Twenty-five hours. That's what Recursive Intelligence claims its new browser saved one early customer each week—per person, according to the company. Time enough that the company in question reportedly scrapped its Clay subscription altogether.
Those numbers, while anecdotal and awaiting formal case study validation, would turn heads in any sales or recruiting department currently buried under mountains of manual grunt work. And that's precisely what the small San Francisco startup is banking on.
On July 29, 2026, Recursive Intelligence emerged with Polar, an AI-powered browser that clicks, types, and navigates the web while logged into your accounts. The announcement came wrapped in what the company disclosed as $5.7 million in fresh seed funding led by Madrona, with checks from a roster of notable angels including Thomas Dohmke, the former CEO of GitHub, and Erik Bernhardsson from Modal.
The pitch is deceptively simple. But the execution? That's where things get interesting—and potentially complicated.
A Browser That Works While You Sleep
Polar isn't a Chrome extension or a sidebar assistant. It's built on Chromium itself, designed to handle sprawling, multi-site workflows that can stretch from minutes to hours. Think of it as a very persistent digital employee who never gets bored.
The use cases the company rattles off feel ripped from the daily grind of modern knowledge work: scraping LinkedIn to build prospect lists, enriching company databases with decision-maker contacts, checking outreach replies twice a day and updating CRM trackers. Or dispatching 15 personalized LinkedIn connection requests every morning before you've finished your coffee.
Recruiters, meanwhile, might use Polar to source candidates, prepare interview packets from job postings, triage overflowing inboxes, or sweep job boards weekly without lifting a finger.
Over the seven months leading up to launch, the browser racked up more than 4.5 million actions across research, recruiting, sales, operations, and marketing functions, according to the company. That's a lot of clicking.
The Benchmark Wars Begin
Naturally, Recursive Intelligence wants you to know their product is fast. Very fast.
According to the company's internal benchmarks, Polar "leapfrogs both OpenAI and Anthropic" on browser-agent tests—bold words in a field where everyone's jostling for pole position. They've published evaluation results showing Odysseys scores of 98.0 out of 100 and BU Bench V1 scores of 95 out of 100.
For context, they position competitors like Aside at 75.5, Browser Use at 70.0, GPT-5.4 at 60.8, and Claude Opus 4.6 at 44.5 on the same tests. The evaluation repository on GitHub breaks down performance across categories like InteractionTests (20/20), WebBenchREAD (20/20), and GAIA (16/20), using Gemini-2.5-flash for judging.
But here's where healthy skepticism kicks in. At least one competitor—Aside—also claims top rankings on similar benchmarks. Which suggests we're watching the familiar early-stage dance of startups cherry-picking metrics that make them look good. Benchmarks in AI have a way of mattering intensely until they don't.
From Composer to Polar: An Architectural Gamble
The product didn't start as Polar. Back in January 2026, it was called Composer—an autonomous browser agent running on Electron, the framework that powers apps like Slack and Visual Studio Code.
Then, in May, the team made what they describe as a significant call: tear it down and rebuild as a "proper Chromium browser." The shift wasn't cosmetic. Running as a full browser rather than an Electron wrapper gave the team what they wanted most: control. Tighter hooks into navigation, session management, and the kind of multi-hour workflows that would choke a lighter-weight implementation.
The rebranded Polar launched with an initial macOS focus, though a Windows 10+ build slipped into beta by late July. Whether that architectural bet pays off in the long run remains an open question—one that will likely be answered by how well Polar handles the messy, unpredictable chaos of real-world web navigation.
MIT Credentials and Madrona Money
Recursive Intelligence bills itself as an applied AI lab, a self-description that's become something of a genre in the current fundraising climate. The three-person founding team carries credentials from MIT, Y Combinator, Jane Street, Citadel, Apple, Perplexity, and Modal, along with various medals and awards name-checked in the announcement. They're operating out of 629 Market Street in San Francisco—a downtown address that carries its own set of implications about ambition and burn rate.
Madrona's involvement lends institutional weight, though as of July 30, 2026, no external confirmation from the venture firm had surfaced publicly. That lag isn't unusual for early-stage announcements, where press releases sometimes run ahead of official portfolio page updates.
Pricing, Platform, and the Privacy Question
Polar operates on a credit-based consumption model. Longer, more complex tasks burn through more credits—a pricing approach that mirrors how cloud infrastructure bills compute. The Team plan runs $50 per seat monthly, bundling 25,000 pooled credits per seat, shared billing, usage analytics, and priority support.
Right now, the browser runs primarily on macOS, with Windows in beta and SOC 2 compliance somewhere down the pipeline. Privacy controls include what the company describes as "zero data retention" arrangements with commercial LLM providers, meaning those providers are contractually prohibited from training on customer data. Polar itself, however, retains operational and observability data necessary to keep things running—a distinction worth noting for compliance-minded teams.
The browser stores profile data locally and keeps run history server-side. It claims compliance with Google API Services User Data Policy for connected accounts like Gmail, Drive, Calendar, and Sheets. Whether that's enough to satisfy enterprise security teams is another matter.
Launching Into a Suddenly Crowded Field
Timing in tech is everything, and Polar is arriving just as everyone else had the same idea.
Google rolled out Chrome "AI Mode" and "Auto Browse" features between January and April 2026. Brave opened AI browsing to early testers in May. Palo Alto Networks positioned its Prisma Browser for what it called the "agentic AI era" in March, leaning hard into security controls for enterprise deployments—a signal that the paranoia is justified.
Because here's the thing: security concerns around AI browsers are real and growing. A Cloud Security Alliance research note from March outlined what they called "agentic attack surfaces," including browser-integrated AI panel hijacks. Multiple research papers throughout the year have flagged prompt injection and data exfiltration risks. Livescience summarized warnings in July that AI web browsers, frankly, aren't ready for broad public use given the security gaps still being patched.
Recursive Intelligence is threading that needle with local storage, server-side retention policies, and compliance frameworks. Whether it's enough to sleep soundly at night is a call each customer will have to make for themselves.
Show Me the Data
The company promises case studies "in the coming weeks" to flesh out that eye-catching 25-hour time savings claim with more customer data. Until then, we're left with the 4.5 million actions across seven months and those benchmark scores as the primary traction signals.
For sales leaders and recruiting teams evaluating AI automation tools right now, Polar represents a specific bet: that the browser itself—not a sidebar, not an extension, but the entire operating environment—is the natural home for agentic work. It's an architectural thesis that makes intuitive sense.
Whether it delivers sustained competitive advantages over lighter-weight competitors, or whether it becomes just another tool in an increasingly cluttered stack, is the question that only time and real-world usage will answer. The next few months should tell us whether Recursive Intelligence is onto something, or whether they've simply built a very sophisticated way to automate tasks that enterprises will ultimately decide are too risky to hand over to an AI.
Either way, someone at Clay is probably keeping an eye on this.
