In the CursorBench 4.0 leaderboard snapshot accessed on 2026-11-16, Gemini 4 Carbon High scored 60.4% and ranked first, Medium scored 60.1% and ranked second, and Low scored 58.6% and ranked fourth. The benchmark targets ambiguous, multi-file engineering tasks in Cursor sessions and does not represent all software development tasks.
Tasks it can help assess: Compare the model configurations listed on the CursorBench 4.0 page on scores, cost per task, token usage, and step count for multi-file engineering tasks with longer workflows.
Tasks the results should not be generalized to: The results do not establish success rates for single-file code generation or for a specific language or repository. They also cannot be generalized to other agents, IDEs, model APIs, or task sets.
Applicable model versions: Gemini 4 Carbon High, Medium, and Low as listed on the leaderboard. The page does not provide underlying model API IDs or revision numbers.
Test environment or client: Cursor's internal evaluation suite, CursorBench 4.0, using Cursor tasks and its Agent execution environment. More specific client versions, task runtime environments, and tool configurations are not specified.
Reasoning tiers and parameters: High, Medium, and Low. Sampling parameters and specific reasoning configurations are not specified.
Cursor describes CursorBench as a set of ambiguous, multi-file tasks drawn from real Cursor sessions. The 4.0 update history says it added long-horizon problems covering editing, refactoring, investigation, intent understanding, task management, and following design requirements. The page says higher scores are better. It also reports average cost, token count, and step count per task.
Costs are calculated using each model's published price per million tokens. They include input, cache-read, cache-write, and output tokens, summed across usage for each task. Cursor notes that results have variance and small score differences may not be statistically meaningful.
A benchmark methodology article published by Cursor on 2026-03-11 describes the source of tasks in an earlier CursorBench version: Cursor Blame links committed code to the Agent request that generated it. The article says the tasks came from real sessions with engineering teams, internal codebases, and controlled sources. It described CursorBench 3.1 at the time, so it does not confirm that task collection and scoring details remained entirely unchanged in 4.0. The current 4.0 page does not disclose the total number of tasks, runs per model, sample distribution, evaluator, confidence intervals, or per-task results.
| Configuration | Rank | Score | Average cost/task | Tokens/task | Steps/task |
|---|---|---|---|---|---|
| Gemini 4 Carbon High | 1 | 60.4% | $9.22 | 39,960 | 71 |
| Gemini 4 Carbon Medium | 2 | 60.1% | $8.49 | 37,067 | 67 |
| Gemini 4 Carbon Low | 4 | 58.6% | $11.89 | 42,663 | 124 |
| Fable 5.2 Max | 3 | 60.0% | $32.79 | 352,794 | 218 |
| Opus 5.5 Max | 5 | 57.8% | $13.43 | 218,363 | 185 |
| Gemini 4 Hydrogen Extra High | 6 | 57.8% | $5.86 | 164,610 | 88 |
| Gemini 4 Argon High | 16 | 52.9% | $4.57 | 109,500 | 118 |
Carbon High scored 0.4 percentage points higher than Fable 5.2 Max. Its listed cost per task was about 72% lower, token usage about 89% lower, and step count about 67% lower. Carbon Medium was 0.3 percentage points below High, with a cost $0.73 lower, 2,893 fewer tokens, and 4 fewer steps. Carbon Low scored below the other two tiers, but its listed cost was higher ($11.89). These results should not be interpreted as showing a stable, monotonic relationship between reasoning tier and cost.
Compared with other Gemini configurations, Carbon High scored 2.6 percentage points above Hydrogen Extra High, at an average cost $3.36 higher, with fewer tokens and steps. It scored 7.5 percentage points above Argon High, at a cost $4.65 higher. The leaderboard does not provide enough information to determine whether these differences are statistically significant.
Leaderboard snapshot: When accessed on 2026-11-16, the CursorBench 4.0 page listed 91 model configurations sorted by score. Rankings and costs may change as the page is updated.
Carbon tier data: High 60.4% / $9.22 / 39,960 tokens / 71 steps, Medium 60.1% / $8.49 / 37,067 tokens / 67 steps, Low 58.6% / $11.89 / 42,663 tokens / 124 steps.
Sample size: Neither the leaderboard nor the methodology notes show the number of tasks, samples per model, or repeated runs, so these are recorded as unspecified. Steps/Task on the leaderboard is an average step count, not a sample size.
Cost basis: Public token prices are used to include input, cache-read, cache-write, and output tokens. This is not independently measured actual billing. The page's change history says costs are updated when prices change.
Version basis: CursorBench 4.0 was introduced in the 2026-09-10 update history, which says new task types were added. The earlier methodology article explicitly cautions that different benchmark versions have different task distributions and should not be compared across versions.
The official leaderboard shows that all three Carbon tiers placed in the top four in the CursorBench 4.0 page data, with High receiving the highest score. The score difference between High and Medium is only 0.3 percentage points. Cursor has cautioned that small score differences may not be statistically meaningful, so the leaderboard alone cannot establish that High is consistently better than Medium.
The leaderboard compares costs, token usage, and steps by model configuration, but does not disclose task sample size, per-task results, or the full 4.0 scoring method. The results apply only to this version's internal Cursor tasks and evaluation environment. Model price updates may also change the cost column. Do not directly convert leaderboard scores into real-world project success rates or rankings across tools.
Visit the CursorBench page, confirm that the version is labeled 4.0, and record the access date. Transcribe the model configuration, score, cost per task, tokens/task, and steps/task from the table. Reproducing the ranking requires Cursor's unpublished task set, execution configuration, number of repeated runs, and scoring method. The within-version comparisons in this note use the page snapshot from 2026-11-16.
Gemini 4 Carbon