Is Cognition really worth $26B today?
In our AI code assistant market deck, you will find everything you need to understand the market
SUMMARY
Is Cognition really worth $26B today? Yes, but only if Devin becomes software labor infrastructure, not just another AI coding tool.
The valuation is not random. Cognition reached a reported $26B valuation in May 2026 after moving through five valuation checkpoints: about $350M in early 2024, $2B in April 2024, nearly $4B in March 2025, $10.2B in September 2025, and $26B in May 2026.
The cleanest support for the valuation is revenue velocity. Cognition disclosed $492M of run-rate revenue in May 2026, after Devin had grown from $1M ARR in September 2024 to $73M ARR in June 2025.
The cleanest concern is the multiple. At $26B of valuation and $492M of run-rate revenue, Cognition trades at roughly 53x revenue, far above normal public SaaS and even above most premium software companies.
Cognition looks expensive even versus Cursor. Cursor’s last confirmed $29.3B valuation and $1B+ annualized revenue imply roughly 29x revenue or lower, while Cognition sits near 53x.
The valuation starts to look more reasonable only if Cognition reaches roughly $1.5B to $2B of ARR quickly. At that level, the multiple compresses into a premium but less extreme range.
The core bull case is that Cognition is not selling developer productivity software. It is trying to sell a new unit of engineering capacity: AI agents that can plan, write, test, and ship work inside real company codebases.
The Windsurf acquisition matters because it changed Cognition from a single-agent company into a broader coding platform. Devin gives the asynchronous agent layer, while Windsurf adds the synchronous IDE layer.
The strongest competitive risk is bundling. Microsoft/GitHub, OpenAI, Anthropic, and Google do not need to beat Devin perfectly; they only need to make autonomous coding feel like a default layer inside tools companies already use.
The open-source risk is also real. OpenHands suggests that the agent harness could become less proprietary over time, which would pressure Devin’s pricing power and gross margins.
The TAM depends entirely on framing. If Devin is a developer tool, $26B is stretched; if it becomes the enterprise layer for converting engineering payroll into metered AI-agent work, the market can be large enough.
Our conclusion is that Cognition’s $26B valuation is aggressive but not absurd in the current AI market. It is priced for near-perfect execution, fast revenue growth, strong enterprise retention, defensible workflow ownership, and successful resistance against bundled agents from larger platforms.
This market map, featured in our AI code assistant market deck, highlights top companies and startups in the AI code assistant market
How did Cognition’s valuation evolve?
Cognition’s valuation has been growing by around 2.0x to 2.55x each time. The company is adding roughly $1.8 billion per month in valuation.
We found five useful valuation checkpoints. The company was valued around $350M in early 2024, then raised $175M at a $2B valuation in April 2024, then reached nearly $4B in March 2025, $10.2B in September 2025, and finally $26B in May 2026.
We can notice a recent acceleration after a slow middle period.
From April 2024 to March 2025, Cognition added only about $186M of valuation per month. From March to September 2025, that jumped to about $1.1B per month. From September 2025 to May 2026, it jumped again to about $1.8B per month.
| Date | Valuation | Event | Multiple vs previous | Value added per month |
|---|---|---|---|---|
| Early 2024 | $350M | Founders Fund-led early round | — | — |
| Apr. 2024 | $2B | $175M round after Devin went viral | 5.7x | ~$1.65B/month |
| Mar. 2025 | ~$4B | 8VC-led round | 2.0x | ~$0.19B/month |
| Sep. 2025 | $10.2B | $400M+ round after Windsurf acquisition | 2.55x | ~$1.08B/month |
| May 2026 | $26B | $1B+ Series D, $492M run-rate revenue | 2.55x | ~$1.83B/month |
What revolutionary things does Cognition actually offer?
Cognition’s real novelty is not “AI that writes code”. A lot of companies do that.
However, Devin offers something more ambitious: turning coding from a chat-assistant workflow into a delegated-labor workflow.
The big shift is that Devin is positioned as an AI software engineer that can plan, write, test, and ship production code inside a company’s own codebase and tooling.
