Hark Handoff: The $6 Billion Bet on an AI That Can Actually Use Your Browser

Hark, a startup that raised $700 million in Series A funding in May at a staggering $6 billion valuation, today launched its agent Hark Handoff, which can use a browser efficiently to complete tasks.

The company claims that Handoff can easily navigate websites that have no official APIs, including Target, Walmart, OpenTable, and LinkedIn. It said that the Handoff agent looks at website structure and visual data to understand if it needs to click buttons or type information. The premise is similar to tons of browser-based agents released before. Issue a command, and it will complete tasks for you, including ordering food or coffee, booking travel tickets, filing returns, shopping for essentials, reserving a table at a restaurant, or researching on your behalf by looking at various sources.

In a video demo, the company's CEO, Brett Adcock, showed that the assistant can take a command to build a bouquet with flowers users specify, along with accommodating fuzzy terms like "some of the florist's choice." Notably, the video shows only part of the process, so we can't really gauge its effectiveness.

How It Works: A Virtual Computer for Every Task

For each request, Hark Handoff creates a dedicated virtual computer equipped with a browser, file system, and terminal. It clicks, scrolls and types through websites instead of waiting for each service to provide an API. This approach allows it to work across websites that lack consumer-facing APIs, a critical feature given that Hark's analysis of nearly 3 million minutes of screen activity found that 74.9% of people's time was spent in browsers, while fewer than one in 1,000 websites offered a public API.

Users can connect existing accounts, giving Handoff access to saved addresses, preferences, and purchase histories. That design makes reliability and account security central product requirements. A browser agent that can merely find a flight is materially different from one trusted to log in, select an itinerary, and make a purchase.

The Technical Bet: Action Prediction Over Token Prediction

Hark's architectural bet is what separates the preview from the crowd. The company says Handoff predicts the next action: a click at a coordinate, a keystroke into a field—rather than the next token, the standard objective for large language models. It is running on a post-trained model for this release and plans to move to pre-training later this year, which Hark argues will let it iterate faster on its data pipeline and training infrastructure than teams grafting agents onto general-purpose LLMs.

The company said that for this release, it's using a post-trained model, and plans to pre-train later this year. Hark said that, with this approach, it can refine its data pipeline, training infrastructure, and techniques more quickly.

Benchmark Claims: Tops the Leaderboard, With Caveats

Handoff scored 97.7% on the human-evaluated Online-Mind2Web benchmark, according to the official leaderboard. The test covers 300 tasks across 136 live websites. Handoff recorded 100% on tasks classified as easy, 95% on medium tasks, and 100% on hard tasks in an evaluation dated August 4th. That result narrowly passed Yutori's Navigator n1.5, which scored 97.3% in June, and an ACT-2 system powered by GPT-5.4, which scored 92.7%. Careerflow.ai conducted the human evaluations for all three entries.

Hark also published results for WebTailBench v2 and an internal evaluation. Handoff scored 68.6 on WebTailBench, behind GPT-5.5 at 72.3 and ahead of Claude Opus 4.8 at 66.9. On Hark's internal test, Handoff led with 83.2, compared with 80.5 for Opus 4.8 and 75.2 for GPT-5.5.

The leaderboard supports Hark's claim that Handoff currently holds the top verified Online-Mind2Web score. It does not establish Handoff as the best internet-use model across every test or production setting. Hark's broader "best ever" description stretches beyond the scope of the independent result. Hark ran those comparisons through its own evaluation harness and used an internal large-language-model judge, leaving Online-Mind2Web as the independently evaluated portion of the release.

Pricing: A Fraction of the Competition

Hark's pricing advantage is far clearer. The company charges USD 0.18 per million input tokens and USD 2.37 per million output tokens, compared to USD 5 and USD 30 for GPT 5.5. Handoff costs less than a tenth of what frontier competitors charge, and the company promises a latency of 0.8 seconds per turn, significantly faster than the 6-6.8 seconds Hark measured for competitors.

Hark claims Handoff operates faster and at a significantly lower cost than competing models such as GPT 5.5 and Opus 4.8.

The Founder: Brett Adcock's Unconventional Path

Adcock, who grew up on a farm in central Illinois, previously built recruiting marketplace Vettery and co-founded electric-aircraft developer Archer Aviation before moving into humanoid robotics at Figure AI. Hark extends the same full-stack approach into consumer AI: models, software and eventually purpose-built devices.

Adcock launched Hark publicly in March with $100 million of his own capital. The Series A round in May was led by Parkway Venture Capital, with participation from Nvidia, AMD, Intel, Qualcomm, Salesforce Ventures, and ARK Invest, putting virtually every major AI-chip maker on a single cap table.

Adcock's experience makes recruiting a natural proving ground for Handoff. It is a workflow Adcock knows, full of repetitive browsing and messaging as well as sensitive decisions that can expose the limits of autonomous software.

The Competitive Landscape

There are many companies working on computer-use agents, including Google, OpenAI, and Anthropic, with VC-funded startups working on browser-based task automation, such as Browser Use, Polar, Strawberry, and Aside. OpenAI, Google, and Anthropic all ship computer-use agents inside their consumer and API products, and a wave of venture-backed startups is chasing the same browser-automation wedge.

Most of these agents share the same failure modes: brittle when sites re-render, slow on multi-step flows, and expensive per task when routed through frontier models. Hark's claim to beat GPT 5.5 and Opus 4.8 on cost is a direct swipe at that pricing dynamic.

A purpose-built action model, if it works, is a defensible position. GPT 5.5 and Opus 4.8 were trained to write tokens, not to click through Target's checkout flow, and the gap in cost per successful task is where a specialist can win.

What's Next: Waitlist and Launch

Hark has opened a waitlist for its platform and plans to release it by the end of the summer. Hark introduced Handoff as a research preview on August 5, 2026. Access is currently application-based through Hark's beta sign-up; Hark says its broader platform will be available by the end of summer.

Hark says the company is entering beta and reviewing applications, while VentureBeat reported that access to the software platform was planned for later in August.