history
the company, sourced and dated — the ioi apartment, the devin launch and debunk, the windsurf weekend, and the $48b arc.
the ioi apartment2023–2024-02
three ioi gold medalists work out of a new york apartment through roughly eight pivots — reportedly starting in crypto — before landing on the autonomous software engineer. the company formally incorporates around january 2024, hires russell kaplan as president, and takes $21m from founders fund.
- 2023cognition foundedscott wu, steven hao, and walden yan — all ioi gold medalists — found cognition in a new york apartment. the founding date is genuinely disputed: august 2023 (wikipedia body, david senra's notes) versus november 2023 (wikipedia's own infobox, wu's linkedin); an outlier fan bio claims january 2023. per wu the team ran ~eight pivots first, the earliest reportedly a crypto project.
- 2024-01incorporated; kaplan joinsthe company formally incorporates — bloomberg's launch-day feature notes it 'didn't even officially exist as a corporation until two months ago.' the same month russell kaplan, scale ai's director of engineering and ml (helia founder, tesla autopilot before that), joins as president.
- 2024-03-12$21m at ~$350mthe $21m round is disclosed alongside the launch — led by founders fund with peter thiel personally involved; angels include elad gil, patrick and john collison, chris re, and sarah guo — at a ~$350m valuation. the round is labeled both 'seed' and 'series a' across sources; the label question never resolves.
the first ai software engineer2024
'introducing devin, the first ai software engineer' goes viral on a 13.86% swe-bench claim — then the internet of bugs debunk, a $2b round inside the debunk fortnight, and general availability at $500 a month in december.
- 2024-03-12introducing devinscott wu publishes 'introducing devin, the first ai software engineer': a tireless teammate with shell, editor, and browser in a sandbox, claiming 13.86% on swe-bench — footnoted as a random 25% subset, devin unassisted while other models were assisted — versus a prior 1.96%. upwork job demos narrated by walden yan; the cta is literally 'hire devin.' the launch tweet goes viral; ashlee vance's 'gold-medalist coders' feature runs the same day.
- 2024-03-13opendevin appears overnightthe open-source replication begins one day after launch: the opendevin repository is created — the project that becomes openhands (~85k stars). the community's answer to a demo it couldn't inspect.
- 2024-03-15the swe-bench reportthe swe-bench technical report publishes the methodology — 570 of 2,294 issues, unassisted — and cognition releases its eval harness plus devin's actual edits on github. unusually transparent for a launch claim; the 25%-subset caveat still does the reputational damage.
- 2024-04the debunkinternet of bugs (carl brown) publishes the frame-by-frame debunk of the upwork demo: the client had asked for setup instructions, the readme already contained them, devin debugged files that didn't exist in the repo, and hours of work compressed to ~36 minutes for a human. the hn thread hits the front page; ai made simple adds that the listing was found by searching 'road damage' — a curated demo. no formal rebuttal from cognition is documented.
- 2024-04-24$175m at ~$2binside the debunk fortnight: a $175m round closes at ~$2b — the information first, founders fund the sole investor per ai business — after the company reportedly turned down offers at ~$1b. the signature dissonance of the era: review scores and valuation fully decoupled.
- 2024-05the waitlist opens, microsoft signs onearly access expands from ~250 accounts — a cognition tweet around april 17 says 'we just asked devin to help us send out the first batch of invitations.' public write-ups by early users appear (ppaolo's in-depth analysis; student andrew kean gao among the first testers). at microsoft build, a partnership is announced: devin on azure, aimed at code migration and modernization.
- 2024-06-05the update cadencethe monthly-update cadence begins — first product update june 5, 'using devin to recover from the crowdstrike outage' july 20, and the o1 evaluation post september 12 ('how we evaluate coding agents'). the engineering blog becomes the company's public interface while access stays gated.
- 2024-12-10general availability at $500devin reaches general availability eight months after the unveil: $500/month flat, no seat limits, slack as the primary interface plus an ide extension and api. the ga post claims project timelines 'accelerated by as much as 16x' (self-reported); early customers named: nubank, vesta, ramp. two days later the devin open source initiative gives free acus to oss maintainers, with devin contributions to anthropic's mcp, llamaindex, and zod.
the reality gap2025-01 → 2025-06
the reality-gap months: answer.ai's 3-of-20 evaluation, the $500→$20 price collapse of devin 2.0, deepwiki, walden yan's 'don't build multi-agents' — and a ~$4b round landing anyway.
- 2025-01-08the answer.ai evaluationanswer.ai publishes the standing independent evaluation: of 20 attempted tasks, 3 successes, 3 inconclusive, 14 failures — 'devin would spend days pursuing impossible solutions rather than recognizing fundamental blockers,' and 'we couldn't discern any pattern to predict which tasks would work.' the register's headline: ''first ai software engineer' is bad at its job.'