Cognition’s own language is much closer to “hire an engineer” than “use an autocomplete tool”.
Cognition’s strongest innovation is not a single feature but the product category framing.
For example, Cursor says “better IDE.” Claude Code says “agentic coding tool.” Codex says “coding partner.” Cognition says, implicitly, “here is a new unit of engineering capacity”. It’s definitely more ambitious.
| Thing | Why it’s revolutionary | Who also proposes it, and to what extent |
|---|---|---|
| Autonomous software engineering agent | Devin is built to take a goal and execute multi-step engineering work end-to-end, not just suggest code. That changes the buyer logic from “developer productivity software” to “software labor leverage.” | Claude Code also operates across a project and can complete development tasks autonomously. OpenAI Codex also claims end-to-end engineering tasks. The gap is that Cognition packaged this earlier and more explicitly as a “software engineer.” |
| Parallel cloud agents | The productivity leap comes from running many agents at once, not making one developer type faster. This is closer to managing a small engineering team than using an IDE plugin. | OpenAI Codex explicitly markets multi-agent cloud workflows with parallel worktrees. GitHub Copilot coding agent can work in the background on issues. Devin’s differentiation is the stronger “remote worker” positioning. |
| Deep codebase context and repo understanding | Devin can work inside large existing repositories, where most enterprise engineering time is spent. This matters because legacy modernization, migrations, and bug fixing are more valuable than greenfield code generation. | Claude Code and Cursor also emphasize whole-codebase understanding. Cognition’s edge is less exclusive here; this is becoming table stakes. |
| Agent + IDE bundle after Windsurf | By acquiring Windsurf, Cognition got both the asynchronous agent layer and the synchronous IDE layer. That means it can serve both “delegate this task” and “help me code right now.” | Cursor is strongest on the IDE-native side. GitHub Copilot is strongest inside GitHub/Microsoft workflows. Cognition’s move is to combine both modes under one company. |
| Enterprise workflow integration | Devin is sold into serious enterprise workflows: modernization, migrations, security remediation, documentation, and PRs. That makes adoption easier for companies that need governance, not just developer enthusiasm. | GitHub Copilot Enterprise, Claude Code, and Codex all compete here. Cognition’s advantage is that its story is unusually focused on replacing chunks of engineering work, not just assisting engineers. |
| AI productivity guarantee / measurable agent output | Cognition is trying to make agent work measurable in human-equivalent engineering hours. That is important because enterprises will not pay huge contracts forever without ROI proof. | Others measure usage and productivity, but Cognition is unusually explicit about selling agents as accountable labor. This is still early and needs independent validation. |
If you want more recent data on this point, please see our latest AI code assistant market report.
As this chart shows, and as featured in our AI code assistant market deck, search interest in AI code assistants has increased significantly
Does Cognition already have big confirmed customers?
Yes, Cognition has big confirmed customers, but most confirmations come from Cognition itself rather than from every customer independently.
Cognition says Devin and Windsurf power or partner with Goldman Sachs, Citi, Dell, Cisco, Ramp, Palantir, Nubank, Mercado Libre, Mercedes-Benz, Santander, Elevance, the U.S. Army, and the U.S. Navy.
One caveat: the public evidence confirms logos, not contract sizes. We know the customer quality is real. We do not know how much of the $492M run-rate revenue is concentrated in a few very large contracts.
What is Cognition’s latest revenue or ARR?
The latest official number is $492M of run-rate revenue, disclosed by Cognition in May 2026.
Cognition said in its Series D announcement that enterprise usage had grown more than 10x since the start of 2026 and that run-rate revenue had grown to $492M.
Before that, Cognition disclosed that Devin alone grew from $1M ARR in September 2024 to $73M ARR in June 2025. It also said Windsurf brought $82M of ARR when acquired, plus 350+ enterprise customers and hundreds of thousands of daily active users.
