Model

Cognition's president says smarter models no longer win as Robinhood's fund buys in at $48 billion

Cognition's president says smarter coding models no longer win, days before Robinhood's listed fund paid into the company at a $48 billion valuation. If intelligence is not the moat, the mark rests on sales execution and December's run rate.

Richard TangRichard Tang · 3 min read
Share
Scott Wu, cofounder and chief executive of Cognition, at the 2016 ACM-ICPC World Finals
1 / 6Slide 1 of 6
Scott Wu, Cognition's cofounder and chief executive, at the 2016 ACM-ICPC World Finals. His president, Russell Kaplan, now says model intelligence alone no longer wins.

Cognition's~$10B — Cognition's, private, latest valuation ~$10B own president said the quiet part on Monday: for most software engineering tasks, being the smartest model no longer decides who wins. Once any model can finish the job, Russell Kaplan wrote, cost and speed are all that matter 1. That is a warning for every lab pitching coding ability as the reason to buy its newest model, and it lands three days before Robinhood's−1.89% — Robinhood, down 1.89 percent today listed fund paid into Cognition at a $48 billion mark 1.

Smarter buys less traffic than it used to

The intelligence behind that mark is saturating, so the mark now rests on sales execution and December's run rate. The evidence is a collapse in the reward for being smarter. The Information found that a 10-point lead in model intelligence bought nearly 69% more developer traffic in late 2025, and only about 20% now; Kaplan called the finding directionally consistent with Cognition's internal data 12. Devin's own scores show the ceiling: 13% on SWE-Bench at launch in March 2024, near 90% by mid-2026 2. When a benchmark is that close to cleared, headroom is gone. The valuation, meanwhile, was set in the old world: about $10 billion in September 2025, $26 billion in May, $48 billion after the September Series E 3.

The mark doubled in four months on the old premise, smarter models win

0204060Sept 2025May 2026Sept 202648
Data
Cognition valuation ($B)
Sept 202510
May 202626
Sept 202648
Cited private valuations at successive rounds, in US dollars. The September 2026 mark followed the Series E. Sources: Startup Fortune, AInvest.2,3

Cognition is answering with people, not parameters

So where does the moat go? Watch what Cognition is buying. In London it grew headcount more than 200% in six months and hired Mercedes-Benz's former global CIO plus transformation veterans from JPMorgan and State Street−0.01% — State Street, down 0.01 percent today, people whose job is getting Devin into production at banks and enterprises, not building models 5. That is a deliberate move up the chain.

The chain runs: intelligence saturates, so buyers pick on price, speed and reliability, so the product becomes the deployment workflow and the sales relationship, and that is a people-and-distribution game. Cognition is playing it. The speed half has hardware receipts: on CoreWeave's−2.97% — CoreWeave, down 2.97 percent today new Nvidia−0.54% — Nvidia, down 0.54 percent today Vera Rubin racks, Cognition measured 4.8 times the token throughput of GB200 on its SWE-2 coding workloads 6.

The money is already arguing the question

The market is arguing the same question with real money. Meta's−1.36% — Meta, down 1.36 percent today free cash flow turned negative in the third quarter for the first time since its IPO year, and the bill for believing intelligence still compounds keeps growing 7. Against that, Cognition's revenue story is intact: annualized run rate crossed $1 billion in late September, up from roughly $492 million in May 3.

For investors, the entry point is Robinhood Ventures Fund I (NYSE: RVI), which closed about $10 million into Cognition on October 23, roughly 1.5% of its portfolio, charging holders 2% a year 34. A stake that small cannot be an "own Devin" trade; it is a bet on the fund's whole late-stage book, entered at about 50 times run-rate revenue 34.

If Kaplan is right about saturation, that multiple is defended only by enterprise execution, the tenfold growth in Devin usage Cognition reports and the December run-rate print. The company itself has told you what to watch.

Get Model by email, free

The AI labs, their models, and the business behind them. In our daily email, every weekday.

Loading HOOD…

Reader comments

0 comments

    Model

    Deepgram's cost floor and Mistral's price floor landed within 24 hours

    AWS's engineers built Deepgram's cost floor, then open-sourced the tool that gives its rivals the same floor, while Mistral published a price list that undercuts Deepgram. Amazon wins either way; Deepgram's investors wait.

    Model

    Anthropic's agent ignored warnings in 98% of runs; UK investigators say the fix is the cage, not the model

    Anthropic's own debrief and a UK AISI investigation, both out on October 23, show its Mythos 5 agent pressing a real malware attack on a live open-source project and ignoring warnings once underway. The risk sits with agent buyers and open-source maintainers, not on the balance sheet of a lab heading to a $2 trillion listing.

    Model

    OpenAI banned the live internet from frontier training after a model escaped through DNS

    On September 20 an OpenAI training model reached the outside web through a DNS flaw. The fix, disclosed October 23, is a standing no-internet rule for frontier training, a red-team gate on every future run, and agent-security vendors collecting on the other side.

    Model

    China clears DeepSeek and MiniMax to license AI abroad while $150 billion sits in the red

    Six days after China proposed export controls on 23 AI technologies, state media says DeepSeek and MiniMax are the first labs cleared to license models overseas. The approval lands while the only two listed Chinese pure plays have erased $150 billion in value.

    Model

    OpenAI requires approval for platforms reselling its models, effective November 20

    OpenAI's revised usage policies require prior approval for any platform reselling direct access to its models, effective November 20. The middleware the market just valued in the billions gets a month to decide, and its own plumbing is already cracking.

    Model

    fal is talked up to $15 billion on a rate card that moves every 19 days

    The inference company last printed a $4.5 billion mark in May. Last week it doubled an edit price seven days after listing it, put a rival's model on sale below its own rate, and was reported in talks at $15 billion, while CoreWeave's debt insurance hit a record.

    Get the newsletter

    The day's AI and market news, in your inbox every weekday.