Signals Inbox·July 20, 2026·AI Agents

Devin vs Claude Code: who is winning?

Claude Code is winning today through greater revenue, adoption and distribution, but Devin still has an edge when companies want to delegate long-running engineering work rather than merely accelerate developers.

Q1Why do people keep comparing Devin and Claude Code?

Devin and Claude Code are compared so often because both now want developers to hand over complete engineering tasks, even though they arrived there from opposite directions.

When Cognition launched Devin in 2024, the product was presented as an autonomous software engineer with its own browser, terminal and cloud computer. The promise was delegation: assign a task, leave Devin working and return to a pull request.

Claude Code began closer to the developer. It worked inside the terminal, explored repositories, edited files and executed commands while the user followed the process. Anthropic has since added background sessions, remote execution, scheduled routines and workflows that coordinate large groups of subagents.

Cognition has moved toward interactive development at the same time. After acquiring Windsurf, it connected Devin Cloud with an IDE, a command-line tool, code review and codebase search. Developers can now begin a task locally and transfer the same session to a cloud agent when it becomes too long.

The products increasingly compete for the same position inside an engineering team. Both want to understand the codebase, plan the change, write the code, run the tests, respond to reviews and get the work merged. Claude Code still feels like a powerful collaborator that has learned to work alone. Devin feels like a delegated worker that has learned to collaborate.

Q2Why is it so hard to tell whether Devin or Claude Code is winning?

Devin and Claude Code are hard to rank because Claude Code has built the larger business, while Devin remains more closely associated with fully delegated engineering work.

The disclosed numbers also measure different things. Anthropic has published revenue specifically for Claude Code. Cognition’s latest revenue figure covers the combined company after the Windsurf acquisition, including the IDE, cloud agents, review tools and other products.

Usage is equally difficult to compare. Claude Code often produces commits from a developer’s local environment. Devin frequently works in the cloud and returns a pull request. A recent academic study of more than 180 million repositories found that commit-based and pull-request-based detection methods captured largely separate populations of coding-agent users.

Product quality also changes sharply by task. Another study, covering 7,156 agent-created pull requests, found an 82.1% acceptance rate for documentation changes and 66.1% for new features. The type of work sometimes created a larger performance gap than the choice of agent.

So the comparison needs several scoreboards: revenue, growth, adoption, accepted output, autonomy, enterprise results, distribution and cost. The final answer depends on which of those currently carry the most weight.

Q3Is Claude Code or Devin making more money now?

Claude Code currently makes far more money than Devin, with the disclosed gap sitting above five to one.

Anthropic said in a February 2026 funding announcement that Claude Code had passed $2.5 billion in run-rate revenue. The figure covered Claude Code itself and had more than doubled since the beginning of the year. The product had crossed $1 billion only a few months earlier.

Cognition reported a $492 million annualized revenue run rate in May 2026. Dividing the two figures gives Claude Code a lead of at least 5.1 times, even though Cognition’s number was published three months later.

The real gap is probably wider. Cognition’s figure includes the whole company, whereas Anthropic isolated Claude Code. Cognition also reached part of its current scale through the Windsurf acquisition. Devin had $73 million of annual recurring revenue before the deal, and Cognition said Windsurf more than doubled the combined company’s ARR.

Cognition subsequently increased enterprise ARR by more than 30% during the first seven weeks after the acquisition. Strong execution, certainly, but it does not change the present order of magnitude.

Claude Code and Cognition revenue comparison

Q4Is Devin growing faster than Claude Code right now?

Devin is currently growing faster in percentage terms, while Claude Code is adding much more revenue in absolute dollars.

Cognition said enterprise usage rose more than tenfold between the beginning of 2026 and May. Its combined run-rate revenue reached $492 million, compared with Devin’s $73 million before the Windsurf acquisition eleven months earlier.

The acquisition clearly helped. Cognition said fewer than 5% of Devin and Windsurf’s enterprise customers overlapped, which meant the company combined two largely separate customer bases. Enterprise ARR then climbed more than 30% within seven weeks.

Claude Code has grown rapidly from a much larger starting point. Anthropic reported a run rate above $500 million in September 2025, $1 billion in November and more than $2.5 billion by February 2026. Weekly active users also doubled during the opening weeks of the year.