- 2025-01-16devin 1.2devin 1.2 — repository-context reasoning, voice messages, enterprise account management, and usage-based billing: the first crack in the $500 flat price.
- 2025-03-17~$4b anywaya growth round of 'hundreds of millions' lands at ~$4b, led by 8vc with lux capital co-leading per cognition; founders fund, khosla, elad gil, and conviction participate. it arrives two months after the answer.ai evaluation — the valuation-decoupled-from-reviews pattern made literal.
- 2025-04-03devin 2.0, $20devin 2.0 — the agent-native ide experience, interactive planning, parallel devins, and devin search/wiki precursors — with the price collapse: a $20 core plan then pay-as-you-go acus at $2.25. techcrunch does the math (~2.25 hours of 'active devin work'); plain ai reads it as admission that 'the market for $500/month autonomous agents was much smaller than the press tour suggested.'
- 2025-04-16openai circles windsurfbloomberg and axios report openai in talks to acquire windsurf for ~$3b — the deal whose collapse will define cognition's summer. windsurf (fka codeium, founded 2021 by mit classmates varun mohan and douglas chen) had raised ~$240m and was last valued at $1.25b.
- 2025-05-05deepwiki; devin 2.1; windsurf's swe-1deepwiki launches — free ai-generated documentation for any public github repo, the code-understanding surface that becomes devin wiki's sibling. ten days later devin 2.1 ships confidence indicators (green/yellow/red); the same day a still-independent windsurf launches its swe-1 model family.
- 2025-06-12don't build multi-agentswalden yan publishes 'don't build multi-agents' — the essay that popularizes 'context engineering' over prompt engineering: 'share context'; 'actions carry implicit decisions'; the field 'still playing with raw html & css.' cognition's most durable intellectual export.
the windsurf twist2025-06 → 2025-08
the 72-hour deal: openai's ~$3b exclusivity lapses, google takes windsurf's founders for $2.4b, and cognition signs for everything left behind — then lays off 30 and offers the rest nine-month buyouts under an 80-hour culture memo.
- 2025-06-03anthropic cuts windsurf's claudeanthropic cuts windsurf's first-party claude 3.x capacity with under five days' notice, amid the openai acquisition rumors — jared kaplan: 'i think it would be odd for us to be selling claude to openai.' windsurf's arr had reportedly touched ~$100m in april and slipped toward ~$82m in the squeeze.
- 2025-07-11google takes the brainhours after openai's exclusivity lapses — reported casualty: microsoft ip-access friction — google announces a $2.4b non-exclusive license plus the hires: ceo varun mohan, co-founder douglas chen, and senior r&d into deepmind to work on agentic coding for gemini. no equity stake. jeff wang (head of business) becomes interim ceo of the ~250-person shell; graham moreno becomes president.
- 2025-07-11goldman 'hires' devinthe same day, cnbc breaks the goldman deployment — cio marco argenti: 'we're going to start augmenting our workforce with devin, which is going to be like our new employee… initially, we will have hundreds of devins [and] that might go into the thousands.' first major bank on devin, per cognition; supervised, on drudgery work like legacy updates; argenti claims 3–4x productivity versus prior tools.
- 2025-07-14cognition takes the bodycognition signs the definitive agreement for everything google left: the ip, product, trademark and brand, the ~250-person team, $82m arr, 350+ enterprise customers, and 'hundreds of thousands' of daus — with full claude access restored. first call friday ~5pm, signed monday morning. wu's memo: '100% of windsurf employees will participate financially… vesting cliffs waived… fully accelerated vesting… there's only one boat and we're all in it together.' price undisclosed; secondaries estimate ~$250m.
- 2025-07-22the mcp marketplacedevin's mcp marketplace launches — the agent plugs into the mcp server ecosystem eight days post-acquisition.
- 2025-08-05the 80-hour memothree weeks after the 'one boat' memo: 30 windsurf-team layoffs and nine-month buyouts offered to the ~200 remaining (decide by august 10); stay means six office days a week and 80+ hour weeks. wu's memo, verbatim: 'we don't believe in work-life balance—building the future of software engineering is a mission we all care so deeply about that we couldn't possibly separate the two.' wu publicly owns it: 'many of us literally live where we work.'
the independent agent lab2025-09 → 2026-09
the independent agent lab: $10.2b → $26b → $48b in twelve months, in-house swe models served by cerebras, devin desktop absorbing windsurf, government contracts, and a self-scored productivity guarantee.
- 2025-09-08series c, $10.2bthe >$400m series c lands at $10.2b post, led by founders fund with lux, 8vc, neo, elad gil, definition, swish, and new names bain capital ventures, hanabi, d1. wu's metrics post is the real document: devin arr $1m (sep 2024) → $73m (jun 2025), lifetime net burn under $20m, >30% combined enterprise-arr growth in the seven weeks post-windsurf, <5% customer overlap — all self-reported. christian lawless (conversion capital), emily cohen (neo), and swyx join full-time.
