Hark says its Handoff agent can browse, buy and recruit for a fraction of frontier AI costs

AI startup Hark has unveiled Handoff, a computer-use agent designed to complete tasks in a web browser, from ordering food to booking flights and contacting job candidates. Carl Franzen reports for VentureBeat that public sign-ups have opened, with initial access planned later this month.

For every task, Handoff creates a separate virtual computer with a browser, file system and terminal. Users can connect existing accounts, allowing the agent to log in and use stored payment details, addresses and account history. Hark says the approach is needed because most websites do not offer public application programming interfaces, or APIs, that software agents can use directly.

Strong claims, limited comparisons

Hark says Handoff scored 97.7 on Online-Mind2Web, a benchmark for web agents evaluated by humans. The company compares that result with scores for older versions of models from OpenAI, Anthropic and Google. It also says its model costs $0.18 per million input tokens and $2.37 per million output tokens.

Those claims need context. Hark did not provide benchmark results against newer leading models, including OpenAI’s GPT-5.6 and Anthropic’s Opus 5. These systems do not currently have published Online-Mind2Web scores, so Hark’s claim of a record result cannot be independently compared with all current rivals.

Some tests also rely on Hark’s own evaluation setup and internal AI judge. On WebTailBench v2, one of the benchmarks Hark cites, GPT-5.5 outscored Handoff.

Further questions concern the model’s foundations and data security. Hark says it has so far focused on post-training rather than training a base model from scratch. The company has not identified the underlying model or detailed who can access files on users’ virtual computers. It describes Handoff as a technical preview and says it will share more security information before broader availability.