The clean revenue bridge is:
| Date | Metric | Source quality |
|---|---|---|
| Sep. 2024 | Devin at $1M ARR | Official Cognition disclosure |
| Jun. 2025 | Devin at $73M ARR | Official Cognition disclosure |
| Jul. 2025 | Windsurf at $82M ARR | Official Cognition disclosure |
| Sep. 2025 | Combined enterprise ARR up >30% in seven weeks post-Windsurf | Official Cognition disclosure |
| May 2026 | Cognition at $492M run-rate revenue | Official Cognition disclosure |
This chart, featured in our AI code assistant market deck, illustrates yearly VC funding for AI code assistant startups
What is Cognition’s valuation multiple on revenue or ARR? Is it high?
At $26B valuation and $492M run-rate revenue, Cognition trades at about 53x revenue.
It is far above public SaaS, even above most premium cloud software names.
Public SaaS multiples vary by index and methodology, but the broad market is nowhere near Cognition. PublicSaaSCompanies reported a median public SaaS revenue multiple of 2.74x as of June 5, 2026. PitchBook’s Q1 2026 enterprise SaaS comp sheet put the median EV / trailing revenue multiple at 3.3x as of March 31, 2026.
Premium public SaaS is higher, but still mostly below Cognition. SEG’s Q1 2026 SaaS report listed Cloudflare at 30.5x, CrowdStrike at 21.7x, Shopify at 13.8x, Snowflake at 12.9x, and Datadog at 11.9x EV / TTM revenue.
Clearly, Cognition is not just priced like a good SaaS company but like a company that could change the labor model of software engineering. That is a much bigger claim than “high-growth SaaS.”
| Company / benchmark | Revenue multiple | Cognition premium |
|---|---|---|
| Cognition | 52.8x | — |
| Cloudflare | 30.5x | 1.7x higher |
| CrowdStrike | 21.7x | 2.4x higher |
| Shopify | 13.8x | 3.8x higher |
| Snowflake | 12.9x | 4.1x higher |
| Datadog | 11.9x | 4.4x higher |
| Public SaaS median | 2.74x | 19.3x higher |
How does Cognition’s multiple compare with Cursor?
Cognition is valued at a higher revenue multiple than Cursor, unless we use very aggressive rumored Cursor numbers.
Cursor’s official November 2025 Series D announcement said it raised $2.3B and had passed $1B in annualized revenue. BusinessWire reported the round at a $29.3B valuation, with enterprise revenue up 100x in 2025 YTD.
Using those official numbers: Cursor: $29.3B valuation ÷ $1B+ annualized revenue = roughly 29x or lower.
Cognition is at 52.8x, so Cognition is roughly 1.8x more expensive than Cursor on the cleanest official comparison.
There is a complication: TechCrunch later reported that Cursor had surpassed $2B in annualized revenue by March 2026, based on a Bloomberg source. If we keep Cursor’s last confirmed $29.3B valuation and update revenue to $2B, Cursor’s multiple falls to roughly 14.7x. In that version, Cognition is about 3.6x more expensive.
There were also reports that Cursor was discussing a new round around $50B while at roughly $2B ARR. That would imply about 25x revenue, still roughly 2.1x below Cognition’s 53x.
This chart, featured in our AI code assistant market deck, breaks down Anyshpere’s playbook in AI code assistants
How fast did Cognition’s revenue grow?
Cognition’s revenue growth is extreme: Devin went from $1M ARR to $73M ARR in nine months, then Cognition reached $492M run-rate revenue by May 2026.
The official milestones are unusually clear. Devin grew from $1M ARR in September 2024 to $73M ARR in June 2025, a 73x increase in nine months. That is about $8M of new ARR per month on average.
Then Cognition acquired Windsurf, which had $82M ARR, meaning the combined base immediately moved from Devin’s $73M to at least roughly $155M ARR before any post-deal growth. Cognition then said combined enterprise ARR rose more than 30% in the seven weeks after buying Windsurf.
By May 2026, Cognition reported $492M run-rate revenue. If we compare that with the reported $37M annualized revenue in May 2025 cited by industry coverage, that is about 13.3x year over year, or roughly $38M of run-rate revenue added per month.
| Period | Revenue movement | Multiple | Approx. monthly add |
|---|---|---|---|
| Sep. 2024 → Jun. 2025 | $1M ARR → $73M ARR | 73x | +$8M/month |
| Jul. 2025 acquisition | +$82M Windsurf ARR | +$82M inorganic | — |
| 7 weeks post-Windsurf | Combined enterprise ARR up >30% | >1.3x | Not fully disclosed |
| May 2025 → May 2026 | $37M → $492M run-rate revenue | 13.3x | +$37.9M/month |
The smart observation here is that Cognition’s growth changed shape.
