cursor.com · benchmark source

Cursor

Cursor’s first-party agent evals for IDE-native multi-file work. CursorBench is vendor-run from real Cursor sessions — useful inside Cursor, not a public reproducible harness.

Open cursor.com3 boards on this publisher

Coding agents

CursorBench

Cursor’s offline correctness suite for ambiguous, multi-file coding-agent tasks sourced from real Cursor sessions rather than public GitHub issues.

Why it matters

It separates frontier models on IDE-style agent work where SWE-bench and similar public suites are increasingly saturated or contaminated.

Use it when

You are choosing a model for Cursor’s agent loop and want a signal aligned with Cursor’s own product evals — then validate on your repo.

The limitation

First-party and not independently reproducible from a public harness. Scores move when Cursor refreshes the problem distribution (compare only within the same version).

Leaderboard snapshot

CursorBench — chart and full table

Same layout as Cursor’s public board: Cost / Tokens / Steps chart, then one full Score · Cost · Tokens · Steps table.

Local leader

Grok 4.6 Extra High

70.8%

Rows shown

Full CursorBench 3.2 table

Snapshot date

2026-08-14

Mirrored from cursor.com/cursorbench

Better is toward the top-right (higher score, lower cost). X-axis is reversed to match Cursor’s public chart.

CursorBench 3.2 full leaderboard with score, cost, tokens, and steps per task
#	Model	Score	Cost / task	Tokens / task	Steps / task
1	Grok 4.6 Extra High	70.8%	$2.81	41,136	46
2	Fable 5 Max	70.5%	$17.32	103,525	72
3	Opus 5 Max	70.0%	$8.23	61,838	78
4	Grok 4.6 High	69.9%	$2.34	32,449	39
5	Opus 5 Extra High	69.3%	$7.35	54,239	72
6	Fable 5 Extra High	68.4%	$11.73	64,971	56
7	GPT-5.6 Sol Max	67.2%	$5.69	28,320	48
8	Grok 4.6 Medium	67.1%	$1.28	17,942	29
9	Opus 5 High	66.7%	$3.91	27,932	48
10	Fable 5 High	66.5%	$8.77	43,747	48
11	Fable 5 Medium	65.2%	$6.80	30,366	41
12	GPT-5.6 Terra Max	64.9%	$2.31	32,969	47
13	GPT-5.6 Sol Extra High	64.5%	$3.88	19,699	38
14	Opus 5 Medium	64.3%	$3.29	23,612	44
15	GPT-5.6 Sol High	63.5%	$2.79	13,867	32
16	Opus 5 Low	62.8%	$2.55	18,529	37
17	Opus 4.8 Max	62.3%	$5.77	71,411	44
18	Fable 5 Low	62.1%	$4.46	18,182	31
19	Gemini 3.7 Flash High	61.6%	$1.20	38,448	99
20	Sonnet 5 Max	61.5%	$4.30	92,882	86
21	GPT-5.6 Luna Max	61.1%	$0.39	87,973	61
22	Grok 4.6 Low	61.0%	$0.70	10,658	23
23	Kimi K3 Max	60.8%	$2.70	38,428	57
24	GPT-5.6 Sol Medium	60.0%	$1.95	9,747	27
25	Kimi K3 High	59.7%	$1.89	26,846	47
26	Opus 4.8 Extra High	59.4%	$4.50	51,121	40
27	GPT-5.6 Terra Extra High	59.2%	$1.15	16,089	29
28	Gemini 3.7 Flash Medium	59.0%	$0.95	30,953	82
29	Sonnet 5 Extra High	58.7%	$2.77	52,871	67
30	GPT-5.5 Extra High	58.4%	$2.85	17,534	32
31	GPT-5.5 High	58.4%	$2.05	12,183	28
32	Opus 4.8 High	58.0%	$3.15	33,548	33
33	GPT-5.6 Luna Extra High	57.7%	$0.23	22,480	48
34	Sonnet 5 High	56.9%	$2.13	39,483	57
35	GPT-5.6 Luna High	56.8%	$0.16	15,141	40
36	Composer 2.5	56.1%	$0.44	14,286	33
37	Opus 4.8 Medium	56.1%	$2.81	28,384	32
38	GLM 5.2 Max	55.0%	$1.76	35,946	58
39	GPT-5.6 Terra High	54.2%	$0.71	9,468	23
40	GPT-5.5 Medium	53.8%	$1.51	8,522	25
41	Gemini 3.7 Flash Low	53.8%	$0.74	20,594	68
42	Gemini 3.6 Flash High	53.5%	$1.56	30,436	64
43	Opus 4.8 Low	53.1%	$2.02	19,624	27
44	GPT-5.6 Sol Low	52.6%	$1.01	5,104	19
45	Sonnet 5 Medium	52.4%	$1.44	26,200	46
46	GLM 5.2 High	51.5%	$1.19	21,829	49
47	Gemini 3.6 Flash Medium	51.2%	$1.48	28,511	62
48	Kimi K3 Low	50.5%	$0.99	13,007	33
49	GPT-5.6 Terra Medium	50.3%	$0.49	6,222	20
50	Kimi K2.7 Code	49.7%	$1.43	31,247	58
51	GPT-5.6 Luna Medium	47.7%	$0.08	7,095	28
52	Sonnet 5 Low	47.7%	$0.87	16,269	33
53	Gemini 3.6 Flash Low	47.4%	$1.13	20,529	50
54	GPT-5.6 Terra Low	46.9%	$0.42	5,312	19
55	GPT-5.5 Low	46.6%	$0.98	5,168	20
56	GPT-5.6 Luna Low	37.6%	$0.03	3,209	17

Avg cost / task applies each model’s published per-million-token pricing to tokens used on each task. This snapshot has no row-level training-data disclosure markers. Small score gaps may not be statistically meaningful — confirm on cursor.com/cursorbench.

Evidence in this catalog

Where CursorBench fits

Closest verified examples we currently carry. A missing score is not a zero.

Model examples

Grok 4.6 Extra High70.8% · $2.81 / task
Top published CursorBench 3.2 row (full table mirrored 2026-08-12).
Fable 5 Max70.5% · $17.32 / task
Within a point of Grok 4.6 Extra High — small gaps may not be meaningful.
Opus 5 Max70.0% · $8.23 / task
Strong correctness at lower average cost than Fable 5 Max.

Harness examples

CursorIDE agent harness
CursorBench scores models inside Cursor’s agent loop, not bare API chat.

Open this board on CursorBench (live)