CursorBench is Cursor's own evaluation of coding agents. Ambiguous, multi-file tasks taken from real Cursor sessions, graded on whether the agent finished the job. The current version, CursorBench 4.0, was published on 10 September 2026. It reports four numbers per configuration: score, average cost per task, average tokens per task and average steps per task.

Which model scores highest on CursorBench?

Opus 5.5 at Max effort, with 57.8%. The next three places are also Opus 5.5 at lower effort levels, so the top of the table is one model rather than four.

Is the highest-scoring model worth the money?

Rarely. Opus 5.5 Max costs $13.43 a task for 57.8%, while Opus 5.5 High costs $3.97 for 56%. That is 3.4 times the price for 1.8 points.

Are small score differences meaningful?

Cursor says explicitly that results are subject to variance and that small differences may not be statistically meaningful. Two configurations a point apart should be treated as tied, and the cost column used to break the tie.

Does a high CursorBench score mean a model is better for my work?

Not necessarily. The tasks are drawn from real Cursor sessions, which makes the benchmark representative of agentic editing in a large codebase, not of writing an isolated function, answering a question about a library, or working in a language the sample under-represents.

Why does the same model appear many times?

Each row is a model at one reasoning effort level. Effort changes how long the model thinks and how many steps it takes, and it moves both score and cost substantially. Opus 5.5 spans 43.7% to 57.8% depending only on effort.

Does effort level cost more even at the same model price?

Yes, because cost is driven by tokens rather than by a per-request fee. Higher effort means more thinking tokens and more tool steps, so the same per-million rate produces a much larger bill.

How often does CursorBench change?

The task set has been revised four times since March 2026, and scores are also restated when a provider changes its prices. Scores from different CursorBench versions are not comparable.

CursorBench scores

Ambiguous, multi-file tasks taken from real Cursor sessions, graded on whether the agent finished the job. Here is the whole CursorBench 4.0 table, plus the part the ranking hides: what each point of score actually costs.

CursorBench 4.0
Published 10 September 2026
52 configurations

Checked against cursor.com on 25 September 2026

The short answer

Opus 5.5 holds the top four places. But the fourth-place configuration scores 52.5% at $2.91 a task while first place scores 57.8% at $13.43. You pay 4.6 times as much for 5.3 points.

Highest score

57.8%

Opus 5.5 Max, $13.43 per task

Best under $1 a task

33.0%

GPT-5.6 Luna Extra High, $0.44 per task

Widest effort spread

14.1 pts

Opus 5.5 from Low to Max, one model, no change of provider

Points per dollar

The same table sorted by dollars spent per percentage point scored. This is not a ranking Cursor publishes, and it rearranges the board completely: the cheapest models rise to the top because their scores, while lower, cost almost nothing to obtain.

Model	Effort	Score	Cost / task	Cost per point
GPT-5.6 Luna	Low	16.0%	$0.03	$0.002
GPT-5.6 Luna	Medium	22.2%	$0.08	$0.004
GPT-5.6 Luna	High	29.4%	$0.25	$0.009
GPT-5.6 Luna	Extra High	33.0%	$0.44	$0.013
GPT-5.6 Terra	Low	25.2%	$0.52	$0.021
Muse Spark 1.3	Minimal	24.3%	$0.56	$0.023
GPT-5.6 Terra	Medium	27.6%	$0.64	$0.023
Composer 2.5	Default	27.7%	$0.68	$0.025

Cheapest per point does not mean best. A configuration that scores 16% is wrong on five tasks in six, and the cost of being wrong is not on this table.

The best score at each budget

A more useful way to read the board. Pick what you are willing to spend on one agent task, and this is the highest score available at that price.

Budget per task	Best configuration	Score	Actual cost
$0.50 or less	GPT-5.6 Luna Extra High	33.0%	$0.44
$1.00 or less	GPT-5.6 Luna Extra High	33.0%	$0.44
$2.00 or less	Opus 5.5 Low	43.7%	$1.17
$4.00 or less	Opus 5.5 High	56.0%	$3.97
$8.00 or less	Opus 5.5 Extra High	56.0%	$6.98
$20.00 or less	Opus 5.5 Max	57.8%	$13.43

Note how flat the middle is. Between $2 and $8 a task the score moves a few points, which is the range most everyday work sits in.

