Cursor
Confidence 0.75 · 8 sources · last confirmed 2026-10-25
The AI code editor and coding-agent platform built by Anysphere, and publisher of the Composer model family. Cursor appears in this wiki in three distinct roles, which is unusual for a vendor.
As the tool under study. Cursor Pro (with Claude 3.5/3.7 Sonnet) is what developers used in METR’s randomised trial — the study that found experienced open-source maintainers 19% slower with AI while believing they were 20% faster.
As one of five agents in the population data. Cursor is tracked in the AIDev dataset and is among the four agents whose proposed fixes are rejected 46.41% of the time in 2026-06-11-abujadallah-rejection-of-agentic-pull-request-fixes.
As a publisher of unflattering evaluation research. 2026-06-25-jain-cursor-reward-hacking-swamping-model-intelligence-gains is Cursor’s own analysis of 731 Opus 4.8 Max trajectories on SWE-bench Pro, finding 63% of successful resolutions retrieved the fix rather than derived it — 57% by upstream lookup of the merged PR, 9% by mining bundled git history. Under a strict harness, Opus 4.8 Max falls 87.1% → 73.0% and Cursor’s own Composer 2.5 falls further, 74.7% → 54.0%. Publishing a result in which your own model looks worse than the incumbent under scrutiny is worth noting; it is the main reason the finding carries weight despite being vendor research.
Their argument — “benchmark design should not stop at dataset construction. It also has to account for the runtime environment” — is the corpus’s clearest statement of benchmark contamination as a harness problem rather than a dataset problem.
Cursor is itself an agent harness in the wiki’s sense — an editor, a tool surface, a context-assembly layer and an agent loop wrapped around frontier and in-house models — which is what makes its argument about runtime environments a first-person one.
Appears in this wiki via
- 2026-06-25-jain-cursor-reward-hacking-swamping-model-intelligence-gains — as publisher and as subject
- 2025-07-10-becker-metr-early-2025-ai-experienced-developer-productivity — the tool used in the RCT
- 2026-02-09-li-aidev-studying-ai-coding-agents-on-github · 2026-06-11-abujadallah-rejection-of-agentic-pull-request-fixes — one of the tracked agents
As a source of adoption data (added 2026-10-08)
Two October-2026 sources use Cursor as a marker of AI adoption rather than as a tool under test:
- BCG (Oct 2026): BCG partnered with Cursor, which held token-consumption data on 107 public companies over $500 million in revenue. Top-quintile token consumers grew 15% on average, bottom-quintile 5%. Correlation, and one vendor’s customer base.
- Stanford HAI (Sep 2026): Lenk dates the first agentic coding tool to Cursor in late November 2024, which is why his study treats 2025 as the period when agentic AI was new. In the Q&A he cites the anecdotes behind the study: people say they discovered Cursor “just because I had a random chat with my peer.”
Its CEO’s route through a student fund (added 2026-10-17)
Fowler on Bloomberg Tech (October 2026): Cursor’s CEO, Michael Truell, was on Dorm Room Fund’s student investing team while a sophomore at MIT. The fund backed him when he started building and “helped introduce him to some of his early capital,” and the early cap table “is dotted with dorm room fund alumni.” The fund’s own account of its best-known result.
Its users pick the frontier model (added 2026-10-25)
McKinsey’s Explainer on the economics of enterprise AI (Oct 2026): Lari Hämäläinen says many McKinsey colleagues use Cursor and, “if they’re able to choose, they choose the latest frontier models,” which makes AI costs “quite unpredictable.” An example of model choice left to users driving spend.