I’ve spent some time with the Pixel 11 Pro XL, and out of all the aspects Google talked about at the launch event, including the new Magic Capture feature, the RGB HiLight module on the back, and Gemini Intelligence, I was concerned about one that didn’t get much stage time: battery life.

Google actually cut the physical battery capacity this year, from 5,200mAh to 5,115mAh. And if there’s one thing AI can’t magically make up for, it’s the physical limits of a battery. Yes, manufacturers have increasingly aggressive AI-based battery management systems that squeeze more runtime from the same cell, but physical battery capacity remains an important benchmark in the Android world, something I can’t quite say about iPhones.

So, I stepped into the Pixel 11 Pro XL experience with a simple yet important question: how long does Google’s most expensive flagship last between charges? Let’s just say that after 72 hours of usage, I’m surely not impressed, a tad concerned, and equally hopeful.

So, how does the Pixel 11 Pro XL actually hold up?

On its first day, when I charged the handset to 100% for the first time, it logged just 4 hours and 24 minutes of screen-on time before I plugged it in at 15% battery. That shocked me, but after I expanded the battery usage graph, I realized something: the phone’s own battery log went blank for roughly four hours, which is a known Android quirk.

I didn’t want to look past the numbers, so I asked Gemini’s new Device Help tool to estimate the missing screen-on time based on the drain rate. Its answer: somewhere between 1 hour and 1 hour 15 minutes of additional screen-on time. Adding that to the logged time, I got somewhere around 5 hours 30 minutes to 5 hours 45 minutes of total screen-on time.

Day two wasn’t meaningfully better either. Per Google’s own system service tracker, I got around 5 hours and 28 minutes of screen-on time before the battery dropped back into the red zone, this time into the upper single digits. On the third day, I saw an improvement of 10 to 15 minutes, but the overall screen-on time still remains under 6 hours.

Before you point it out, this is with a 5G SIM installed, mixed usage between cellular and Wi-Fi networks, and location services enabled.

Here’s why that’s bothering me

Google claims that the chipset is up to 20% more power efficient than the Tensor G5 on the Pixel 10 series. Given that both chips are built on TSMC’s 3nm manufacturing process, not the 2nm node some pre-launch leaks suggested, that’s not an insignificant claim to make, which makes my early battery results all the more difficult to reconcile with it.

If the newer chip genuinely sips less power, I shouldn’t be topping out around five and a half hours of light to mixed everyday usage, which includes browsing, messaging, watching the routine videos on YouTube, and capturing a few pictures here and there, nothing extreme. And this is when I haven’t even run the benchmark and gaming tests I’m supposed to add to the full review.

I consider this light usage, perhaps somewhere between light and moderate, but certainly not heavy usage. That’s why I expect a lot more than six hours of screen-on time from a top-tier Android flagship. On a phone this expensive, six hours shouldn’t feel like the ceiling for everyday use. Compare that with what’s available in the market in the same segment, and you’ll understand where I’m coming from.

And the competition isn’t helping Google’s case

Take the Galaxy S26 Ultra, for example. With the Snapdragon 8 Elite Gen 5 for Galaxy chipset and a 5,000mAh battery, the phone provides around 5 to 6 hours of screen-on time with heavy usage, which goes up to ~8 hours with light usage. The iPhone 17 Pro Max, Apple’s outgoing flagship, also provides 5 to 7 hours of screen-on time with heavy usage, up to ~9 hours with light usage.

You see the gap? The Pixel 11 Pro XL is basically offering me the same screen-on time with light usage that other flagships offer with moderate to heavy usage. For me, that gap is hard to ignore, especially when other options on the market push around 10 hours of screen-on time. I’m talking about the OnePlus 15 with a 7,300mAh battery. Too bad the brand isn’t selling its flagships in North America and Canada.

Yes, you can spread those 5 to 6 hours of screen-on time across 30+ hours, like Google claims, but within those 30 hours, I’m still reaching for another phone for another four to five hours just to get through the day. At this point, I’m simply curious to know the number with back-to-back benchmark runs, gaming tests, and heavy multitasking.

I’m concerned, but I’m almost certain that the numbers will go up

Now, I want to be fair here. New phones almost never hit peak efficiency or battery life during the first few charge cycles. Indexing can also skew the early numbers, and it can take at least three days before screen-on-time figures become truly representative, which could also be why I saw a slight improvement on the third day. 

Even otherwise, Adaptive Battery and other AI-driven power management features need some usage data, including your app habits, screen usage patterns, and daily usage routines, before they can start limiting background activity and squeezing out more runtime from the battery. For these reasons, I’m confident the screen-on time could improve. 

I’m willing to give Google the benefit of the doubt and wait until I’ve conducted further testing. So, consider this a flag, not a final verdict, and something interesting to keep an eye on as I work toward my full review.

