▣ RealtimeCUA Bench

REALTIME GUI · FIXED COHORT / 13F1B74

Every score has a replay.

Compare GPT-6.1 Sol and Claude Haiku 5.5, then watch what happened in each task.

17task variants
34verified replays
2models
1×elapsed-time playback

SAMPLED PERFORMANCE

Leaderboard

Download data ↓

One valid run per model per task. Scores are task points, not success rates. The aggregate is the equally weighted mean over these 17 task variants.

Scoring & comparison notes

    Source frame and video manifest ↓ · Original replay collection ↗

    INSPECT THE EPISODES

    Task replays

    GPT 2829 / HAIKU 2844

    TaskGPTHaiku

    Select a score to watch that model’s run.

    Native video learning · six tasks, tutorials and model videos / 原生应用视频学习 ↗

    OTHER TASKS & BUDGETS

    More task videos

    Existing Blender and TicketVault pilots use different budgets and protocols. Watch them here; their scores are separate from the leaderboard above.