The first phase was Devin product-market fit. The second phase was Windsurf inorganic scale. The third phase is the real test: whether Cognition can turn that combined product suite into durable enterprise expansion.
So far, the revenue curve supports a premium valuation. But it also creates the risk: at 53x revenue, even “very fast growth” may not be enough.
Are there companies trading at a 50x revenue multiple like Cognition?
Yes, but the list is very narrow: Cognition’s 53x revenue multiple looks less crazy only when we compare it to the most aggressive AI re-rating stories, not to normal SaaS.
The cleanest public-market example is Palantir. Depending on whether we use market cap / TTM revenue or enterprise value / revenue, Palantir is roughly in the 55x to 65x revenue zone. That is the closest public comp because investors are not valuing Palantir as generic software anymore. They are valuing it as an AI operating system for governments and enterprises.
Then we have private AI comps, but the data is less clean. OpenAI has reportedly traded around the 50x to 65x revenue zone depending on which valuation and annualized revenue number we use. xAI has also reportedly traded in a very high revenue multiple range, sometimes above 50x, but its revenue base and valuation references are less transparent.
Anthropic, despite being more strategically important than Cognition, often looks lower on revenue multiple if we use the latest very large reported revenue run-rate numbers.
| Company | Approx. valuation / market cap | Approx. revenue base | Implied revenue multiple | Why it matters |
|---|---|---|---|---|
| Cognition | $26B | $492M run-rate revenue | 52.8x | AI software engineering agent |
| Palantir | ~$348B market cap | ~$5.2B TTM revenue | ~60x+ market cap / revenue | Public AI software comp closest to Cognition’s multiple |
| OpenAI | ~$800B+ reported valuation | ~$13B reported annualized revenue | ~60x+ | Frontier AI platform, not SaaS |
| xAI | ~$75B+ reported valuation | ~$1B reported revenue / run-rate | ~75x | Frontier AI / distribution / model bet |
| Cloudflare | ~$88B market cap | ~$2.3B TTM revenue | ~38x | Premium AI/cloud infrastructure, but below Cognition |
| CrowdStrike | ~$173B market cap | ~$5.1B TTM revenue | ~34x | Premium cybersecurity SaaS, still far below Cognition |
| AppLovin | ~$189B market cap | ~$6.2B TTM revenue | ~31x | AI ad platform with very high margins, still below Cognition |
| Cursor / Anysphere | $29.3B last confirmed valuation | $1B+ annualized revenue | ~29x or lower | Closest AI coding peer, materially cheaper than Cognition |
The pattern is clear: 50x revenue is not impossible today, but it is not a software multiple. It is an “AI category-winner” multiple.
Companies at this level usually have at least one of three things: a monopoly-like AI narrative, explosive revenue acceleration, or a belief that the company is capturing a much larger budget than its current category suggests.
If you want more recent data on this point, please see our latest AI code assistant market report.
This chart, featured in our AI code assistant market deck, illustrates yearly funding for AI code assistant startups
Does Cognition’s 53x multiple look crazy compared with those companies?
Cognition’s multiple is expensive, but not obviously insane if we believe Devin can become software labor rather than developer tooling.
Against public SaaS, Cognition looks extreme. Cloudflare, CrowdStrike, Snowflake, Datadog, and other premium SaaS companies usually trade far below 53x revenue.
Even excellent SaaS businesses with strong retention, high gross margins, and durable enterprise contracts rarely sustain this kind of multiple unless they are also being treated as an AI platform.
Against Palantir, the comparison is more interesting. Palantir is much larger, public, profitable, strategically embedded, and growing fast for its scale. Cognition is smaller, private, less proven, and probably less profitable.
