Brett Adcock's fourth startup, Hark, has emerged from stealth with its first product: Handoff, a computer-use agent (CUA) designed to autonomously navigate the open web on behalf of users. Think ordering dinner on DoorDash, booking flights on United or Delta, or messaging job candidates on LinkedIn — end-to-end, without human hand-holding. Sign-ups opened publicly today at hark.com, with availability planned for later this month.

The Core Pitch: Performance and Price

Handoff operates by spinning up a dedicated virtual computer for each request, complete with its own browser, file system, and terminal. Users can connect existing accounts so the agent can log in and act using saved addresses, payment methods, and browsing history.

Hark's rationale for the product is grounded in a pointed observation: despite users spending 75% of their screen time in a browser, fewer than 1 in 1,000 websites have publicly accessible APIs. That gap is exactly where a browser-native agent can wedge in.

On pricing, the case is genuinely compelling:

  • $0.18 per million input tokens and $2.37 per million output tokens for Handoff
  • Versus $5 and $30 for OpenAI's GPT-5.5
  • That's roughly a tenfold cost advantage — and it holds even against Anthropic's newer Opus 5, which carries the same $5/$25 list price as its predecessor

Per-turn model latency is claimed at 0.8 seconds, versus 6–6.8 seconds cited for competing models.

The Benchmark Numbers — and Their Asterisks

Hark claims Handoff posted the top-ever score on Online-Mind2Web (OM2W), a third-party human-evaluated leaderboard for web agents, with a score of 97.7 versus:

  • 92.8 for OpenAI GPT-5.4
  • 84.1 for Anthropic Claude Opus 4.8
  • 69 for Google Gemini 2.5 Pro

But here's the catch: those are last generation's models. GPT-5.6, Anthropic's Opus 5, DeepSeek V4, Kimi K3, and Qwen3.8-Max are all absent from the comparisons — because none have published OM2W results. There's no way to independently verify whether Handoff leads the current field.

The omission is significant. On OSWorld 2.0, a related full-computer-control benchmark, Anthropic's Opus 5 scores 70.6% versus the 55.7% posted by Opus 4.8 — the version Hark chose as its comparison point. Frontier models have made their biggest recent gains precisely in computer use.

The latency comparison carries similar caveats: Hark measured competing models in its own harness, with reasoning set to the highest — and slowest — level. No independent measurements exist. And on WebTailBench v2, one of three benchmarks in Hark's own results table, GPT-5.5 scores 72.3 to Handoff's 68.6 — meaning Hark doesn't lead across the board even in its own chosen suite.

Asked whether Hark plans to publish comparisons against newer models, the company didn't specify.

What We Don't Know Yet

Hark's training pipeline involves supervised fine-tuning followed by asynchronous reinforcement learning using the GRPO algorithm. But the company acknowledges it has only done post-training so far — pre-training is planned for later this year. That means Handoff is built on an undisclosed base model Hark didn't train.

For enterprise buyers, the more pressing unknown is security. The dedicated virtual computers Handoff uses can generate files — and it's unclear who can access them. A Hark spokesperson said security is "a primary focus" but that more details will come when the product hits general availability "at the end of the summer."

Adcock's Track Record

Adcock previously co-founded Vettery (sold for ~$100M in 2018), Archer Aviation, and humanoid robotics unicorn Figure AI — where he remains CEO simultaneously. Hark raised a $700 million Series A in May 2026 at a $6 billion valuation, led by Parkway Venture Capital with participation from Nvidia, AMD, Intel Capital, Qualcomm Ventures, Salesforce Ventures, and ARK Invest. Adcock seeded the company himself with $100 million.

His promotional style has drawn scrutiny before. A 2025 Fortune report found that Figure's BMW partnership was far more modest than Adcock's "fleet" framing suggested — though the relationship has since grown substantially, with BMW crediting the Figure 02 robot with supporting production of more than 30,000 BMW X3 vehicles over a 10-month period.

None of that history means Handoff's numbers are fabricated. The pricing advantage is real and verifiable. The benchmark edge — against current-generation models — remains unproven.