Hark Handoff: The $700M Browser Agent That Predicts Clicks, Not Words

What tasks can Handoff automate?

According to Hark, Handoff handles a wide range of multi-step online workflows :

• Ordering food (e.g., DoorDash)
• Booking travel (e.g., United, Delta flights)
• Shopping (e.g., Target, Walmart)
• Restaurant reservations (e.g., OpenTable)
• Messaging and recruiting (e.g., LinkedIn)
• Research, filing returns, and other form-heavy tasks

Architecture: predicting the next action, not the next token

Hark deliberately pivoted away from the standard LLM paradigm. Instead of predicting the next word in a text sequence, Handoff's model is architected to predict the next graphical user interface action — a click, a scroll, or a keyboard input at a specific screen location . The current release uses a post-trained model (fine-tuned on an existing base), with plans to transition to pre-training from scratch later in 2026 . Hark says this approach allows it to iterate more quickly on data pipelines and training methods before committing to a full pre-training run .

Benchmark performance: 97.7 on OM2W

Hark claims Handoff achieved the top score to date on the Online-Mind2Web (OM2W) benchmark, a human-evaluated industry standard for web-browsing AI performance . The reported scores are:

• Hark Handoff: 97.7
• GPT-5.5: 92.8
• Opus 4.8: 84.1
• Gemini 2.5 Pro: 69.0

Pricing is listed at $0.18 per million input tokens and $2.37 per million output tokens — less than one-tenth the cost of GPT-5.5 ($5 input / $30 output) . Hark claims Handoff is both faster and dramatically cheaper than frontier models from OpenAI, Anthropic, and Google for web-use tasks . These are company claims on a third-party benchmark; independent verification of real-world task completion rates and cost comparisons is not yet available.

Current status

As of August 2026, Handoff is in a closed research preview phase. Users can sign up for a waitlist at Hark's website. A full public release is expected by end of summer 2026 . The tech stack is still evolving: the model is post-trained for now, with pre-training planned before year-end .