Hark Opens Signups for a Cut-Rate Browser Agent, Handoff

Hark, the startup led by Figure AI founder Brett Adcock, opened public signups Thursday for Handoff, a computer-use agent that drives a virtual browser, file system and terminal to book flights, message job candidates and handle other web tasks without a human in the loop. The company says the tool beats frontier-model pricing roughly tenfold while topping a widely cited web-agent benchmark. Both claims deserve scrutiny before they shape any enterprise buying decision.

Hark’s headline figure comes from Online-Mind2Web, a benchmark where human judges score whether an agent actually finished a browsing task rather than merely attempted it. Handoff posted 97.7 there, a number the company sets against three rivals it tested itself: 92.8 for GPT-5.4, 84.1 for Opus 4.8, and 69 for Gemini 2.5 Pro, per Hark’s own comparison.

Handoff’s price undercuts those same rivals by a wide margin. Hark bills 18 cents for every million input tokens Handoff processes. A million output tokens costs $2.37, well below the roughly $5 and $30 that GPT-5.5 charges for the same categories, putting Handoff under a tenth of that rate.

Every comparison Hark supplied used last generation’s models. GPT-5.6 and Claude Opus 5, the labs’ current flagship systems, have not posted results to Online-Mind2Web, so the “top-ever” claim cannot be checked against today’s actual competition. On OSWorld 2.0, a related computer-control benchmark, Opus 5 already scores roughly 70.6 percent against 55.7 percent for the older Opus 4.8 model, Hark’s own point of reference.

Two of the three benchmarks behind Handoff’s headline score ran inside Hark’s own testing harness, scored by Hark’s internal software rather than an outside judge. On the third, WebTailBench v2, GPT-5.5 actually outscored Handoff, 72.3 to 68.6.

Hark has not named Handoff’s base model or the data mix used to fine-tune it. The company told VentureBeat that Handoff reflects fine-tuning and reinforcement learning layered onto an existing base model, and that training its own foundation model from scratch will not start until later this year. Asked who could reach the virtual machines and files an agent generates, a spokesperson said security is “a primary focus” but that the product remains a technical preview, with more detail coming at summer’s end.

The bigger open question is not benchmark integrity but custody. Handoff logs into a user’s accounts and acts using saved addresses, payment methods and order history, so a bad instruction or a hijacked session could place an order or move money before a person reviews it. Neither the announcement nor Hark’s research preview says who is liable when that happens, a gap that matters more once an agent handles real money than when it clears a benchmark.

Whether the price gap survives contact with Handoff’s real serving costs is a separate question. Anthropic has not raised Opus 5’s price above what Opus 4.8 charged, so Handoff’s tenfold discount holds against Anthropic’s current flagship, not just the older model Hark chose to cite, provided the performance claims hold too. Parkway Venture Capital led the $700 million round Hark closed this spring, a Series A valuing the company at $6 billion, with Nvidia, AMD, Qualcomm Ventures, Intel Capital, Salesforce Ventures and ARK Invest also writing checks, plus another $100 million from Adcock himself. That war chest buys room to price below cost while Hark builds a user base, and the discount is either a real cost advantage or a subsidy Hark has not disclosed.

Adcock’s public claims have drawn scrutiny before. Fortune’s Jason Del Rey reported in April 2025 that Figure’s BMW deployment, which Adcock described as a robot fleet running end-to-end operations, amounted to a single unit practicing part-picking during off hours; BMW has since said the partnership grew to support more than 30,000 vehicles over ten months. Adcock also failed to appear for a live demo he had committed to at an industry conference two months later, per TechCrunch, and avoided the BMW questions when he did take the stage.

None of that proves Handoff underperforms. It does mean enterprise buyers should press Hark for results against GPT-5.6 and Opus 5 before signing anything, and should get contractual clarity on who bears liability for agent actions taken with stored payment credentials before connecting a live account.

Reported by VentureBeat on August 6, 2026.