Cursor quietly adds the Chinese model that found a hole in its editor
SpaceX-owned Cursor put Zhipu's GLM 5.3 in its model picker this week with no announcement, at $1.40 a million input tokens. The silence is the tell: application-layer valuations built on resold frontier intelligence are meeting a Chinese commodity price floor.
Vincent Jiang · 2 min read
A settings line, not a press release
Cursor, the coding editor SpaceX agreed in June to buy for $60 billion in stock, added Zhipu's GLM 5.3 to its model picker this week with no announcement: a documentation page and a settings line, nothing more 12. The model lists at $1.40 per million input tokens and $4.40 per million output, flat across a one-million-token context window 1.
The model found the hole first
Days after GLM 5.3 shipped on 14 August, a Z.ai developer advocate said it had flagged a "potentially serious vulnerability" in Cursor itself 3; the flaw, as described, would let an attacker write files anywhere on a user's machine 1. Anthropic's Frontier Red Team then showed the model building working exploits in 50 of 410 attempts, with safeguards that give way in 64 to 100 percent of tests depending on the trick 4.
The quiet add is a habit
Composer 2, Cursor's in-house model, turned out to be a fine-tune of Moonshot's Kimi K2.5, caught when a developer read the model ID in API traffic; co-founder Aman Sanger called the omission "a miss" 5. Composer 2.5, still on that base, lists at $0.50 per million input tokens, an eighth of Claude Opus 5.5 or GPT-5.6 Sol, a twentieth of Claude Fable 5.1 16. GLM 5.3 slots in one rung above it 1.
The bottom of Cursor's price ladder is Chinese or Chinese-derived
Data
| Value | |
|---|---|
| Composer 2.5 | $0.5/M |
| GLM 5.3 | $1.4/M |
| Grok 4.7 | $2/M |
| GPT-5.6 Sol | $4/M |
| Claude Opus 5.5 | $4/M |
| Claude Fable 5.1 | $10/M |
The bill is the business
Cursor sold to SpaceX because it was severely compute-constrained, on annualized revenue above $2 billion 2. OpenAI pulls its models on 12 November; they carried roughly 5 percent of traffic, and Anthropic says it will pour in more compute for Claude instead 7. Cheap weights are not a preference here. They are the cost line.
The bid these prices sit against
Cognition raised $2 billion at $48 billion on 8 September on $900 million of company-claimed run-rate revenue, training its own model on open-source bases to cut the same bill, with 2026 burn put near $800 million 8. Vercel's gateway moved 61 percent of model spend to open weights in September, up from 40 percent in July 9. Anthropic filed to list at more than $2 trillion on $4.6 billion of 2025 revenue, an $8 billion operating loss, and up to $42 billion in Broadcom credit 10.
What to watch
Two dates price the question: OpenAI's 12 November cutoff in Cursor, and Cognition's year-end run-rate against its $4 to $5 billion guide 78. When coding intelligence costs $1.40 a million tokens at the floor, every application-layer valuation is a bet on something other than the tokens.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.
