How long one DeepSWE task takes
Shorter is better. Pass@1 and repriced task cost remain beside the measured mean wall time.
| Model | Effort | Mean duration | Pass at one | Cost per task | Mean output tokens |
|---|
The data behind the signal.
Data archive
Artificial Analysis score explorers35 Agentic and Intelligence Index leaders against cost and time.
Time: the missing DeepSWE metricMeasured wall-clock time beside pass@1 and cost per task.
DeepSWE time/price Pareto explorerweighted x-axis for a cost vs time sensitive model choice
CursorBench 3.2 Pareto explorer60 coding-agent configurations against cost, tokens, and steps.
The evidence, with the missing dimensions.
TrigrD Data turns public datasets into readable comparisons. We keep source values, derived values, and opinion separate so people and agents can inspect the same evidence. Write to us at contact@trigrd.tech.
DeepSWE v1.1
Time: the missing DeepSWE metric
The public artifact records end-to-end duration for every scored attempt. This page covers 15 selected model families across labs, score bands, and price bands.
Score against time
Pass@1 starts at zero so the score differences stay in proportion. Use the configuration picker to compare individual model and effort combinations.
DeepSWE v1.1
DeepSWE time/price Pareto explorer
Keep score on the y-axis. Compare it with cost, time, or a weighted x-axis for a cost vs time sensitive model choice.
Choose the trade-off that matters to you
The Pareto line connects configurations that no lower-burden point can match on score.
Cost per task · lower is better
Artificial Analysis · 2026-09-04 snapshot
The score leaders, with their burden.
Compare 35 Artificial Analysis score leaders on two indices. Each model/effort configuration is its own datapoint, with cost, time, and a weighted choice of both on the x-axis.
AA Intelligence Index against cost and time
Higher Intelligence Index is better. The same selector exposes current cost, time per task, and a normalized weighted burden for the selected leaders.
Current cost per task · lower is better
| Rank | Model | Effort | Intelligence Index | Cost per task | Time per task |
|---|
AA Agentic Index against cost and time
Higher Agentic Index is better. Move left to favor the lower task burden, or use the weighted view to set your own cost/time preference.
Current cost per task · lower is better
| Rank | Model | Effort | Agentic Index | Cost per task | Time per task |
|---|
CursorBench 3.2 · 2026-09-02 snapshot
Coding-agent scores, with the work attached.
Compare all 60 current CursorBench configurations. Keep score on the y-axis and test cost, tokens, steps, or your own cost/steps weighting on the x-axis.
CursorBench 3.2 score against task burden
Higher score is better. The Pareto line connects configurations that no lower-burden point can match on score.
Average cost per task · lower is better
| Rank | Model | Effort | Score | Average cost per task | Tokens | Steps |
|---|