- 2025-09-29rebuilt on sonnet 4.5devin is rebuilt on claude sonnet 4.5 and devin agent preview ships — '2x faster, 12% better on our jr. developer evals' (self-reported).
- 2025-10-29swe-1.5 and the swe-bench disclaimerswe-1.5 — a frontier-size in-house model served by cerebras at up to 950 tok/s, claimed 6x faster than haiku 4.5 and 13x than sonnet 4.5 on swe-bench pro (self-reported). the post's buried confession matters more than the benchmarks: 'we stopped reporting swe-bench numbers in 2024' — the company that minted the benchmark's pr value disclaimed it. (swe-grep and swe-grep-mini, fast retrieval models, shipped october 16.)
- 2025-11-14the 18-month review'devin's 2025 performance review: learnings from 18 months of agents at work' — the retrospective claiming devin embedded at 'thousands of companies' (self-reported). by late 2025 the team is reportedly ~200; linkedin later shows 300–400 across 21 countries.
- 2026-01infosys, devin review, londonthe january push: an infosys partnership puts devin across its org and client base (01-07); devin review launches as the 'ai to stop slop' (01-21); the london office opens — first in europe — alongside a cognizant partnership (01-28). the enterprise-distribution era starts.
- 2026-02-24devin 2.2 and cognition for governmentdevin 2.2 ships, billed as 'the most important update to devin since launch'; 'cognition for government' launches the next day. two days later 'how cognition uses devin to build devin' — the dogfooding account behind the 90%+ internal-code claims.
- 2026-03-19devins managing devins'devin can now manage devins' — delegation to parallel managed devins — plus scheduled devins. the agent starts supervising agents; swe-1.6 previewed earlier in the month.
- 2026-04-15devin inside windsurf'devin in windsurf' — the promised product integration ships: the devin agent inside the windsurf ide. 'local agents make you faster. cloud agents do the work while you're not there.' the same fortnight: swe-1.6, cognition japan (takumi masai), swe-check (an rl bug-detection model with applied compute), new self-serve plans, and a mercedes-benz deployment with the devin cli.
- 2026-04-22the multi-agents revisionwalden yan's follow-up, 'multi-agents: what's actually working,' narrows the 2025 thesis: multiple agents may contribute intelligence to a task while writes stay single-threaded.
- 2026-05-27series d, $26bthe series d: >$1b at $26b post ($25b pre), co-led by lux capital, general catalyst, and 8vc, with ribbit, atreides, and layer global new. run-rate revenue $492m versus $37m a year earlier; wu claims >90% of cognition's internal code is written by devin (self-reported); the positioning word is 'the independent agent lab.' customers named include goldman, mercedes-benz, citi, and the us army and navy.
- 2026-06-02devin desktopdevin desktop launches — 'the next generation of windsurf,' built around devin cloud and the agent command center. the acquisition's endgame: the ide becomes the agent's front end; by mid-2026 the phrasing is 'devin desktop, formerly windsurf.'
- 2026-06-04the productivity guaranteethe ai productivity guarantee: an in-house estimator converts devin's output into engineer-hours, and if delivered value falls short of the price, 'cognition will fund your usage up to $10m.' outcome-priced agent labor — self-measured, which is both the point and the caveat.
- 2026-07fedramp, genesis, the exec benchthe institutional summer: devin security swarm and the vulnerability remediation program (07-01/02); swe-1.7 (07-08); fedramp high (class d) in-process and the 'making fable cheaper than opus' post — naming the in-house fable model family (07-13); the windsurf anniversary post (07-14); acqui-hires of tierzero and the interaction company, makers of poke (07-20/23); an mou with the department of energy's genesis mission (07-22); the ltm partnership covering 260+ clients including 26 of the fortune 500 (07-28). russell kaplan announces the exec bench: paul grewal (ex-coinbase clo), art l. (ex-brex), jared palmer.
- 2026-09-08series e, $48bthe series e: >$2b at $48b led by new investors a16z and accel plus founders fund, general catalyst, avenir, and a ~30-firm syndicate including nvidia (also a devin customer). run-rate revenue ~$900m per company statements relayed by press. marc andreessen's post supplies the coda: 'in the past year, devin has gone from writing 13% of cognition's production code to more than 90%' — and a16z is 'proud to back him for the second time,' the first being lunchclub.
- 2026-09-09swe-2, rsa-260, awsthe september coda: a 'factoring rsa-260' research post (devin plus a researcher built a top gpu lattice siever); the swe-2 model launch; the dioxus team (jonathan kelley et al.) joins; the fusion harness for the in-house 'fable' and 'astra' models lands in devin desktop and cli (claimed up to 39% more efficient); and a multi-year aws strategic collaboration agreement (09-15).
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.