DesktopBench

Leaderboard

Jun 25, 2026 / 200 tasks / 3 models

Ranking agents on economically valuable professional software work, built from real workflows and scored with resettable environments, tasks, and programmatic verifiers.

Rank 1

Opus 4.8

Anthropic · Proprietary

Score: 62.0±1.0%
Success / Total: 124 / 200 avg
Runs: 2
Max Steps: 100

Run	Success / Total	StoreOps	AccountingOps	SupportOps	SalesOps	OfficeOps
Run 1	126 / 200	30 / 40	25 / 40	28 / 45	21 / 35	22 / 40
Run 2	122 / 200	29 / 40	24 / 40	27 / 45	20 / 35	22 / 40

Rank 2

OpenAI 5.4 mini

OpenAI · Proprietary

Score: 58.0±1.0%
Success / Total: 116 / 200 avg
Runs: 2
Max Steps: 100

Run	Success / Total	StoreOps	AccountingOps	SupportOps	SalesOps	OfficeOps
Run 1	118 / 200	28 / 40	23 / 40	26 / 45	19 / 35	22 / 40
Run 2	114 / 200	27 / 40	22 / 40	25 / 45	19 / 35	21 / 40

Rank 3

RLVR open-weight model

Open weight · RLVR trained

Score: 45.5±1.0%
Success / Total: 91 / 200 avg
Runs: 2
Max Steps: 100

Run	Success / Total	StoreOps	AccountingOps	SupportOps	SalesOps	OfficeOps
Run 1	93 / 200	23 / 40	18 / 40	20 / 45	15 / 35	17 / 40
Run 2	89 / 200	22 / 40	17 / 40	19 / 45	14 / 35	17 / 40

Rank Model Provider Runs Score

1 Opus 4.8 Proprietary Anthropic 2 62.0±1.0%

Success / Total: 124 / 200 avg Max Steps: 100

Run	Success / Total	StoreOps	AccountingOps	SupportOps	SalesOps	OfficeOps
Run 1	126 / 200	30 / 40	25 / 40	28 / 45	21 / 35	22 / 40
Run 2	122 / 200	29 / 40	24 / 40	27 / 45	20 / 35	22 / 40

2 OpenAI 5.4 mini Proprietary OpenAI 2 58.0±1.0%

Success / Total: 116 / 200 avg Max Steps: 100

Run	Success / Total	StoreOps	AccountingOps	SupportOps	SalesOps	OfficeOps
Run 1	118 / 200	28 / 40	23 / 40	26 / 45	19 / 35	22 / 40
Run 2	114 / 200	27 / 40	22 / 40	25 / 45	19 / 35	21 / 40

3 RLVR open-weight model RLVR trained Open weight 2 45.5±1.0%

Success / Total: 91 / 200 avg Max Steps: 100

Run	Success / Total	StoreOps	AccountingOps	SupportOps	SalesOps	OfficeOps
Run 1	93 / 200	23 / 40	18 / 40	20 / 45	15 / 35	17 / 40
Run 2	89 / 200	22 / 40	17 / 40	19 / 45	14 / 35	17 / 40