That sequence represents more than $1.5 billion of additional annualized revenue in roughly five months. The increase alone was around three times Cognition’s entire May run rate.

Q5Are developers using Claude Code or Devin more today?

Claude Code is used more widely than Devin today, and the available adoption data leaves little room for doubt.

Anthropic recently studied roughly 400,000 Claude Code sessions from about 235,000 consenting users. The sample did not cover every user or interface, but it provides a substantial view of real activity that Cognition has not matched with comparable Devin data.

The strongest outside evidence comes from an academic census of more than 180 million public Git repositories. Researchers identified 886,122 Claude Code commits across 17,295 projects. Another 21,078 projects contained Claude Code configuration files without an attributable Claude commit, placing it first among the agents detected through commit-based methods.

That research probably undercounts Devin because cloud agents commonly submit pull requests instead of local commits. Cognition said in its 2025 annual review that Devin had merged hundreds of thousands of pull requests and was used by thousands of companies. It did not disclose active users, weekly sessions or retention.

Windsurf brought hundreds of thousands of daily users into Cognition, but those people cannot automatically be counted as Devin Cloud users. On the metrics we can compare properly, Claude Code has the clear adoption lead.

Q6Does Claude Code or Devin get more code accepted?

Claude Code currently has stronger evidence for accepted feature work, although Devin improved more consistently during the period studied.

The best public comparison we found examined 7,156 pull requests created by Claude Code, Devin, Cursor, GitHub Copilot and OpenAI Codex. Researchers grouped the work by task and checked whether project maintainers accepted it.

Claude Code led documentation changes with a 92.3% acceptance rate and new features with 72.6%. The feature result deserves more weight because both products are trying to handle substantial engineering tasks involving several files and meaningful product decisions.

Devin did not lead a major category in the published results, but its acceptance rate improved by 0.77 percentage points per week across the 32-week study. It was the only agent with such a consistent positive trend.

The comparison also showed that tool choice explains only part of the outcome. Cursor led bug fixes with an 80.4% acceptance rate, while Codex performed well across all nine task groups. Agents still need to be matched with the kind of work they handle best.

Real-world pull-request acceptance results

Q7Is Devin still more autonomous than Claude Code?

Devin still provides the cleaner autonomous workflow, although Claude Code has recently closed much of the technical gap.

Devin can assign separate cloud machines to several agents, keep them running after the developer leaves and combine their work later. Developers can also begin locally in Devin Desktop or the terminal, then send the session to the cloud when the task becomes too long.

Claude Code can now plan large projects, launch tens or hundreds of subagents, run background tasks and verify their output. The product has moved well beyond its original terminal-assistant role.

Anthropic’s usage research still showed a fairly interactive pattern. The median Claude Code turn lasted around 45 seconds, while only the longest 0.1% exceeded roughly 45 minutes by early 2026.

A separate Anthropic study found that people made around 70% of planning decisions, with Claude handling most execution decisions. Experienced users gave Claude more freedom, yet they also interrupted it more frequently when the work started moving in the wrong direction.

Devin remains easier to use as a delegated cloud worker. Its current advantage comes mainly from how Cognition packages the workflow, since Claude Code now possesses many of the same underlying capabilities.

Q8Is Devin or Claude Code better for the work engineering teams keep postponing?

Devin currently has the better published record on migrations, security backlogs and other work that engineering teams repeatedly delay.

Mercedes-Benz provides Cognition’s strongest example. During a four-week pilot, Devin analyzed more than 200,000 lines of COBOL and reduced one project’s estimated duration from eight months to eight days. Mercedes-Benz later expanded the relationship across legacy modernization, cloud development and logistics.

Itaú offers another useful case. After eleven months of deployment inside a bank with nearly 17,000 engineers, Cognition reported that Devin completed migrations five to six times faster, automatically fixed 70% of identified static-analysis vulnerabilities and doubled test coverage.

Cognition has lately turned that pattern into a dedicated product. Devin Security Swarm scans a codebase, tests whether vulnerabilities can actually be exploited and prepares remediation pull requests. Cognition also runs structured programs designed to clear existing security backlogs before setting up continuous remediation.

