HN 日本語サマリー

← 一覧へ戻る
AI・機械学習

CursorBench 3.1

CursorBench 3.1 (cursor.com)

11 pointsby handfuloflight2 コメント

要約

CursorBench 3.1は、AIモデルのコード関連タスクにおけるパフォーマンスとコストを評価するためのベンチマークです。Fable 5、Opus 4.8、GPT-5.5、Sonnet 5などの様々なモデルを、コード理解、バグ発見、計画、コードレビューといったタスクで比較しています。結果は、モデルのスコアとタスクあたりの平均コストを示しており、パフォーマンスと価格のトレードオフを分析できます。

全文翻訳

CursorBench 4.0

曖昧で、複数のファイルにまたがる実際のCursorセッションからのタスクでエージェントを評価します。スコアが高いほど優れています。

CursorBenchについてさらに詳しく

Fable 5.1、Opus 5.5、Opus 5、Grok 4.7、Grok 4.6、GPT-5.6 Sol、GPT-5.6 Terra、GPT-5.6 Luna、Sonnet 5.5、Sonnet 5、Haiku 5.5、Gemini 4 Argon、Gemini 3.8 Flash、Muse Wonder、Muse Spark 1.3、GLM 5.3、GLM 5.3 Flash、Composer 2.5のスコアと、タスクあたりの平均コストを比較した散布図と折れ線グラフ。

[画像: CursorBench 4.0のスコアとタスクあたりの平均コストを比較したグラフ。モデル名がX軸、スコアが左Y軸、コストが右Y軸にプロットされている。]60%CursorBench 4.0 score55%50%45%40%35%30%25%20%$18$15$12$9$6$3$0Average cost per taskGemini 4 ArgonOpus 5.5Fable 5.1Sonnet 5.5Muse WonderHaiku 5.5GPT-5.6 SolGrok 4.7