The full CursorBench 4.0 table

All 52 configurations as published, in rank order. Tokens and steps are per task and explain most of the cost column.

#	Model	Effort	Score	Cost / task	Tokens / task	Steps / task
1	Opus 5.5	Max	57.8%	$13.43	218,363	185
2	Opus 5.5	Extra High	56.0%	$6.98	101,083	109
3	Opus 5.5	High	56.0%	$3.97	53,078	68
4	Opus 5.5	Medium	52.5%	$2.91	37,954	54
5	Fable 5.1	Max	51.8%	$17.28	117,236	128
6	Fable 5.1	Extra High	51.6%	$13.01	87,294	101
7	Fable 5.1	High	49.2%	$9.08	58,438	77
8	Fable 5.1	Medium	46.8%	$7.05	45,411	63
9	Opus 5	Max	46.6%	$11.95	85,384	106
10	Grok 4.7	Extra High	46.3%	$6.01	70,141	88
11	Opus 5	Extra High	46.1%	$11.43	80,094	103
12	Fable 5.1	Low	45.1%	$5.44	34,795	51
13	Opus 5	High	44.7%	$9.00	61,405	86
14	Grok 4.7	High	43.9%	$4.69	56,382	71
15	Opus 5.5	Low	43.7%	$1.17	15,811	28
16	Opus 5	Medium	43.3%	$6.94	45,272	72
17	GPT-5.6 Sol	Max	41.7%	$8.23	42,944	99
18	Grok 4.7	Medium	41.6%	$3.49	36,683	60
19	Muse Spark 1.3	Max	41.6%	$2.64	52,005	98
20	Grok 4.6	Extra High	41.4%	$6.10	49,814	56
21	GPT-5.6 Terra	Max	41.3%	$5.14	60,814	107
22	Opus 5	Low	40.7%	$4.87	31,995	57
23	Grok 4.6	High	40.4%	$5.20	41,387	48
24	Gemini 3.8 Flash	High	39.6%	$4.70	162,565	324
25	GPT-5.6 Sol	Extra High	37.7%	$4.40	24,729	55
26	Muse Spark 1.3	Extra High	37.5%	$2.10	40,891	83
27	Gemini 3.8 Flash	Medium	37.3%	$4.06	128,364	290
28	Grok 4.6	Medium	36.1%	$3.48	24,893	40
29	GPT-5.6 Luna	Max	35.9%	$1.03	87,284	208
30	GPT-5.6 Sol	High	35.7%	$2.85	16,174	41
31	Sonnet 5	Max	34.1%	$7.17	149,257	140
32	GPT-5.6 Terra	Extra High	33.6%	$1.81	23,436	43
33	Grok 4.6	Low	33.4%	$2.25	16,307	32
34	Muse Spark 1.3	High	33.4%	$1.66	30,654	69
35	Grok 4.7	Low	33.1%	$1.58	15,677	40
36	GPT-5.6 Luna	Extra High	33.0%	$0.44	40,598	98
37	Muse Spark 1.3	Medium	32.6%	$1.49	27,255	64
38	Sonnet 5	Extra High	32.0%	$4.55	83,373	102
39	GPT-5.6 Sol	Medium	31.1%	$1.77	10,111	32
40	Sonnet 5	High	30.8%	$3.48	61,146	85
41	GPT-5.6 Terra	High	30.7%	$1.11	13,162	33
42	GPT-5.6 Luna	High	29.4%	$0.25	23,368	64
43	Muse Spark 1.3	Low	29.3%	$0.93	17,483	47
44	Sonnet 5	Medium	28.0%	$2.31	39,114	65
45	Composer 2.5	Default	27.7%	$0.68	17,347	41
46	GPT-5.6 Terra	Medium	27.6%	$0.64	7,307	25
47	GPT-5.6 Terra	Low	25.2%	$0.52	5,914	23
48	GPT-5.6 Sol	Low	24.6%	$0.87	4,885	21
49	Muse Spark 1.3	Minimal	24.3%	$0.56	10,620	34
50	Sonnet 5	Low	24.1%	$1.39	23,772	46
51	GPT-5.6 Luna	Medium	22.2%	$0.08	7,642	32
52	GPT-5.6 Luna	Low	16.0%	$0.03	3,288	18

