Bỏ qua đến nội dung chính
Back to home
AI tools-ai 3 min read

Hark Launches Handoff: An Ultra-Low-Cost Browser-Automating AI Agent

Hark has announced Handoff, an ultra-low-cost AI agent designed to automatically control browsers, though its performance benchmarks have raised some skepticism.

Tier 2 · sources 99% confidence Reviewed
Sources venturebeat.com

Artificial intelligence startup Hark, founded by entrepreneur Brett Adcock, has officially launched its first product named Handoff. It is a 'Computer Use Agent' (CUA) capable of autonomously navigating the web to perform tasks such as ordering food, booking flights, or sending recruitment messages. According to Hark's announcement on August 5, 2026, the tool is now open for public beta registration and is expected to officially launch later this month.

Key Developments

Hark claims Handoff achieved a record score of 97.7 on the Online-Mind2Web (OM2W) benchmark, surpassing major models like OpenAI's GPT-5.4 (92.8 points) and Anthropic's Claude Opus 4.8 (84.1 points). What is particularly drawing attention is the model's ultra-low operating cost. Hark states they can offer the service at less than one-tenth the cost of competitors—specifically $0.18 per million input tokens and $2.37 per million output tokens—with a latency of just 0.8 seconds per response.

To execute requests, Handoff spins up a separate virtual machine featuring its own browser, file system, and command line. Users can connect their existing accounts to allow the agent to log in automatically and complete transactions using saved addresses or payment methods. The startup points out that while humans spend up to 75% of their device time on browsers, fewer than 1 in 1,000 websites have public APIs, creating a massive barrier that only this type of AI agent can effectively bridge.

Technical Analysis & Technology

On the technical side, Hark utilizes a supervised fine-tuning workflow combined with asynchronous reinforcement learning via the GRPO algorithm. However, the company admits it has only completed the post-training phase, while pre-training is slated for later this year. This means Handoff is currently running on a foundational model developed by a third party, whose identity Hark is keeping confidential.

Although the benchmark figures are impressive, analysts have pointed out several discrepancies in Hark's data. The compared competitors are older generations like GPT-5.5 or Claude Opus 4.8, ignoring the latest updates like GPT-5.6 or Opus 5. Notably, in the WebTailBench v2 test, GPT-5.5 actually outperformed Handoff with a score of 72.3 compared to 68.6.

Expert Opinions & Assessments

Founder Brett Adcock is a familiar name, having co-founded Vettery, the flying taxi company Archer Aviation, and the humanoid robotics startup Figure AI. However, Adcock's historically exaggerated promotional style has drawn skepticism. For instance, he previously claimed Figure AI had a 'fleet of robots' operating at a BMW plant, but a representative for the automaker later clarified it was only a single robot undergoing testing outside of standard production hours. Nevertheless, the partnership has since made more tangible progress.

For enterprise customers, data security within Handoff's virtual machines remains a major question mark. When asked about access to sensitive data, a Hark representative offered a vague response, stating that 'security and privacy are top priorities, but this is currently just a technical preview,' promising to share more details once the product is commercialized.

Impact & Future Outlook

The arrival of Handoff marks a distinct shift in AI models from pure chatbots to agents capable of independent action. If Hark can prove its tenfold price reduction is viable at scale, it will exert massive competitive pressure on the world's leading AI research labs.

For the Vietnamese tech community, the rise of computer-use agents like Handoff opens up numerous opportunities to automate repetitive browser-based workflows without building complex APIs. However, users should exercise extreme caution when granting personal account and payment credential access to these early-stage AI systems until more objective, independent security evaluations are available.