Model
1Opus 5.5 Max57.8%$13.43218,363185
2Opus 5.5 Extra High56.0%$6.98101,083109
3Opus 5.5 High56.0%$3.9753,07868
4Sonnet 5.5 Max55.5%$7.05271,920170
5Sonnet 5.5 Extra High53.1%$2.81100,15878
6Gemini 4 Argon High52.9%$4.57109,500118
7Gemini 4 Argon Medium52.8%$4.39103,600115
8Opus 5.5 Medium52.5%$2.9137,95454
9Fable 5.1 Max51.8%$17.28117,236128
10Fable 5.1 Extra High51.6%$13.0187,294101
11Gemini 4 Argon Low50.6%$4.0994,200110
12Muse Wonder Max50.4%$5.2549,692114
13Fable 5.1 High49.2%$9.0858,43877
14Haiku 5.5 Max48.4%$1.12325,934163
15Muse Wonder Extra High48.2%$4.2439,56496
16Sonnet 5.5 High47.8%$1.2037,39141
17Fable 5.1 Medium46.8%$7.0545,41163
18Opus 5 Max46.6%$11.9585,384106
19Grok 4.7 Extra High46.3%$6.0170,14188
20Opus 5 Extra High46.1%$11.4380,094103
21Muse Wonder High46.1%$3.5133,61284
22Muse Wonder Medium45.2%$2.8227,12475
23Fable 5.1 Low45.1%$5.4434,79551
24Opus 5 High44.7%$9.0061,40586
25Haiku 5.5 Extra High44.3%$0.56143,81393
26Grok 4.7 High43.9%$4.6956,38271
27Opus 5.5 Low43.7%$1.1715,81128
28Opus 5 Medium43.3%$6.9445,27272
29GLM 5.3 Max42.6%$5.0596,387166
30Haiku 5.5 High42.3%$0.3277,05759
31GPT-5.6 Sol Max41.7%$8.2342,94499
32Muse Wonder Low41.7%$1.9118,88056
33Grok 4.7 Medium41.6%$3.4936,68360
34Muse Spark 1.3 Max41.6%$2.6452,00598
35Grok 4.6 Extra High41.4%$6.1049,81456
36GPT-5.6 Terra Max41.3%$5.1460,814107
37Opus 5 Low40.7%$4.8731,99557
38Grok 4.6 High40.4%$5.2041,38748
39Gemini 3.8 Flash High39.6%$4.70162,565324
40Sonnet 5.5 Medium39.2%$0.5216,03622
41GLM 5.3 High38.0%$3.2460,031114
42GPT-5.6 Sol Extra High37.7%$4.4024,72955
43Muse Spark 1.3 Extra High37.5%$2.1040,89183
44Gemini 3.8 Flash Medium37.3%$4.06128,364290
45Haiku 5.5 Medium36.9%$0.1742,65939
46GLM 5.3 Flash Max36.8%$0.3956,410118
47Grok 4.6 Medium36.1%$3.4824,89340
48GPT-5.6 Luna Max35.9%$1.0387,284208
49Sonnet 5.5 Low35.8%$0.3711,66818
50GPT-5.6 Sol High35.7%$2.8516,17441
51Muse Wonder Minimal34.2%$0.849,76133
52Sonnet 5 Max34.1%$7.17149,257140
53GPT-5.6 Terra Extra High33.6%$1.8123,43643
54Grok 4.6 Low33.4%$2.2516,30732
55Muse Spark 1.3 High33.4%$1.6630,65469
56GLM 5.3 Low33.3%$2.0431,98381
57Grok 4.7 Low33.1%$1.5815,67740
58GPT-5.6 Luna Extra High33.0%$0.4440,59898
59Muse Spark 1.3 Medium32.6%$1.4927,25564
60Sonnet 5 Extra High32.0%$4.5583,373102
61GPT-5.6 Sol Medium31.1%$1.7710,11132
62GLM 5.3 Flash High31.1%$0.2535,10484
63Haiku 5.5 Low30.9%$0.0823,38225
64Sonnet 5 High30.8%$3.4861,14685
65GPT-5.6 Terra High30.7%$1.1113,16233
66GPT-5.6 Luna High29.4%$0.2523,36864
67Muse Spark 1.3 Low29.3%$0.9317,48347
68Sonnet 5 Medium28.0%$2.3139,11465
69Composer 2.527.7%$0.6817,34741
70GPT-5.6 Terra Medium27.6%$0.647,30725
71GLM 5.3 Flash Low26.9%$0.1517,83158
72GPT-5.6 Terra Low25.2%$0.525,91423
73GPT-5.6 Sol Low24.6%$0.874,88521
74Muse Spark 1.3 Minimal24.3%$0.5610,62034
75Sonnet 5 Low24.1%$1.3923,77246
76GPT-5.6 Luna Medium22.2%$0.087,64232
77GPT-5.6 Luna Low16.0%$0.033,28818

Changelog

Oct 7, 2026
Reporting
  • Updated Sonnet 5.5 results to account for adjusted pricing.
Sep 10, 2026
Tasks
  • CursorBench 4.0
    • Introduced new long-horizon problems focused on edit, refactor, investigation, intent understanding, managing jobs, and design adherence.
Aug 11, 2026
Reporting
  • Updated Sonnet 5 results to account for adjusted pricing.
Jul 30, 2026
Reporting
  • Updated GPT-5.6 Terra and Luna results to account for adjusted pricing.
Jul 9, 2026
Reporting
  • Updated GPT-5.6 Sol, Terra, and Luna results to account for cache write costs.
Jul 8, 2026
Tasks
  • CursorBench 3.2
    • Introduced instruction following and advanced tool use problems.
May 19, 2026
Tasks
  • CursorBench 3.1
    • Introduced problems focused on codebase understanding, bugfinding, planning, and code review.
    • Improved grading criteria for some edit tasks.
Mar 11, 2026
Tasks
  • CursorBench 3.0
    • Initial set of tasks focused on edit, refactor, and bugfix problems.

Avg cost / task is computed by applying each model's published per-million-token pricing (input, cache read, cache write, and output) to the tokens it used on each task. Results are subject to variance; small differences in scores may not be statistically meaningful.