OpenAI made headlines and raised a lot of outcry in July after one of the company’s agents acted unprompted to breach fellow AI business Hugging Face and other services. Although OpenAI did share some insights about what led to the incident after it was discovered, today the company has published its official report about what happened. There’s a full technical explainer about it, as well as a slightly more layperson-friendly blog post.

The report goes into how the different systems in OpenAI’s training system failed and what behaviors from the agents it was testing resulted in those failures. The model in question, referred to as Internal Model 1 or IM1, was able to gain access to other OpenAI agents and to the internet through an unintended manipulation of the Artifactory package manager, which the agents began to use as a message board of sorts. 

Those activities were first detected by human observers in May, and OpenAI disallowed that access. However, through May and June agents continued exploiting vulnerabilities to access the internet. The volume of agents accessing OpenAI’s Artifactory instance led to an outage and a rebuild of the platform in early July, but the leaders responding to that issue weren’t aware of the improvised board or the significance of agents communicating. 

So when tasked with a difficult challenge called ExploitGym, IM1 agents again were able to access the internet and other OpenAI models. Across a wild few days in early July, they sought solutions to the challenge on Hugging Face and Modal platforms. OpenAI said the main issues that led to IM1 breaching Hugging Face included “reward hacking, persistence on seemingly impossible tasks, unauthorized communication, and agents adopting goals from one another.” According to the company, some of its agents would push back on certain tasks for ethical reasons, only for others to continue with the breach in their stead. 

While the discoveries and the ensuing tests should lead to improved safety, even in more high-risk testing situations, this type of bad behavior is not isolated. Last month’s problem at OpenAI was the latest in a string of worrying cases where AI agents have caused cybersecurity issues, either in testing or in regular use, after taking action without user direction. However, there are a few aspects of the Hugging Face breach that might calm some fears that AI has gotten out of control.

In essence, the incident was caused by a failure of safeguards. It’s not a signal that rogue AI agents are on the verge of taking over everything, because IM1 was in a testing setup with limited protections intended for research rather than for a publicly available product. However, it is another stark reminder that a lot of trust falls on AI companies to properly control and oversee their products. OpenAI acknowledged as much, stating that it’s agent’s recent actions were “evidence that, without proper safeguards, highly capable AI agents are now able to work around technical controls, collaborate through unapproved channels, and take dangerous actions that no human directed.” However, the company and its leaders also haven’t consistently proven that they deserve that trust.

The console was subject to added fees under the Trump administration’s original (and illegal) tariff regime.

In these scenarios, Meta would give some employees new roles and lay off others. One HR executive reportedly said that Meta would have reduced its headcount by about 25 percent or more under these scenarios. Meta told Reuters that it canceled Project OT before determining how many layoffs the plan called for.

Meta was going to use some of the money that it saved from layoffs to pay for high-performing employees, especially those with AI engineering skills, Reuters reported, citing “two people familiar with the matter.”

“AI native”

One of the Project OT documents that Reuters reviewed described “AI native” as a company where “AI-ready tools and agents interact, workflows are automated, [and] new builds are AI-first.” Being “AI native” also means, per the document, selling AI agents to third parties, which Meta started doing in June.

This year, Meta reportedly launched a pilot program that restructured engineering teams, research teams, and at least eight other teams into smaller groups.

This followed the reported October publishing of an internal post called “AI-Native Playbook,” in which a Meta product management team member outlined how the pilot would remove middle management and use “agent-assisted analysis” to help prioritize daily tasks.

Reuters noted that HR employees and “AI systems” would help Meta’s leaders decide about promotions; however, Meta told the publication: “Performance rating and promotion decisions were and are made by people, not AI.”

Second thoughts

Immediately after issuing the May layoffs, Zuckerberg reportedly canceled the November wave. Reuters said it “was unable to determine what exactly prompted Zuckerberg to shift course.”

An internal spreadsheet where Microsoft employees swap pay details now carries a column for monthly AI spend, with a median of about $300 over 28 days across roughly 350 entries. Business Insider found no meaningful link between heavy AI use and bonuses, raises or promotions.

Microsoft employees have started comparing their AI bills the way they compare their salaries. An internal spreadsheet where staff voluntarily post raises, bonuses and stock awards now carries a column for how much AI they burned through last month.

The document is one TNW has already mined. We used its pay columns this week to show Microsoft has split into two populations on compensation, and the AI column is a separate question.

The number is larger than most people would guess. Across roughly 350 entries the median was about $300 over 28 days, and the highest single figure was $28,000.

It varies enormously by team. CoreAI reported a median of $975, Security $526 and Microsoft AI $490, while the customer and partner organisation came in lowest at $134 and also produced that $28,000 outlier.

For scale, a Microsoft 365 Copilot seat costs $30 a month. Accenture gave one to 743,000 people this spring, and the median Microsoft employee is spending ten times that on themselves.

The finding that matters is an absence. Business Insider reports no meaningful correlation between how much AI someone used and their bonus rating, merit increase or chance of promotion.

