducing Devin 2.2." cognition.ai/blog/introducing-devin-2-2 (February 24, 2026).
  12. ^1 ^2 ^3 ^4 ^5 ^6 ^7Cognition AI. "Funding, growth, and the next frontier of AI coding agents." cognition.ai/blog/funding-growth-and-the-next-frontier-of-ai-coding-agents (September 2025).
  13. ^1 ^2 ^3VentureBeat. "Devin 2.0 is here: Cognition slashes price of AI software engineer to $20 per month from $500." venturebeat.com (April 2025).
  14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16Contrary Research. "Cognition Business Breakdown & Founding Story." research.contrary.com/company/cognition (February 2026).
  15. ^1 ^2Microsoft Customer Stories. "Cognition developed an AI engineer that transforms enterprise software with Microsoft Azure." microsoft.com/en/customers/story/24262-cognition-ai-azure (May 29, 2025).
  16. ^Internet of Bugs. YouTube video analysis of Cognition's Devin demo (2024).
  17. ^1 ^2 ^3Answer.AI evaluation of Devin autonomous agent performance (2024).
  18. ^The Register. "'First AI software engineer' is bad at its job." theregister.com (January 23, 2025).
  19. Gennady Korotkevich (@que_tourist). Post on X regarding ryanbAI IOI gold medal. x.com (July 2025).
  20. SiliconANGLE. "Cognition, creator of the AI software engineer Devin, in talks to raise 'hundreds of millions' at $25B valuation." siliconangle.com (April 23, 2026).
  21. Lenny's Newsletter. "Inside Devin: The world's first autonomous AI engineer." lennysnewsletter.com (2024).
  22. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Cognition AI. "Introducing SWE-1.5: Our Fast Agent Model." cognition.ai/blog/swe-1-5 (October 29, 2025).
  23. ^1 ^2Cerebras Systems. "Case Study - Cognition x Cerebras." cerebras.ai/blog/case-study-cognition-x-cerebras (October 2025).
  24. ^1 ^2 ^3Cognition AI. "DeepWiki: AI docs for any repo." cognition.ai/blog/deepwiki (May 5, 2025).
  25. ^Cognition AI. "The DeepWiki MCP Server." cognition.ai/blog/deepwiki-mcp-server (2025).
  26. ^1 ^2 ^3 ^4TechCrunch. "Three weeks after acquiring Windsurf, Cognition offers staff the exit door." techcrunch.com (August 5, 2025).
  27. ^1 ^2IBM Think. "Meet Devin the AI Software Engineer, Employee #1 in Goldman Sachs' 'Hybrid Workforce'." ibm.com (2025).
  28. ^1 ^2American Banker. "Citi is rolling out agentic AI to its 40,000 developers." americanbanker.com (2025).
  29. Cheeky Pint (Stripe). "Cognition CEO Scott Wu on acquiring Windsurf, AI replacing engineers, and the Moneyball-ification of everything." cheekypint.substack.com (August 2025).
  30. ^Colossus. "The Wu Tapes: Q&A with Cognition's Scott Wu." colossus.com/article/scott-wu-tapes-cognition (May 2026).
  31. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8The San Francisco Standard. "Inside the grind: The SF startup racing to build an AI software engineer." sfstandard.com (March 24, 2026).
  32. ^1 ^2Cognition AI. "Devin's 2025 Performance Review: Learnings From 18 Months of Agents At Work." cognition.ai/blog/devin-annual-performance-review-2025 (November 14, 2025).
  33. ^1 ^2 ^3Hg Capital Orbit Podcast. "A golden age of software engineering with Russell Kaplan, Founding President of Cognition." hgcapital.com (2025).
  34. ^1 ^2 ^3 ^4 ^5 ^6TechCrunch. "AI coding startup Cognition raises $1B at $25B pre-money valuation." techcrunch.com/2026/05/27/ai-coding-startup-cognition-raises-1b-at-25b-pre-money-valuation (May 27, 2026).
  35. ^1 ^2 ^3 ^4 ^5Cognition AI. "More Devins in More Places." cognition.ai/blog/series-d (May 27, 2026).
  36. ^1 ^2Bloomberg. "AI Coding Startup Cognition Raises $1 Billion at $26 Billion Value." bloomberg.com/news/articles/2026-05-27/ai-coding-startup-cognition-raises-1-billion-at-26-billion-value (May 27, 2026).
  37. ^Benzinga. "Cognition Raises $1 Billion At $26 Billion Valuation, Led By Lux, General Catalyst, 8VC." benzinga.com (May 28, 2026).
  38. ^1 ^2 ^3 ^4 ^5Cognition AI. "Do it all with Devin: Announcing our Series E." cognition.ai/blog/series-e (September 8, 2026).
  39. ^1 ^2Cognition AI. "Cognition Crosses $1B in Annualized Revenue Run Rate." cognition.ai/blog/1b-run-rate (September 25, 2026).
  40. ^Bloomberg. "AI Startup Cognition in New Funding Talks at $40 Billion Value." news.bgov.com/private-equity/ai-startup-cognition-in-new-funding-talks-at-40-billion-value (August 12, 2026).
  41. ^Cognition AI. "Introducing SWE-grep and SWE-grep-mini: RL for Multi-Turn, Fast Context Retrieval." cognition.ai/blog/swe-grep (October 16, 2025).
  42. ^1 ^2Devin. "Plans and Pricing." devin.ai/pricing (accessed November 2026).
  43. ^Cognition AI. "Introducing SWE-2: Pushing the Pareto Frontier." cognition.ai/blog/swe-2 (September 10, 2026).
  44. ^1 ^2 ^3Cognition AI. "Introducing Devin Desktop." cognition.ai/blog/introducing-devin-desktop (June 2, 2026).
  45. ^1 ^2Anthropic. "Claude 3.7 Sonnet and Claude Code." anthropic.com/news/claude-3-7-sonnet (February 24, 2025).
  46. ^Windsurf. "SWE-1: Our First Frontier Models." windsurf.com/blog/windsurf-wave-9-swe-1 (May 15, 2025).
  47. ^Microsoft. Kevin Scott keynote transcript, Microsoft Build 2024. news.microsoft.com/wp-content/uploads/prod/2024/05/KevinScott_transcript_KEY01_Build20204.pdf (May 21, 2024).
  48. ^CNBC. "AI startup Cursor raises $2.3 billion funding round at $29.3 billion valuation." cnbc.com/2025/11/13/cursor-ai-startup-funding-round-valuation.html (November 13, 2025).
  49. ^Cursor. "Cursor is now a part of SpaceX." cursor.com/blog/joining-spacex (August 14, 2026).
  50. ^Cognition AI. "Introducing SWE 1.6: Improving Model UX." cognition.ai/blog/swe-1-6 (April 7, 2026).

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

13 revisions · v13 · 6,149 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check of the full article against its cited sources (165 claims, 44 sources fetched); 14 material + 50 minor fixed

Cite this page: AI Wiki. "Cognition AI." aiwiki.ai, updated 28 Oct 2026, fact-checked 2 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/cognition_ai

Edit article