But Cognition is growing much faster from a smaller base. Palantir is the “de-risked AI software operating system.” Cognition is the “high-risk AI labor substitution option.”
However, Palantir gets a similar multiple with far more proof. Palantir has multi-billion-dollar revenue, public financials, profitability, government depth, and a long enterprise sales track record. Cognition has explosive revenue, strong logos, and a compelling category, but much less transparency.
The supportive argument for Cognition is that it is growing faster than almost every public comp.
Cognition went from roughly $37M annualized revenue in May 2025 to $492M run-rate revenue in May 2026, so roughly 13x year over year. Palantir, Cloudflare, CrowdStrike, and AppLovin are growing fast, but they are not growing 13x.
The skeptical argument is that Cognition’s growth is partly inorganic because of Windsurf. It did not simply grow Devin from zero to $492M alone. It combined Devin with Windsurf, absorbed Windsurf’s $82M ARR base, then accelerated the combined business.
That is still impressive, but it is not the same quality as purely organic ARR growth.
So, we believe 53x is not crazy in the current AI market, but it is priced for near-perfect execution.
At what ARR does Cognition’s valuation start to look reasonable?
At $26B valuation, Cognition needs roughly $1.7B to $2.6B of ARR for the multiple to feel reasonable by elite software standards.
Right now the math is simple: $26B valuation / $492M run-rate revenue = 52.8x revenue. That is the uncomfortable starting point.
If we define “reasonable” as premium AI software, then a 20x to 30x multiple can be defended. If we define “reasonable” as elite public SaaS, then we want something closer to 10x to 15x.
If we define “normal SaaS,” then it would be lower, but that is probably the wrong benchmark for Cognition.
| Cognition ARR / run-rate revenue | Implied multiple at $26B valuation | How it would feel |
|---|---|---|
| $492M | 52.8x | Current level; priced for extreme future growth |
| $750M | 34.7x | Still very expensive; comparable to top-tier AI/cloud public names |
| $867M | 30.0x | Aggressive but defensible if growth remains exceptional |
| $1.0B | 26.0x | Still rich, but no longer absurd for a category leader |
| $1.3B | 20.0x | Starts to look reasonable for premium AI software |
| $1.7B | 15.3x | Reasonable for elite software if growth and margins are real |
| $2.6B | 10.0x | Very reasonable if Cognition is still growing fast |
| $5.2B | 5.0x | Normal public SaaS multiple territory |
The right threshold is probably $1.5B to $2B ARR.
At that level, the valuation would compress into the 13x to 17x range, which is still premium but no longer requires heroic assumptions.
The more important question is time. If Cognition reaches $1.5B ARR in 12 to 18 months, the current valuation looks aggressive but understandable.
If it takes three or four years, the valuation looks much harder to justify because the market will have demanded proof on margins, retention, and competition by then.
This chart, featured in our AI code assistant market deck, compares the main business model options for AI developer tools platforms
Is Cognition profitable?
Cognition has not disclosed profitability, and we should assume it is not meaningfully profitable today.
There is no public evidence that Cognition is profitable. The company talks about revenue, usage, customers, product progress, and fundraising. It does not disclose operating margin, free cash flow, gross margin, burn, net revenue retention, or payback period. That silence is heavy.
Most likely, Cognition is not GAAP profitable. The company is scaling enterprise sales, absorbing Windsurf, hiring aggressively, serving compute-heavy agent workloads, and competing in a market where pricing is moving fast.
Those are not the conditions of a mature profitable SaaS company.
What gross margin can we estimate for Devin?
Devin is probably not 85% to 90% gross-margin SaaS today. It is more likely a 30% to 60% gross-margin product now, with a path to 60% to 75% if inference costs fall and enterprise pricing holds.
Classic SaaS has beautiful gross margins because serving one more user is cheap. Devin is different. Every task can consume model inference, VM runtime, browser actions, networking, test runs, retries, and long-context reasoning. The product is not just software access. It is software plus active compute.
The pricing model tells us the truth. Devin uses Agent Compute Units, which measure the computing resources required to complete tasks.
That means Cognition itself knows per-seat pricing is dangerous. If it sold unlimited autonomous engineering work for a flat subscription, heavy users could destroy margins.