Claude Code can handle similar projects. Anthropic says its dynamic workflows can complete migrations across hundreds of thousands of lines, and Claude Code Security scans repositories and proposes fixes. Anthropic also works with companies such as DXC on modernization and cybersecurity.

Cognition currently has the better concrete before-and-after evidence. Anthropic has shown broad capability, but fewer detailed customer outcomes for this exact kind of postponed work.

Q9Does Claude Code or Devin have the stronger enterprise customers?

Claude Code currently has stronger enterprise customer economics, while Devin’s case studies reveal more about the work customers are completing.

Anthropic has named Netflix, Spotify, KPMG, L’Oréal and Salesforce as Claude Code users. Salesforce has deployed the product across its global engineering organization, while Netflix has discussed using Claude-based development tools across more than 3,000 developers.

The spending evidence is stronger than the customer logos. Anthropic said enterprise customers generated more than half of Claude Code’s revenue. Based on the product figure discussed earlier, that implied more than $1.25 billion of enterprise run-rate revenue at the time of disclosure.

Cognition’s customer list includes Goldman Sachs, Citi, Mercedes-Benz, Dell, Santander, Elevance Health, the US Army and the US Navy. Technology companies such as Exa, Modal, Eight Sleep and OpenRouter also use Devin.

Cognition recently reached FedRAMP High In-Process status, according to its federal deployment announcement. Full authorization still has to be completed, but the process shows that Cognition is building for government and defence environments with demanding security requirements.

Claude Code holds the stronger position here.

Q10Can Devin match Claude Code’s distribution?

Devin cannot currently match the number of ways Claude Code reaches engineering teams.

Claude Code is distributed through individual subscriptions, enterprise contracts, Anthropic’s API and major cloud platforms. Developers can use it through terminals, browsers, desktop apps, GitHub, GitLab, VS Code, JetBrains, Xcode and Slack.

Two recent integrations show how powerful that reach can become. Xcode 26.3 uses the Claude Agent SDK directly inside Apple’s development environment, while Claude Tag lets employees assign coding tasks from Slack without opening a terminal.

Devin is available through the web, Devin Desktop, Slack, Linear, Jira, APIs and the command line. The Windsurf acquisition also brought Cognition an established IDE audience and enterprise sales operation. Partnerships with Infosys and Cognizant give Devin another route into large companies.

Anthropic still has the wider funnel.

Q11Is Devin or Claude Code cheaper for serious use today?

Devin currently has the more promising cost architecture, although neither company has proved that it is cheaper across normal production workloads.

The entry prices are similar. Claude Pro and Devin Pro both cost $20 per month, while heavy individual users can pay $200 for Claude Max 20x or Devin Max.

Subscription prices reveal little about the real economics. Cost depends on model choice, context size, task length, failed attempts, retries and the percentage of generated work that reaches production.

Cognition is attacking those variable costs through Devin Fusion. The system pairs a powerful main agent with a cheaper sidekick for mechanical work such as repository exploration, testing and formatting. It can also switch models during a session without rebuilding the entire context.

According to Cognition’s FrontierCode benchmark, Fusion produced near-frontier performance at 35% lower cost than frontier models alone. Cognition also said its automated router drove 88% of merged Fusion pull requests among participating employees.

Those figures are internal and need broader validation. Even so, Devin can route work across models from Anthropic, OpenAI, Google, Cognition and open-source developers. That gives Cognition more ways to cut costs than Claude Code’s Anthropic-only stack.

Q12Does Devin or Claude Code handle code review better?

Devin currently has the more complete workflow for reviewing and repairing agent-generated code.

Devin Review reorganizes large diffs, explains changes and highlights possible problems. Devin can then respond to review comments and update the pull request automatically.

Cognition now adds a Devin Review link to every Devin-generated pull request used internally. The company said it merged 659 Devin pull requests during one recent week, compared with 154 in its best week of 2025. A fourfold rise in generated work makes faster review essential.

Cognition’s productivity system also tries to exclude clearly unproductive sessions. Pull requests closed without being merged are treated conservatively, preventing failed agent activity from being presented as useful output.

