Cursor Introduces Composer 2.5

Original Article ↗ Hacker News Discussion ↗

Model & Technical Approach

  • Composer 2.5 is built on Moonshot’s open Kimi K2.5 checkpoint, with extra RL and coding-focused fine-tuning.
  • Several comments note this isn’t a “from scratch” model; the “new from scratch” training is said to be a separate, larger model on SpaceX/xAI’s Colossus 2 cluster.
  • Some argue much of coding capability comes from RL and harness design, not just base model quality.

Performance, Benchmarks & Real-World Use

  • Benchmarks claim near-Opus / GPT frontier performance at ~1/10–1/16 the cost; many are skeptical, citing past Composer 2 claims that didn’t match real-world usage.
  • Multiple developers report Composer 2.5 (especially the fast variant) feels weaker than Opus / Claude Code / GPT for planning, code quality, bug avoidance, and session-level behavior.
  • Others say Composer 2.x is “good enough” and very fast for many day-to-day coding tasks, especially as a sub-agent or for autocomplete.

Pricing, Limits & Economics

  • Confusion around Cursor’s pricing tiers; some on $20 personal plans never hit limits, others report hitting caps or large bills on team plans.
  • Several teams report costs “skyrocketing” after switching to team plans or after fast mode became default; some companies are moving to Claude Code or Codex for cost reasons.
  • Debate over whether cheaper high-quality tokens will compress revenue or instead expand usage (Jevons-like effects).

Product, UX & Harness Quality

  • Strong praise for Cursor’s tab completion and coding harness when it works; some say it remains best-in-class for integrated coding workflows.
  • Many complaints about constant UI changes, regressions, bugs, lag, memory use, and weaker integration with GitHub and agents than competitors.
  • CLI and alternative harnesses (e.g., via Zed) exist but are described as immature or less capable.

Moat, Data & Strategy

  • Ongoing argument whether Cursor has a moat:
    • Skeptics: “still a VS Code fork,” open models are commoditized, big tech has more data.
    • Supporters: the IDE+harness+RL on rich coding/edit data could be defensible, especially with early large-scale usage.
  • Concerns that user code and interaction data are likely used for fine-tuning; some hope users “wake up” to this.

xAI / SpaceX Context & Future

  • Training on Colossus 2 and rumored acquisition by xAI are seen as giving massive compute and cash, but also raising questions about sustainability and strategy.
  • Some view Cursor’s ambition as impressive; others see it as a necessary move to escape low-margin dependence on external APIs.