Hark launches Handoff²: 2x faster, 73%→82% success, near Opus 5.5 accuracy
adcock_brett · x · 2026-11-13
Hark founder Brett Adcock announces Handoff², the next-generation computer-use model, available to all users on all plans.
Key numbers
- The model itself is 46% faster; with backend optimizations, average speed is 2x the original — up to 3.8x faster on flight booking and 5x+ on long research tasks
- Internal browsing benchmark (478 tasks across 238 websites): success rate up from 73% to 82%, with 46% less time per task
- Near-frontier accuracy (within 5 points of Opus 5.5 max) at 2x the speed; +80% better at solving CAPTCHAs
- Since launching in October, Handoff has logged 5.6M+ browsing sessions and handled over $20M in purchases — groceries, flights, dentist appointments, refunds, job applications and more
Per the official post, the efficiency gains come from co-evolving the harness with reinforcement learning.
Related event: Figure Releases Handoff²: Faster Computer Use with Near-Frontier Accuracy(2 posts)→
More from coding & agent
- Edge0 launches open-weight device-native 100B models, claims 86.5% on Terminal-Bench 2.1 — kimmonismus · 2026-11-13
- Google Cloud donates Agent Substrate, gVisor, OpenRL to CNCF as Agent Substrate hits 1.0 GA — rakyll · 2026-11-13
- SGLang Summit: agents now write ~95% of merged PRs as v1.0 and Miles v0.2 ship — hsu_byron · 2026-11-13
- Real-time chess tracking from iPhone video with RF-DETR and Stockfish — measure_plan · 2026-11-13
- OmniMind: an open-source MCP server for handing off tasks between AI agents — IrfanCodesBTW · 2026-11-13
- Claude Design Crit in Canvas: Screenshots Every Component State, Fixes All Comments in Parallel — max__drake · 2026-11-13