Hark Handoff AI Agent Claims Leading Web Performance at One-Tenth the Cost

Summary

Hark, an AI startup founded in 2026 by entrepreneur and roboticist Brett Adcock, announced its first product, Handoff, on August 5. The company describes Handoff as a computer-use agent that can independently complete browser-based tasks such as ordering food through DoorDash, booking United or Delta flights, and contacting job candidates on LinkedIn. Public sign-ups opened at hark.com on the announcement date, with initial availability expected later in August as part of Hark’s broader software platform. Hark argues that browser automation addresses a major gap for AI agents because people spend an estimated 75% of their screen time in browsers, while fewer than one in 1,000 websites offer publicly accessible APIs.

According to Hark, Handoff achieved 97.7 on the human-evaluated Online-Mind2Web benchmark, compared with 92.8 for OpenAI’s GPT-5.4, 84.1 for Anthropic’s Claude Opus 4.8, and 69 for Google’s Gemini 2.5 Pro. Hark also quotes prices of $0.18 per million input tokens and $2.37 per million output tokens, alongside per-turn latency of 0.8 seconds. For comparison, the company cites GPT-5.5 pricing of $5 and $30 per million input and output tokens. Each task runs in a dedicated virtual computer containing a browser, file system, and terminal. Users can connect accounts, allowing the agent to access saved addresses, payment methods, and account history.

Those performance claims require qualification. VentureBeat notes that the Online-Mind2Web comparison excludes newer systems including OpenAI’s GPT-5.6, Anthropic’s Opus 5, DeepSeek V4, Kimi K3, and Qwen3.8-Max because no results for them were available on the public leaderboard. Newer models have made substantial computer-use gains elsewhere: Opus 5 reportedly scores about 70.6% on OSWorld 2.0, versus 55.7% for Opus 4.8. Handoff also did not lead every disclosed test, scoring 68.6 on WebTailBench v2 against GPT-5.5’s 72.3. Two of Hark’s three evaluations were conducted in its own harness and graded by an internal LLM judge, while the latency comparisons used competitors at their slowest, highest-reasoning settings. No independent latency results were provided.

Hark says Handoff underwent supervised fine-tuning followed by asynchronous reinforcement learning using GRPO, but the company has only conducted post-training and plans to begin pre-training later in 2026. It has not identified the underlying base model or detailed the mix of proprietary and open training data. Security is another unresolved issue because the virtual computers may hold files, credentials, payment information, and browsing history. A spokesperson said privacy and security are priorities but characterized Handoff as a technical preview, with further information promised when the product reaches the market around the end of summer.

Hark raised a $700 million Series A in May 2026 at a $6 billion valuation after Adcock contributed $100 million. Investors include Parkway Venture Capital, Nvidia, AMD, Intel Capital, Qualcomm Ventures, Salesforce Ventures, and ARK Invest. Adcock remains CEO of both Hark and humanoid-robotics company Figure AI; Hark’s models are reportedly being trained on Figure robots, although the companies are not expected to merge. His history of ambitious promotion at Figure gives skeptics additional reason to seek independent verification, even though BMW later reported that Figure 02 supported production associated with more than 30,000 BMW X3 vehicles over 10 months. The next tests for Handoff will be real-world availability, independent performance measurements, comparisons with current frontier models, and detailed enterprise security disclosures. Its claimed cost advantage appears significant, but whether its quality, speed, and safeguards hold at scale remains uncertain.