AI Browsers Are Here, but Why? Grok 4, Perplexity's Comet, and More
Grok 4 breaks benchmarks. AI browsers break the web?
Welcome to The Median, DataCamp’s newsletter for July 11, 2025.
In this edition: xAI releases Grok 4, the browser as we know it is changing, AI slows down developers by 19%, and more.
This Week in 60 Seconds
xAI Releases Grok 4, Its Most Powerful Model Yet
xAI has launched Grok 4 in two variants: Grok 4 and Grok 4 Heavy. There’s no smaller or faster “mini” version available. While both models top benchmarks like Humanity’s Last Exam and ARC-AGI, the context window—128k in the app and 256k via API—may fall short for long-context production workflows. Our team at DataCamp tested Grok 4 and compiled the results in this blog. We’re also taking a closer look at Grok 4 in this week’s Deep Dive.
Perplexity Launches Comet, an AI-Powered Browser
Perplexity has introduced Comet, a new web browser that integrates AI assistance directly into the browsing experience. Built on Chromium, Comet offers features like summarizing content, comparing products, and automating tasks such as scheduling meetings or sending emails. Currently, Comet is available to Perplexity Max subscribers, with broader access coming soon.
OpenAI to Launch AI-Powered Browser
OpenAI is preparing to release its own AI-powered web browser in the coming weeks, according to reports. Built on Chromium, the browser will feature a native ChatGPT interface and integrate OpenAI’s Operator agent, enabling it to perform tasks like booking reservations and filling out forms autonomously. In the Deeper Look section, we’ll examine the growing trend of AI browsers and unpack what’s really driving companies to build them.
Study Finds AI Slows Down Experienced Developers
A new randomized controlled trial found that experienced open-source developers became slower when using AI. Despite expecting a 24% speed boost and later estimating a 20% gain, developers using tools like Cursor Pro and Claude 3.5/3.7 were 19% slower on average across 246 real tasks. Instead of coding, time was lost prompting, waiting, and cleaning up AI output. While the study is thoughtfully designed, it reflects a specific context—experienced devs working in familiar, complex repos—so its findings may not generalize across all teams or workflows.
Cursor Faces Backlash Over Pricing Change
Cursor (the most popular AI-powered code editor) drew criticism after changing its pricing structure, leaving many users surprised by how quickly credits ran out, especially when using models like Claude. CEO Michael Truell issued an apology and offered refunds, but the backlash has led many developers to look elsewhere. Our team is actively working on tutorials covering alternatives like Cline or Roo Code, so keep an eye on our Substack notes or the DataCamp blog.
Microsoft Launches Phi-4-mini-flash-reasoning for Speed-Critical Tasks
Microsoft has introduced Phi-4-mini-flash-reasoning, a compact 3.8B parameter model optimized for math reasoning in latency-sensitive environments like mobile, edge, and education tech. Thanks to a new decoder architecture called SambaY, the model achieves up to 10x higher throughput and 2–3x lower latency compared to its predecessor, while retaining strong long-context performance. It’s available now on Azure AI Foundry, Hugging Face, and NVIDIA’s API catalog.
New Course: Building Scalable Agentic Systems
A Deeper Look at This Week’s News
Why Is Everyone Building an AI Browser?
Perplexity has just launched Comet, an AI-powered browser, and rumors suggest OpenAI is preparing to release one of its own. On the surface, this looks like a race to build a faster, smarter browsing experience. But underneath, there’s more: a race for the data that fuels LLMs.
What are AI browsers about?
Perplexity’s Comet positions itself not as a traditional browser but as a “thought partner for the web.” It lets users ask questions about any page, summarize content, compare products across tabs, and automate tasks like booking meetings or sending emails—all without switching tools.
Source: Perplexity
OpenAI’s upcoming browser is expected to do much of the same, but with tighter integration of its Operator agent: a tool that navigates websites autonomously, mimicking clicks, form entries, and scrolls.
Why not let users pull the data themselves?
The timing isn’t coincidental. As we pointed out in our previous newsletter, Cloudflare has recently rolled out a major policy change: all new domains using its infrastructure now block known AI crawlers by default.
