CursorBench 3.1

Perceptions of Anthropic / Claude models

• Several commenters complain recent Anthropic models overthink, burn tokens, and spawn unnecessary sub‑agents, sometimes ignoring instructions not to browse or read extra docs.
• Others report Opus 4.8 works well when configured carefully, but becomes very slow at high “effort” levels, especially for implementation vs planning.
• Concerns that increased token usage (including a new tokenizer that inflates token counts) might be economically motivated, especially as older models are deprecated.

Composer 2.5 vs GPT‑5.5, Opus, and others

• Many users like Composer 2.5 as a fast, cheap “daily driver” for typical web and app work, especially when tasks are broken into small steps and plans come from stronger models.
• Others find Composer 2.5 shallow, “college‑project” quality, or even net‑negative productivity on harder or non‑standard tasks (e.g., physics, kinematics, tricky engineering), saying it confidently makes subtle mistakes and doesn’t ask for clarification.
• Some claim its code quality is close to Opus or GPT‑5.5 for routine CRUD and UI work; others say that’s “farcical” and GPT‑5.5 or Opus are clearly superior.
• There is repeated emphasis that language/stack, problem domain, and horizon length strongly affect perceived quality.

Benchmark validity and bias

• Strong skepticism toward CursorBench because Cursor’s own model performs nearly as well as frontier models at much lower cost.
• People contrast this with third‑party evals (e.g., DeepSWE, FrontierCode), where Composer 2.5 scores much lower on long‑horizon tasks.
• Some note any vendor is incentivized or naturally inclined to design benchmarks aligned with its training distribution.
• Others argue all benchmarks are flawed: harness design, lack of caching, and unrealistic task style can distort results.

Speed, cost, and context

• Many prioritize wall‑clock time: Opus is frequently described as painfully slow compared to Composer, forcing parallel sessions or off‑hours usage.
• GPT‑5.5 is often seen as the strongest pure programmer, but context limits and pricing models complicate comparisons.
• Large context windows (Opus 1M, optional larger GPT) matter most for long, planning‑heavy sessions; compaction is said to harm those workflows.

Graph and presentation critiques

• Multiple people dislike the reversed cost axis as unintuitive and potentially manipulative.
• Some call for Pareto‑frontier style plots (cost vs performance, plus speed) and more transparent, multi‑benchmark reporting.