Hark Handoff: The Browser Agent That Topped the OM2W Benchmark at a Tenth of the Price

#hark #handoff #browseragent #webnavigatio

TL;DR

Brett Adcock's newest venture Hark (after Figure AI and Archer Aviation) put its browser agent Handoff into research preview on 5 August 2026. Handoff posted a 97.7% score on Online-Mind2Web, ahead of OpenAI's GPT-5.4 (92.8) and Anthropic's Claude Opus 4.8 (84.1), and bills at $0.18/$2.37 per million input/output tokens — less than a tenth of GPT-5.5's $5/$30.

The headline is the benchmark score, but the durable signal is economic: what matters is no longer whether an agent can click through a checkout, but whether running thousands around the clock costs less than the labor they replace.

Why browser agents matter now

Web navigation was long the unglamorous corner of the agent race: models reasoned, coded, and planned long before they could reliably fill out a real form. Through 2026 the field converged on a specific test — can an agent operate a website the way a human does, with no custom API and no hand-crafted script? (Source : Business Wire — Hark Announces Handoff, the World's Best Web Browsing AI Agent)

If agents can do this reliably, the addressable market is the clerical economy — which is why a $6 billion valuation on a research-preview product is not obviously irrational, and why price per token matters more than the leaderboard.

The benchmark: a record with a caveat

Hark reports 97.7% on Online-Mind2Web (OM2W), the most-cited benchmark for measuring whether an agent can operate a site like a human. By Hark's own comparison, that beats OpenAI's GPT-5.4/ChatGPT 5.4 at 92.8, Anthropic's Claude Opus 4.8 at 84.1, and Google's Gemini 2.5 Pro at 69. (Source : Startup Fortune — Hark's New AI Agent Claims to Beat Anthropic and OpenAI at Browsing the Web)

Two caveats matter more than the raw number. First, the figures are self-reported by the vendor topping the list, and the newest frontier models have not yet been benchmarked against it. Second, the gap to the nearest rival is thin: Yutori, founded by ex-Meta researchers, hit 97.3% in late June 2026 — a 0.4-point delta inside any benchmark margin of error. (Source : Enterprise DNA / TechCrunch — Hark Handoff Browser Agent, $700M Series A, Enterprise 2026)

The honest read: Handoff sits in a tight cluster at the top — the first browser agent to clear 97% — but the score reads as "competitive leader," not "decisive breakthrough."

The architecture: predicting actions, not tokens

Handoff does not call APIs. It runs on what Hark calls a "virtual computer" inside its own infrastructure: the model sees the screen, reads layout and pixels, then clicks, fills forms, and scrolls. Hark's framing is that the model does not predict the next word — it predicts the next action, a click or keystroke at a precise point on screen. (Source : Business Wire — Hark Announces Handoff, the World's Best Web Browsing AI Agent)

The distinction is more than marketing. An API-driven agent fails the moment it leaves the paved path of structured endpoints; a pixel-and-layout agent treats the whole web as one interface, so Handoff can be pointed at DoorDash, United, LinkedIn, and a tax form without per-site engineering. (Source : AI Weekly — Hark Previews Handoff, Its Browser Agent for Real Websites)

The trade-off is clear: reading raw pixels is costlier and more fragile than parsing JSON. Handoff's pricing bets its infrastructure keeps that cost low enough to win anyway.

The real story: the 10x price gap

Handoff charges $0.18 per million input tokens and $2.37 per million output tokens; GPT-5.5 costs $5 and $30 for the same volumes — more than 27x on input and 12x on output. (Source : Startup Fortune — Hark's New AI Agent Claims to Beat Anthropic and OpenAI at Browsing the Web)

This is where the economics of always-on agents get decided. A browser agent booking flights and filing taxes burns tens of thousands of tokens per task. At frontier prices a single task can cost dollars, collapsing the margin against a minimum-wage human; at a tenth of the price it drops into tens of cents — the point where round-the-clock agents become a line item, not a lab experiment.

The funding supports the thesis: a $700 million Series A at a $6 billion valuation in May 2026 signals investors are pricing browser agents as a platform, not a feature. (Source : Enterprise DNA / TechCrunch — Hark Handoff Browser Agent, $700M Series A, Enterprise 2026)

What this changes for the enterprise

Hark's running test cases point to where the money is: ordering meals end-to-end on DoorDash and Uber Eats, comparing prices and booking flights across United, Delta, and American, sourcing candidates on LinkedIn, and filing taxes. Commercial launch is planned before the end of summer 2026, with the current phase a waitlist beta. (Source : AI Weekly — Hark Previews Handoff, Its Browser Agent for Real Websites)

For enterprises, the question is not which agent scores highest this quarter — that ranking will churn — but whether to build against the API layer or the screen layer. Handoff argues for the screen layer, because it removes the integration cost that has quietly been the largest hidden line item in agent deployment. If the 97% claim holds independently and the price survives launch, running a persistent web agent gets an order of magnitude cheaper in a single release.

FAQ

What is Hark Handoff?
A browser agent from Hark, founded by Brett Adcock (Figure AI, Archer Aviation). It operates real websites by reading the screen and clicking, typing, and scrolling — no APIs. Research preview began 5 August 2026. (Source : Business Wire — Hark Announces Handoff, the World's Best Web Browsing AI Agent)

What is Online-Mind2Web (OM2W)?
The benchmark for whether an agent can use a website as a human does, across real multi-step tasks.

How does Handoff's 97.7% compare?
It leads OpenAI's GPT-5.4 (92.8), Anthropic's Claude Opus 4.8 (84.1), and Google's Gemini 2.5 Pro (69). Yutori reached 97.3% in late June 2026, a 0.4-point gap within benchmark noise. (Source : Startup Fortune — Hark's New AI Agent Claims to Beat Anthropic and OpenAI at Browsing the Web)

What does Handoff cost?
$0.18 per million input tokens and $2.37 per million output tokens, versus $5 and $30 for GPT-5.5 — under a tenth of the cost. (Source : Startup Fortune — Hark's New AI Agent Claims to Beat Anthropic and OpenAI at Browsing the Web)

When is Handoff available?
Currently a waitlist beta, with commercial launch planned before the end of summer 2026. (Source : AI Weekly — Hark Previews Handoff, Its Browser Agent for Real Websites)

DEV Community