It also introduced a Pay Per Crawl feature, requiring AI companies to compensate site owners for access. These moves reflect mounting frustration from publishers who feel their content is being scraped without permission or payment.
AI browsers offer a workaround. If bots are blocked, why not let users pull the data themselves, assisted by AI? Every AI-driven summary or form fill looks like a regular user browsing a page. It’s harder to detect, harder to stop, and legally murkier.
Convenience vs. visibility
There’s a clear appeal to AI-powered browsers. They reduce friction, help users cut through clutter, and surface relevant information faster than most people can navigate a complex website. For anyone overwhelmed by tab overload or repetitive online tasks, these features offer tangible benefits.
At the same time, there’s a shift in how decisions are made during browsing. When AI summarizes content or offers suggestions, it shapes what the user sees and interacts with. That’s not necessarily a bad thing, but it does raise questions about transparency. As browsing becomes more mediated, users may lose some visibility into how results are prioritized or why certain actions are recommended.
Grok 4 Essentials
xAI has released Grok 4 in two configurations: Grok 4 and Grok 4 Heavy. This section covers the essentials, including key benchmarks, usage limitations, and access. For a more detailed breakdown, including our own testing of Grok 4 in math, coding, and long-context tasks, check out the full blog post.
Grok 4
Grok 4 is a single-agent model built for structured reasoning. It’s designed to tackle high-difficulty tasks in math, programming, finance, and other domains that rely on step-by-step logic.
In benchmarks, it scored 41.0% on Humanity’s Last Exam with tools enabled and outperformed Claude, Gemini, and GPT-4 on several STEM evaluations.
Source: xAI
However, its context window is limited to 128k tokens in-app and 256k via API, which may be a constraint in long-context production workflows. Multimodal input is technically supported but currently unreliable.
Grok 4 Heavy
Grok 4 Heavy runs multiple Grok 4 agents in parallel, each solving the same task independently before comparing answers. Once the agents generate their responses, the system synthesizes a final answer based on agreement or majority logic.
It scored 50.7% on Humanity’s Last Exam and showed strong gains in ARC-AGI and other reasoning-heavy tests. That said, it’s compute-intensive and aimed at niche use cases like research, scientific modeling, and financial simulation.
Access and roadmap
Grok 4 is accessible through the X app, grok.com, and the xAI API.
xAI has also outlined its upcoming releases:
August: A coding-specialized model
September: A more capable multimodal agent
October: A video generation model, reportedly in training on 100,000 GPUs
Industry Use Cases
AI-Generated Band “The Velvet Sundown” Tops Spotify Charts
“The Velvet Sundown,” a ’60s-style band, gained rapid popularity on Spotify with their debut album “Floating on Echoes.” However, it was later revealed that the band members do not exist and that their music is entirely AI-generated. The project, directed by human creators, sparked discussions about authorship and identity in the digital age, with mixed reactions from listeners and music experts.
FDA’s AI Tool ‘Elsa’ Enhances Food Safety Inspections
The U.S. Food and Drug Administration (FDA) has introduced “Elsa,” an AI tool designed to improve the agency’s ability to analyze safety data, identify labeling inconsistencies, and prioritize food safety inspections. Operating behind the scenes, Elsa assists FDA staff in processing complex reports more efficiently, potentially improving the speed and accuracy of responses to food safety risks and recalls.
GlobeScribe Introduces AI Translation for Fiction Publishing
UK-based startup GlobeScribe.ai has launched an AI translation service targeting fiction writers and publishers. The service claims to produce translations nearly indistinguishable from human work. While some reviewers praised the AI’s fidelity and tone, the translation community has raised concerns about the potential loss of cultural nuance and the commodification of literary translation.
Tokens of Wisdom
It's surprising that we have been able to develop AI without understanding how it does what it does.
—Ken Liu, Award-Winning Science Fiction Writer
Listen to Ken Liu discussing what science fiction can tell us about the future of AI in this DataFramed podcast episode.
Amazing work!!!