Cognition seeks $1 billion at $47 billion while its own benchmark crowns its rivals
Investors are paying roughly 52 times run-rate for a coding pure-play whose own launch table shows cheap open-weight Chinese models beating its new model two to one, and which now sells Anthropic's frontier model inside Devin.
Vincent Jiangtwo-minute read
Cognition~$10B — Cognition, private, latest valuation ~$10B, the maker of the Devin coding agent, is set to raise about $1 billion at a roughly $47 billion valuation, per Bloomberg, about 81% above its May mark of $26 billion, with nearly $10 billion of investor interest chasing the round 1. The round is not closed, and the run-rate above $900 million is not audited revenue: it annualizes recent weeks, no full-year figures exist, and the growth from $73 million has never been separated from Windsurf, acquired in July 2025 with $82 million of annual recurring revenue 1.
Paying 52 times for a model its own table ranks fourth
At $47 billion on a run-rate above $900 million, investors are paying about 52 times, against roughly 2.7 times for public go-to-market software 1. Five days into November, Cognition launched SWE-2 Flash, a model it calls "fully American," at one-fifth SWE-2's price and three times the speed, and published its own benchmark table with the launch 2.
On Terminal-Bench 4.0, SWE-2 Flash scores 20.0% 2. GLM 5.3 Flash scores 41.8 and DeepSeek V4.1 Flash 42.3 2. The open-weight models the table does not count as rivals beat the launch model about two to one, and GLM-5.3-Flash sells at $0.15 per million input tokens on Z.ai's pricing page 5.
On Cognition's own launch table, open-weight models beat SWE-2 Flash two to one
Data
| Value | |
|---|---|
| SWE-2 Flash | 20% |
| SWE-2 | 27.3% |
| GLM 5.3 Flash | 41.8% |
| DeepSeek V4.1 Flash | 42.3% |
| Inkling-Small (base) | 0.9% |
The moat now carries a rival's frontier model
On November 9, Cognition put Anthropic's$350B — Anthropic, private, latest valuation $350B Claude Fable 5.2 inside Devin Desktop and CLI 4. Anthropic puts Fable 5.2 at 49.1% on FrontierCode 2.0, Cognition's own coding benchmark, where Cognition's table lists its SWE-2 at 23.8 24. The company selling its own model as the moat now distributes the model that tops its leaderboard, and the scores cross harnesses: a benchmark result does not transfer automatically into Devin's environment 4.
Holders carry the risk if the mark does not hold
Cognition's holders carry the risk if the $47 billion mark does not hold 7. Anthropic's IPO buyers take the next swing: about 825 million shares at $70 to $78 for a $1.33 trillion valuation, days after shipping the model Cognition now resells, priced at $8 and $40 per million tokens 7.
Devin resells frontier tokens at $8 a million against 15 cents from open weights
Data
| Value | |
|---|---|
| GLM 5.3 Flash | $0.15 |
| Claude Fable 5.2, $8 rate | $8 |
| Claude Fable 5.2, $40 rate | $40 |
The productivity story underneath is real: pull requests rose eighteenfold on doubled headcount, ninefold per engineer, with about 80 engineers among roughly 500 staff and no product managers or QA 8. But an agent built to rent frontier intelligence is a spread business, and the spread is being priced down weekly by open weights 5.
Get Deals by email, free
The acquisitions, rounds and contracts that move AI money. In our daily email, every weekday.
Reader comments
0 comments