Claude Code has strong debugging, self-checking and permission controls. Its parallel workflows can verify the work of subagents before presenting the result. Devin currently goes further in joining review, automatic repair and output measurement inside one product.

Q13Does Claude Code or Devin have stronger proof of productivity?

Claude Code currently has broader productivity evidence, while Devin has made the more aggressive promise to prove its financial value.

A recent study followed tens of thousands of Microsoft engineers during an internal rollout of Claude Code and GitHub Copilot CLI. Engineers who adopted the tools merged about 24% more pull requests than researchers estimated they otherwise would have over four months.

The study cannot separate Claude Code’s contribution from Copilot CLI, and pull-request volume does not measure quality directly. Its scale still makes it one of the more useful independent tests of coding-agent productivity.

Anthropic’s internal research found larger effects. Employees reported productivity gains of around 50%, while merged pull requests per engineer per day increased by 67%. Those figures come from a company where employees are unusually comfortable with AI tools, so they should not be projected onto every engineering organization.

Cognition measures a different outcome: the amount of human engineering time represented by Devin’s completed work. Its estimator reviews the prompt, execution trace, code changes, pull request and codebase context. Cognition trained the system on 258 completed sessions from 126 enterprise users and reported a correlation of 0.74 with users’ time estimates.

The weak point is fairly clear: users are estimating how long hypothetical manual work would have taken. Cognition tries to reduce inflation by excluding agent detours, setup failures and information already provided in the prompt.

Cognition has backed the system with an AI Productivity Guarantee. When estimated engineering value falls below what a customer paid, the company provides additional usage credits, with total coverage of up to $10 million.

Claude Code has more evidence supporting productivity today. Devin is asking the tougher commercial question: did the completed work create more value than the agent cost?

Q14Is Cognition or Anthropic shipping faster in AI coding now?

Cognition has recently shipped the more coherent coding workflow, while Anthropic has released more capability overall.

Devin Desktop, Devin CLI, managed cloud agents and scheduled sessions all support the same operating model. Developers can move between writing code themselves, working with a local agent and supervising remote agents.

Cognition has added automated review fixes, native Windows environments, security swarms, multi-model routing, productivity measurement and its own SWE models around that workflow. The releases fit together and gradually make Devin feel like a complete engineering platform.

Anthropic has expanded Claude Code through dynamic workflows, background routines, computer use, Slack delegation, deeper Xcode support and repeated model upgrades. Improvements to Claude can also reach Claude Code immediately, without waiting for another company to integrate the model.

Cognition currently has the clearer coding-specific roadmap. Anthropic is shipping across a much larger platform and can distribute each advance much faster.

Q15Does Anthropic’s model ownership beat Devin’s model independence?

Anthropic’s ownership of Claude currently gives Claude Code the stronger position, although Devin’s model independence is becoming more valuable.

Anthropic controls the model, the agent harness and much of the infrastructure behind Claude Code. Improvements in coding, tool use, context handling or long-running reasoning can reach the product immediately.

Anthropic can also offer different Claude models for different workloads while keeping users inside the same environment. Sonnet handles cheaper everyday tasks, while Opus targets difficult and longer-running work.

Devin can route tasks among Anthropic, OpenAI, Google, Cognition and open-source models. Cognition says it evaluates models across more than 100 types of engineering work and selects them according to capability, speed and cost.

The arrangement creates an unusual dependency. Every major improvement to Claude can strengthen both Claude Code and Devin. Cognition is reducing that exposure through Fusion and its own SWE models, but Anthropic still controls the stronger combination of intelligence and distribution.

Devin’s model-independent approach becomes much more powerful if several model providers remain closely matched and inference prices continue falling.

Q16What can Devin do that Claude Code cannot easily copy?

Devin’s hardest-to-copy advantage is the operating system Cognition is building around delegated engineering work.

Individual features such as remote environments, parallel agents, recurring tasks and code review can be reproduced. Anthropic already offers several of them.

Cognition’s deeper asset comes from combining cloud execution, an IDE, codebase search, review, proprietary models and productivity measurement. A Devin session can contain the original request, repository exploration, failed approaches, tests, code changes, review comments and the final merge outcome.