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. Read from cursor.com/cursorbench on 25 September 2026.

How the benchmark has changed

Four task-set revisions since March 2026, plus restatements when a provider changed its prices. Scores from different versions are not comparable, which is why an old screenshot of this leaderboard proves very little.

CursorBench 4.010 September 2026
Long-horizon problems covering edit, refactor, investigation, intent understanding, managing jobs, and design adherence.
CursorBench 3.28 July 2026
Instruction following and advanced tool use problems.
CursorBench 3.119 May 2026
Codebase understanding, bugfinding, planning and code review problems, plus better grading on some edit tasks.
CursorBench 3.011 March 2026
Initial task set: edit, refactor and bugfix problems.

What the numbers do not say

A benchmark published by the company selling the product is worth reading and worth discounting. CursorBench is unusually transparent about its own limits: it names the variance problem, it publishes cost alongside score rather than score alone, and it lists every revision to the task set. Those are not the habits of a marketing chart. They are also not independent verification.

Three things the table cannot tell you. It cannot tell you how a model behaves in your language and framework, because the task mix is drawn from Cursor sessions in aggregate. It cannot tell you how a model feels to work with, which is what most people actually switch models over: how often it asks before acting, how much it over-edits, whether it gives up. And it measures completed tasks rather than the cost of failures, so a model that fails loudly is treated the same as one that fails quietly and leaves a broken branch behind.

The practical use of this page is narrower than a ranking. Use it to find the effort level where a model stops improving, and to notice when a cheaper model at high effort beats an expensive one at low effort. Both of those are decisions the table supports well.

Frequently asked questions

What is CursorBench?: CursorBench is Cursor's own evaluation of coding agents. Ambiguous, multi-file tasks taken from real Cursor sessions, graded on whether the agent finished the job. The current version, CursorBench 4.0, was published on 10 September 2026. It reports four numbers per configuration: score, average cost per task, average tokens per task and average steps per task.
Which model scores highest on CursorBench?: Opus 5.5 at Max effort, with 57.8%. The next three places are also Opus 5.5 at lower effort levels, so the top of the table is one model rather than four.
Is the highest-scoring model worth the money?: Rarely. Opus 5.5 Max costs $13.43 a task for 57.8%, while Opus 5.5 High costs $3.97 for 56%. That is 3.4 times the price for 1.8 points.
How is CursorBench cost per task calculated?: 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.
Are small score differences meaningful?: Cursor says explicitly that results are subject to variance and that small differences may not be statistically meaningful. Two configurations a point apart should be treated as tied, and the cost column used to break the tie.
Does a high CursorBench score mean a model is better for my work?: Not necessarily. The tasks are drawn from real Cursor sessions, which makes the benchmark representative of agentic editing in a large codebase, not of writing an isolated function, answering a question about a library, or working in a language the sample under-represents.
Why does the same model appear many times?: Each row is a model at one reasoning effort level. Effort changes how long the model thinks and how many steps it takes, and it moves both score and cost substantially. Opus 5.5 spans 43.7% to 57.8% depending only on effort.
Does effort level cost more even at the same model price?: Yes, because cost is driven by tokens rather than by a per-request fee. Higher effort means more thinking tokens and more tool steps, so the same per-million rate produces a much larger bill.
How often does CursorBench change?: The task set has been revised four times since March 2026, and scores are also restated when a provider changes its prices. Scores from different CursorBench versions are not comparable.