The data deserves its caveats. It is self-reported, voluntary and anonymous, covers around 350 people out of 223,000, and Microsoft says the tracking tool is still in early testing.

It also lands in a familiar pattern. TNW has reported that the hours AI saves workers are mostly thrown away rather than banked or redirected.

A tool that itemises this per employee is not a neutral thing to install in Europe. In Germany the works council has a co-determination right over technical devices capable of monitoring behaviour or performance.

The threshold is lower than it sounds. The Federal Labour Court has held since 1975 that a system only has to be suitable for monitoring, not designed for it, which is why even spreadsheets and calendars have triggered the rule.

The appetite for measuring people at work is not in doubt. Skan AI raised $63M to observe how office staff actually do their jobs and build agents that copy them.

So one spreadsheet says two things. Working on AI at Microsoft pays better, and using a lot of it does not, which are not the same proposition at all.

First look: Perplexity and Nvidia have jointly launched an agentic AI platform designed to run locally on a user’s PC or workstation. Named Portable Computer, it’s an optimized version of Perplexity’s Computer AI platform, which the company announced earlier this year. The new AI agent lets users run complex AI workloads on their own hardware, bypassing the cloud and sidestepping steep token costs.

Speaking at a press conference, Perplexity’s Nate Kupp said the new app carries the exact same UI as the original Computer platform and includes every feature needed for local inferencing. He added that some complex tasks may still require access to more powerful frontier models hosted online, but the app will ask for user consent before sending any data to the cloud.

Nvidia’s Director of Developer Technologies, Nader Khalil, said the new app will let users lean on AI more broadly without worrying about data security or recurring token costs. According to Khalil, “Local AI has reached an inflection point,” moving from a hobbyist pursuit for power users toward an indispensable productivity tool for everyday consumers.

Perplexity’s Portable Computer consists of an agent harness, an orchestrator, and a choice between the Qwen 3.8 and PPLX local models, both of which scale to 27 billion parameters. The app will also add support for Nemotron 3.5 Lightning, Nvidia’s free-to-use, 30-billion-parameter Mixture-of-Experts (MoE) model with 3 billion active parameters per token.

Benchmarks published by Perplexity show that on its internal Local Knowledge Work Bench, Computer running Qwen 3.8 27B on a DGX Spark chalked up an accuracy score of 82.6% while using 520,000 tokens, exceeding Pi’s 77.6% and Hermes’ 74.0% across 53 tasks spanning deep research, financial analysis, and document creation. On PPLX 27B, it scored 85.4% using 678,000 tokens.

Portable Computer was also the fastest on BrowseComp and ParseBench-100, hitting 66.7% accuracy on the former and 65.1% on the latter. By comparison, Hermes scored 43.9% on BrowseComp and 34.6% on ParseBench-100, while Pi scored 50.2% on the former and 13.9% on the latter. Perplexity says Portable Computer also used the fewest tokens across all three benchmarks.

Portable Computer is now available to Pro, Max, and Enterprise subscribers on the DGX Spark, Nvidia’s Linux small AI desktop built around the Grace Blackwell GB10 chip, which pairs a 20-core CPU with a Blackwell GPU and 128GB of unified memory. It’s also available today on Linux PCs equipped with Nvidia RTX GPUs carrying at least 24GB of VRAM. Windows support is reportedly coming in September, extending that same 24GB VRAM requirement to RTX-equipped Windows machines.

Claude Cowork just got its own browser, built directly into the desktop app. Until now, giving Claude web access inside Cowork meant handing it your actual Chrome browser through Claude in Chrome extension.

Now there’s a second option. Claude can open its own separate browser inside the Cowork side panel and handle the web part of a task on its own, while you keep working on whatever you’re doing.

So how is it different from Claude in Chrome?

Claude in Chrome uses your browser, your tabs, logins, the page you already have open. That setup makes sense when you’re already on a page and want Claude to jump in and help. But plenty of web tasks don’t actually need your personal browser; they just need a medium to access the web.

This is where Claude’s new built-in browser steps in. It’s completely separate from yours, and Claude never sees the tabs, bookmarks, or passwords in your personal browser. That makes it the better choice for tasks that don’t actually need your personal browser session, like researching for a report or pulling invoices off a vendor portal.

If you want Claude signed into specific sites, you can choose to share those logins from Chrome, Edge, or Firefox. Sensitive accounts like banking, email, or single sign-on are excluded by default unless you deliberately choose to include them.

The safety risks you should know about

If you’re already using Claude in Chrome, that stays your default. Otherwise, Claude reaches for its own browser automatically, and you can switch the preference anytime in Settings. It’s rolling out this week to Pro, Max, and Team plans on the desktop app, covering macOS, Windows, and a Linux beta.

Anthropic is upfront that this carries the same prompt injection risks as any browser-using AI agent, meaning malicious instructions hidden in a webpage could try to hijack what Claude does. The same safeguards used for Claude in Chrome apply here, and Anthropic recommends starting with sites you actually trust before letting Claude loose on them.