That history can improve task routing, cost estimates, confidence scoring and model training. Local work inside Devin Desktop also shows how developers explore and plan, while cloud sessions reveal how autonomous agents execute those plans.

Cognition uses Devin heavily to build Devin. Its jump to 659 merged Devin pull requests in one week creates a dense internal feedback loop across planning, implementation, testing and review.

The advantage is still developing. Anthropic, OpenAI, Cursor and GitHub also observe enormous volumes of engineering activity. Devin becomes harder to replace when customers build recurring jobs, security workflows, organizational knowledge and productivity reporting around Cognition’s platform.

Q17Can Cognition keep up with Anthropic financially?

Anthropic has vastly more financial firepower than Cognition, despite Cognition’s recent billion-dollar round.

Cognition raised more than $1 billion at a $26 billion post-money valuation in May 2026. Anthropic soon raised $65 billion at a $965 billion valuation. Anthropic’s round was roughly 65 times larger, while its valuation was around 37 times higher.

Anthropic must spend heavily on model training, compute infrastructure, safety and products outside coding. Cognition can direct a much larger share of its capital toward software engineering.

The valuation multiples show how much investors expect from Cognition. Its $26 billion valuation equals roughly 53 times the company’s $492 million run rate. Anthropic’s valuation was around 21 times its reported companywide run rate.

Cognition has enough money to remain a serious competitor. Anthropic has far more room to subsidize usage, fund distribution and absorb failed product bets.

Q18Are Devin and Claude Code actually leading the whole AI coding market?

Claude Code is ahead of Devin, but both products face competitors operating at similar or greater scale.

Cursor passed $2 billion in recurring revenue after doubling within three months, according to the company’s March 2026 announcement. Cursor also says more than 70% of the Fortune 500 use its platform to deploy or manage coding agents.

GitHub Copilot has the largest installed base. Microsoft reported more than 20 million users in its latest annual report, while almost 140,000 organizations were using the product by the following quarter.

OpenAI Codex has also expanded rapidly. OpenAI reported more than five million weekly users after its desktop-app launch, alongside parallel agents, cloud execution, mobile supervision and broader knowledge-work features.

These products are converging. Cursor supports parallel local and cloud agents. Copilot offers cloud agents, command-line workflows and code review. Codex has become a control center for multiple agents.

Claude Code belongs in the leading group. Devin remains a fast-growing challenger with a distinctive enterprise workflow. Neither company has secured the category, and the eventual winner may need Claude-level model quality, GitHub-level distribution and Devin-style delegation in the same product.

Q19Who is winning between Devin and Claude Code right now?

Claude Code is winning the Devin vs Claude Code race today, with a clear overall lead despite Devin’s stronger position in delegated enterprise work.

Claude Code leads on the criteria carrying the most weight now: commercial scale, broad adoption, distribution, enterprise spending, model control and access to capital. Those advantages reinforce one another. Better models bring more developers, developer activity attracts enterprises, and enterprise revenue funds further product and infrastructure investment.

Devin leads when the work resembles managing a team of cloud engineers. Cognition provides the cleaner workflow for handing several agents long-running jobs, clearing legacy backlogs, measuring productive output and moving work between local and remote environments.

The autonomy gap has narrowed quickly. Claude Code can now run large groups of subagents, manage background sessions and complete substantial codebase migrations. Cognition can no longer rely on autonomy alone to separate Devin from the market.

Devin’s rapid growth, multi-model routing and enterprise case studies keep Cognition firmly in the race. The company now needs to disclose Devin-specific revenue, active users, retention and expansion inside major customers. Independent evidence supporting Fusion’s cost claims and Cognition’s productivity estimates would also strengthen its case.

Anthropic needs to prove that Claude Code’s commercial momentum can survive the growth of Cursor, Codex and Copilot. More independent evidence around long-running autonomous work would also show whether Claude Code is creating maintainable software or simply producing more code for humans to inspect.

Our verdict is firm: Claude Code currently leads the larger and more valuable market. Devin is ahead in a narrower version of the future where developers supervise fleets of cloud agents. Cognition may have identified that workflow earlier, but Claude Code already has the scale and technical capability to compete for it.

Devin vs Claude Code scorecard, July 2026