A rough way to think about it:
| Product type | Typical gross margin profile | Why |
|---|---|---|
| Traditional SaaS | 75% to 90% | Low marginal cost per user |
| Cloud infrastructure | 50% to 70% | Compute and data center costs matter |
| AI API / model inference resale | 0% to 50% | Heavy model costs, especially if using third-party models |
| AI coding assistant with usage caps | 40% to 75% | Better if usage is controlled and enterprise priced |
| Devin-style autonomous agent | 30% to 60% today; maybe 60% to 75% later | Agent loops are expensive, but usage-based pricing can protect margins |
Our estimate for Devin today: gross margin is probably closer to intelligent compute than pure SaaS. We would not model it like Datadog or ServiceNow yet.
This chart, featured in our AI code assistant market deck, illustrates how market revenue is distributed across customer segments in the AI code assistant market
What is Cognition’s TAM?
Cognition’s TAM is small if we call it developer tooling, but massive if we call it engineering labor substitution.
If Cognition is just an AI coding tool, the market is probably $10B to $30B. That includes AI IDEs, code assistants, code review, test generation, and agentic coding tools.
In that framing, Cognition is already large: at $492M run-rate revenue, it may already represent several percentage points of the current AI coding-tools market. That makes the $26B valuation hard to defend unless Cognition becomes one of the two or three dominant vendors.
If Cognition is a broader developer productivity platform, the TAM expands to roughly $50B to $100B+. That includes workflows around planning, coding, testing, documentation, security remediation, migrations, and engineering knowledge management.
This is the “premium software platform” case. It can support a large company, but it still does not fully explain a 53x revenue multiple unless growth stays extreme.
The real bull case is labor substitution. Global software engineering labor spend is plausibly in the $1T to $2T+ range when we include millions of developers, fully loaded compensation, contractors, outsourced engineering, QA, maintenance, and modernization.
Cognition does not need to replace all of that. If AI agents capture just 5% of a $1.5T engineering labor pool, that is a $75B annual revenue opportunity.
So the clean range is this: $30B TAM if Devin is a developer tool, $100B to $300B if it becomes an enterprise engineering automation layer, and $1T+ only if it truly substitutes labor.
So, the valuation only works under the labor-substitution framing. If Devin is “better Cursor,” $26B is stretched. If Devin becomes the system companies use to convert engineering payroll into metered AI-agent work, the TAM is easily large enough.
If you want more recent data on this point, please see our latest AI code assistant market report.
Can OpenAI, Anthropic, Google, and Microsoft crush Cognition’s value?
Yes. We would put the probability at roughly 45% that the hyperscalers compress Cognition’s valuation by more than 70% within three years, but only 20% that they make the business truly worthless.
The key point is that they do not need to “beat Devin” feature by feature.
They only need to make autonomous coding agents feel like a bundled layer inside ChatGPT, Claude, GitHub, VS Code, Google Cloud, and enterprise AI platforms.
The probability has gone up sharply in the last year. OpenAI Codex is now a cloud coding agent that works in parallel, opens pull requests, runs in its own cloud environment, and is distributed through ChatGPT. OpenAI says Codex usage grew more than 10x after launch, served 40T+ tokens in three weeks after GPT-5-Codex, and is used by companies including Duolingo, Vanta, Cisco, and Rakuten. That makes OpenAI a direct Devin competitor, not a model supplier sitting below the market.
Microsoft/GitHub is the most structurally dangerous. GitHub Copilot coding agent is generally available to paid Copilot users, works asynchronously, uses GitHub Actions, opens draft PRs, and sits exactly where software teams already manage issues and code. GitHub also made the Copilot SDK generally available in June 2026, giving developers access to the same agentic runtime behind Copilot. That turns GitHub from “coding assistant” into agent infrastructure.
Anthropic is also closer than people think. Claude Code reads codebases, edits files, runs commands, and delivers committed code. Anthropic is also building Managed Agents with sessions, harnesses, sandboxes, permissions, credential vaults, and execution tracing. That is not a toy CLI. That is the same production scaffolding Cognition wants to own.
