Hark Handoff Outperforms GPT-5.5 at 90% Lower Cost
Once_ina_Lifetime · reddit · 2026-08-10
A Reddit user sparked discussion around Hark Handoff, a Computer Use Agent (CUA). According to shared metrics, the model outperforms GPT-5.5 and Opus 4.8 across various web benchmarks (like Online Mine 2Web and WebTail Bench) while costing only $2.37 per million tokens—a 90% cost reduction.
Regarding its training methodology, the model uses an undisclosed base combined with Supervised Fine-Tuning (SFT) and asynchronous reinforcement learning, specifically leveraging the GRPO algorithm (similar to DeepMind's AlphaGo approach) to accelerate time to market. Despite advertised low latency, the poster notes that page rendering and state resolution still add roughly 10 seconds of overhead, expressing skepticism about real-world latency but remaining bullish on the CUA sector overall.
More from coding & agent
- Epic Fail: Claude Code Review Agent Unexpectedly Switches Git Branches — 4310sy · 2026-08-10
- Pitfall: Don't Use Legacy Code as the Spec for Claude Refactoring — arekusandr_ · 2026-08-10
- GHOST: An Open-Source Self-Hosted OSINT Investigation CRM — tom_doerr · 2026-08-10
- OpenCodex: Run Any LLM with Claude Code and OpenAI Codex — vista8 · 2026-08-10
- Deep Dive: Context Management in Cline vs. Kilo vs. Qwen Code for Long Tasks — Ok-Shower7286 · 2026-08-10
- Open Source ElatoAI Runs Realtime Voice AI on ESP32 with 100+ Models — tom_doerr · 2026-08-10