Browser Agents 2026: A $300mn Race With No Clear Winner
Browserbase has $67.5mn raised at a $300mn valuation. Browser Use has 79K GitHub stars and Anthropic as a customer. Skyvern's $2.7mn seed beat them both on WebVoyager. The gap between funding, mindshare, and benchmark performance is wider than the press suggests.
The four browser-agent startups with disclosed funding — Browserbase, Browser Use, Firecrawl, and the now-defunct Reworkd — raised roughly $74mn between them across 2024 and 2025. The benchmarks to compare them now exist. The leaders by funding, by mindshare, and by benchmark are three different companies, and the press has tended to lump them.
The funding leader: Browserbase
Browserbase raised a $40mn Series B in June 2025 at a $300mn valuation, bringing total funding to $67.5mn. The company processed 50M sessions in 2025 across 1,000+ customers. The product surface includes a managed cloud Chromium fleet, the Stagehand open-source SDK (10K+ stars), and Director — a no-code automation product launched alongside the Series B.
Browserbase’s bet is on infrastructure: the LLM does not run on Browserbase, the LLM customer’s code does. Browserbase rents the browser. That is a clean-edge SaaS sell, with cost-per-session economics that resemble traditional IaaS. It also positions Browserbase to be the substrate underneath whichever LLM client wins the agent layer — Operator, Mariner, or something else.
The risk: if the foundation labs ship managed browsers themselves (and OpenAI’s Operator, already does), the substrate gets disintermediated.
The mindshare leader: Browser Use
Browser Use raised a $17mn seed in March 2025 led by Felicis, joining YC W25. The GitHub repo has 79,000+ stars. Customers include Airbnb, Amazon, and Anthropic.
The product is different from Browserbase: Browser Use is an open-source Python framework for letting an LLM control a browser, with a hosted runtime as the commercial product on top. The framing is “AI agent SDK,” not “browser infrastructure.” Developers integrate the SDK into their own backend; the browser execution happens locally or in a hosted Browser Use runtime.
Browser Use’s bet is that the LLM-controlled browser is itself a primitive worth owning, not just the underlying Chromium fleet. The 79K stars is evidence the bet is working at the developer level. Whether it converts to enterprise revenue against Browserbase’s better-funded sales motion is the next twelve months’ question.
The benchmark leader: Skyvern
Skyvern raised a $2.7mn seed — a fraction of the funding the others took. Skyvern v2.0 hit 85.85% on WebVoyager, at or above the larger competitors and within striking distance of OpenAI’s CUA (87%).
The benchmark gap is small enough that it should be read with caution. WebVoyager is one test set, success rates are sensitive to evaluation methodology, and frontier-lab models can quickly recapture the lead with a model update. But the data does suggest that the architectural differentiation between the funded and the lean players is smaller than the funding rounds imply. The technology is converging.
The foundation labs: quietly winning OSWorld
The frontier labs shipped browsing agents within 90 days of each other in late 2024 / early 2025. Anthropic’s Computer Use (Oct 22, 2024), Google’s Mariner (Dec 11, 2024), OpenAI’s Operator (Jan 23, 2025).
The benchmark spread on WebVoyager and OSWorld tells the strategic story:
- WebVoyager: OpenAI CUA 87%, Google Mariner 83.5%, Anthropic Computer Use 56%
- OSWorld: CUA 38.1%, Anthropic 22.0% (humans 72.4%)
- WebArena: CUA 58.1% on simulated e-commerce / CMS tasks
The startups are competitive on web tasks. They are not competitive on multi-step OS tasks, where OpenAI’s CUA dominates. That gap is unlikely to close by undercutting on price or shipping a better SDK; it requires foundation-model capability the startups do not have.
What that means is the venture-backed browser-agent class is fundamentally a distribution and developer-experience play, not a capability play. Capability is downstream of the foundation models. The startups win by being where the developer is, with the cleanest SDK, at the lowest cost.
What the architectural convergence means
All four players — Browserbase, Browser Use, Firecrawl, and the surviving fragments of Reworkd — converge on the same primitive: cloud Chromium with an LLM-driven control plane. Differentiation will not come from the architecture; it will come from packaging.
The packaging axes that will matter:
- MCP integration depth: discussed previously. The agent that picks tools through MCP routes around vendor branding entirely.
- Per-session economics: the cost-per-action gap between the cheapest and most expensive vendor is roughly 3x. That gap is wide enough to drive vendor selection in production.
- Stealth posture: browsers that pass DataDome and Cloudflare Turnstile out of the box capture more of the high-value-target use cases. The vendors that don’t ship a stealth path become read-only tools for low-protection sites. Bright Data’s Scraping Browser is the most-cited managed alternative when the buyer wants the stealth layer outsourced to the vendor.
- Recording and replay: enterprise buyers increasingly want session recordings for compliance and debugging. Browserbase ships this; Browser Use is shipping it. The smaller vendors mostly do not yet.
The unresolved question is what happens to the category when Operator and Mariner ship developer APIs at scale. Right now both are gated to specific surfaces (ChatGPT Pro, Gemini Advanced). When the foundation labs open up to general API consumption, the venture-backed substrate plays compress hard.
The current scale of the field — $74mn funded, 79K GitHub stars on the leader, multiple production benchmarks — is enough to support several survivors. The next venture round will consolidate to two or three. The open question is whether any of them survive Operator becoming a developer API.
Sources