Google is less dangerous today, but not irrelevant. Jules is an asynchronous coding agent, out of beta, powered by Gemini, and Google said developers had already used it for 140,000+ public code improvements during beta. Google also has the cloud, model, Android, Workspace, and enterprise distribution layers to keep improving it.
This chart, featured in our AI code assistant market deck, shows how AI coding assistant technology has evolved over time
Can OpenHands commoditize Devin?
Yes, OpenHands is a real threat to Devin’s pricing power, but not yet to Cognition’s enterprise brand.
OpenHands attacks the layer investors want to believe is defensible: the agent harness. OpenHands can plan, write, execute, and apply code changes across a codebase. Its repo also points to an enterprise version with source-available code and paid licensing for longer use.
The problem for Cognition is that the core loop is becoming obvious: model call, tool call, shell, file edit, test, retry, memory, sandbox, permissions. Claude Code research points the same way. The product is not magic. It is a loop plus strong engineering around permissions, context, storage, subagents, and isolation.
OpenHands wins if three things happen: models improve, agent scaffolding standardizes, and enterprises prefer self-hosted or customized agents over premium Devin pricing.
Our probability: 30% that OpenHands materially compresses Devin’s margins within three years. That is not the base case, but it is too high to ignore when Cognition’s valuation assumes proprietary workflow software.
The key bear signal is benchmark direction. On long-horizon software evolution tasks, even GPT-5 with OpenHands reached only 21% resolution, far below simpler single-issue benchmarks. Agents are still weak, but the race is shifting toward scaffolding, evals, and runtime design.
OpenHands is not the killer but the price anchor. It tells every enterprise buyer that Devin-like architecture is not infinitely proprietary.
If you want more recent data on this point, please see our latest AI code assistant market report.
Are Lovable and Base44 serious threats to Devin?
Lovable and Base44 are not direct Devin killers, but they are a serious threat to the low-end version of Cognition’s TAM story.
Lovable and Base44 serve a different buyer. Devin is for engineering teams working inside real codebases. Lovable and Base44 are for people who want to create apps from prompts, often without being engineers. That is not the same job.
But the threat is real in one specific way: they steal the “AI can create software” narrative from below. Lovable reportedly reached $200M ARR in 12 months and later industry tracking claimed $400M ARR in 14 months. Base44 reached $100M ARR inside Wix, with Wix saying users were building everything from simple apps to customized business software.
That matters because some software demand does not need Devin. A sales ops team that wants an internal dashboard, a founder who wants an MVP, or a department that wants a lightweight workflow app may choose Lovable, Base44, Replit, Bolt, or v0 instead of asking engineering to delegate work to Devin.
Still, this is mostly a TAM segmentation issue, not a direct enterprise replacement. Lovable and Base44 are dangerous for new app creation and non-technical builders. Devin is more relevant for existing repositories, migrations, bugs, tests, PRs, and enterprise engineering workflows.
Lovable threatens the story that Cognition owns software creation. It does not yet threaten the story that Cognition owns enterprise engineering labor.
In our AI code assistant market deck, we identify pain points entrepreneurs should prioritize
Who is the most dangerous competitor to Cognition today?
Microsoft/GitHub is the most dangerous competitor to Cognition today. OpenAI is second. Anthropic is third.
Microsoft/GitHub is the most dangerous because it owns the workflow. GitHub has the repositories, issues, pull requests, Actions, enterprise procurement, security controls, and Copilot user base. If coding agents become a GitHub-native feature, Cognition has to convince companies to add a separate vendor for work that GitHub already touches every day.
OpenAI is second because Codex is now attacking Devin head-on: parallel cloud agents, PR generation, ChatGPT distribution, frontier coding models, and enterprise customers. OpenAI can also subsidize usage, bundle into ChatGPT Business/Enterprise, and improve the model faster than most application companies can adapt.
Anthropic is third because Claude Code is probably the strongest developer-loved agent experience outside OpenAI/Copilot, and Anthropic is building the managed-agent infrastructure that can generalize beyond coding. Claude Code also has the right product shape: terminal, IDE, codebase understanding, file edits, tests, commands, and committed code.
Google is fourth. It has the resources, but Jules has not yet shown the same developer pull or enterprise urgency. OpenHands is fifth: it is not the biggest revenue competitor, but it is the best commoditization signal.
What is the bear case for Cognition?
The bear case is that Devin becomes a feature, not a company.
In the bear case, OpenAI, GitHub, Anthropic, and Google all reach “good enough” autonomous coding within 12 to 24 months. Enterprises already using GitHub Copilot, ChatGPT Enterprise, Claude, Azure, or Google Cloud choose bundled agents because procurement is easier, security review is done, and pricing is lower.
At the same time, OpenHands and other open-source scaffolds make the agent harness feel less proprietary. Customers stop believing they need Cognition’s full-stack product. They use Codex or Copilot for standard work, Claude Code for hard repo tasks, OpenHands for self-hosted workflows, and Lovable/Base44 for non-engineer app creation.
The damage is not that revenue goes to zero immediately. The damage is that the market stops valuing Cognition at 53x revenue. If Cognition’s gross margin looks compute-heavy, if retention is not exceptional, or if growth slows below the current extreme curve, the valuation can compress fast.
Our probability for the bear case: 40% over three years.
If you want more recent data on this point, please see our latest AI code assistant market report.
This chart, featured in our AI code assistant market deck, illustrates how revenue is distributed geographically across Europe, Asia, North America, Africa, and South America in the AI code assistant market
What is the bull case for Cognition?
The bull case is that Cognition becomes the first real AI engineering workforce company.
In the bull case, Devin is not a coding assistant and not a vibe-coding tool. It becomes the enterprise system for delegating engineering work: migrations, test generation, bug fixes, dependency upgrades, security remediation, codebase documentation, and eventually feature work.
The strongest bull signal is revenue velocity. Cognition reported $492M run-rate revenue by May 2026 after Devin grew from $1M ARR in September 2024 to $73M ARR in June 2025, then added Windsurf’s $82M ARR and large enterprise distribution. Few software companies ever show that steep a commercial curve.
The bull case also depends on workflow defensibility. Cognition needs to own the orchestration layer around AI engineering work: repo context, permissions, evals, task routing, audit logs, PR review, memory, enterprise integrations, and measurable output. If it owns that layer, the model underneath becomes a cost input, not the product.
Our probability for the bull case: 30% over three years.
OUR METHODOLOGY
This analysis tests whether Cognition’s reported $26 billion valuation is economically plausible based on the evidence available today. We compare the headline valuation with Cognition’s valuation history, latest run-rate revenue, revenue multiple, peer benchmarks, competitive threats, TAM framing, and bull/bear scenario logic.
We separated Devin’s standalone growth from Cognition’s post-Windsurf scale because the acquisition changed the shape of the business. That distinction helps separate product-market fit from platform consolidation.
For valuation multiples, we used Cognition’s latest official run-rate revenue because it is the clearest current figure disclosed by the company. Where figures came from reports rather than company announcements, we treated them as secondary signals.
The probabilities in the competition and bull/bear sections are scenario judgments, not sourced forecasts. We weighted them around distribution control, bundling risk, open-source commoditization, revenue velocity, and how defensible the workflow layer appears.
Key sources used for this analysis include: Cognition’s Series D announcement and $492M run-rate revenue disclosure, Cognition’s Windsurf acquisition announcement and $82M ARR disclosure, Cognition’s funding and Devin ARR growth update, OpenAI Codex cloud agent documentation, OpenAI’s Codex product page, GitHub’s Copilot coding agent general-availability announcement, GitHub’s Copilot coding agent press release, Anthropic’s Claude Code product page, Claude Code documentation, Google’s Jules announcement, OpenHands’ official site, and the OpenHands GitHub repository.
This chart, featured in our AI code assistant market deck, illustrates yearly VC funding for AI code assistant startups
Related blog posts
- How strong is fundraising in the AI code assistant market right now?
- The startups that have raised the most funding in the AI code assistant market
- The most highly valued startups in the AI code assistant market
