Blogroll

We don't often think of our monitors as something that holds the rest of our computer back, but your monitor is the component you use as a window into your computer. It's one of the peripherals you spend the most time using, by far. So when certain specs and missing features start limiting what your PC can offer you, whether for work or play, then it's probably time to start looking for a new screen as your most important upgrade.

In the battle to improve the capabilities of artificial intelligence, Google and OpenAI have both released new versions of their flagship chatbots: Gemini 3 and GPT-5.1. Google calls Gemini 3 its "most intelligent model yet," while OpenAI pitches GPT-5.1 as "smarter" and "more conversational."

When you compare the two, it really comes down to what you want from your environment-killing chatbot. On the AI side, both Gemini 3 and GPT-5.1 feel like real sequels instead of point upgrades. Google says its Gemini 3 model builds upon the agent-y, multimodal ideas from Gemini 2 and 2.5, fusing them into a single model that is better at reasoning, handling long-context multimodal work, and planning ahead. It also now powers features like Google's new Antigravity developer platform and a Deep Think mode for genuinely challenging problems.

SEE ALSO: White House pulls back on AI laws executive order

GPT-5.1, meanwhile, reportedly keeps the raw intelligence of GPT-5 but makes it way more direct and enjoyable to talk to. Instant and Thinking both adjust how much they "think" based on the question; they follow instructions more consistently, and they let you dial in the exact tone and personality you want across every chat.

So, to be as useful as possible, here's a short breakdown of the key differences between the two models.

Gemini 3 versus GPT-5.1: Price

When it comes to cost, the two models occupy different price points. GPT‑5.1 from OpenAI has API pricing at about $1.25/ 1 M input tokens and $10.00/ 1 M output tokens. Meanwhile, the Gemini 3 Pro from Google lists token-based tiers at roughly $2.00 input / $12.00 output per million tokens for contexts of up to approximately 200,000 tokens, and about $ 4.00 and $18.00 for much larger contexts beyond that.

On the consumer subscription side, Gemini 3 offers a Pro tier at around $19.99 / month and an Ultra/Enterprise tier with custom pricing (reported to be up to ~$250/month for full features), while GPT-5.1-related consumer plans fall in the approximately $20/month and up range, depending on workflow needs.

Gemini 3 versus GPT-5.1: Model rankings

According to the latest LMArena rankings, Gemini 3 is sitting at the very top of the board with a score of 1324, making it the current fan favorite across thousands of votes. It’s not just edging out the competition, either. The next closest model is Gemini 2.5 Pro at 1249, which means Gemini 3 is leading by a meaningful margin.

GPT-5.1, listed on the board as GPT-5-chat, comes in further down with a score of 1222, placing it solidly in the upper tier but clearly behind Gemini 3’s surge. What’s interesting is the company GPT-5.1 has in those rankings: it’s surrounded by previous GPT generations and models, such as o3 and Claude Opus, all clustered tightly in the low 1200s.

Gemini 3, meanwhile, is the only model that breaks out of that pack entirely. The vibe from LMArena voters is pretty unmistakable. Gemini 3 is the model to beat right now, while GPT-5.1 is respected and well-liked but not commanding the same kind of top-of-the-charts energy.

Gemini 3 versus GPT-5.1: Key features

When it comes to multi-modal and long-context capabilities, Gemini 3 clearly outperforms. According to Google, it supports massive contexts (hundreds of thousands of tokens), handles images, text, and code with seamless transitions, and introduces a Deep Think mode for higher-order reasoning and planning across domains.

Meanwhile, GPT‑5.1 focuses more on conversational quality and reliability. The chatbot maintains strong performance in pure text generation and instruction following, with tone adjustments and personality fine-tuning built in. However, tests conducted by Tom's Guide indicate that it doesn’t yet match Gemini 3’s breadth in modality or long-form reasoning.

In terms of accessibility and feature set for typical users, GPT-5.1 offers familiar workflows, including chat-based input, strong coherence over shorter contexts, and a refined interface for controlling tone, style, and persona. OpenAI says this makes it great for writing, Q&A, and interactive assistance.

On the other hand, Gemini 3 is built to scale for grander tasks — including multimodal inputs, more complex planning, and longer spans of context — which gives it the edge when tackling multi-step workflows, combining visuals and text, or generating highly technical outputs.

Disclosure: Ziff Davis, Mashable’s parent company, filed a lawsuit in April against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.

Best Black Friday unlocked phone deals at a glance Best Google Pixel deal Google Pixel 10 (256GB) $699 (save $200) Get Deal Best Samsung deal Samsung Galaxy Z Fold7 $1,599.99 (save $200) Get Deal Best Motorola Deal Motorola Moto G Stylus $284.99 (save $115) Get Deal

If you've been limping along with a sad phone that will no longer hold a charge, it's time to make the upgrade. Even if you try to limit the amount of time spent looking at your phone, it's a necessary piece of tech these days. An unlocked phone allow you the ability switch plans, carriers, or swap to a new phone if you'd like. Plus, they're excellent for installing an e-sim for international travel.

While Black Friday isn't officially here yet, the major retailers aren't waiting and they're packed with great deals on unlocked phones. Options include basic budget models that are perfect for kids or anyone who doesn't need fancy features. Or you can go with some of the latest and best designs. Keep in mind iPhones tend to be the exception here, and it's not common to find a great deal on an unlocked iPhone. Carriers tend to offer the best value for those.

Before we fill up on pie later this week, shop these unlocked phone deals before they disappear.

Best Black Friday unlocked phone deal Opens in a new window Credit: Google Google Pixel 10 $699 at Amazon
$899 Save $200   Get Deal Why we like it

Google's lineup of Pixel phones rarely disappoints. Instead, they offer a great value for those aren't tied to iOS. The Pixel 10 comes with all the essentials we've come to love about Pixel phones like solid battery life, a bright display, and a telephoto camera lens. In Mashable's review of the Pixel 10, Tech Reporter Alex Perry noted it's a great upgrade if you current phone needs an update and he was able to get about 26 hours of life before it asked for a recharge, not including time spent sleeping. The $699 sale price applies to the 256GB version, which is $200 off and matches the record-low. But you can also go with less storage and spend $599 on the 128GB Pixel 10 which is also $200 off the normal.

More unlocked phones on sale for Black Friday

• Motorola Moto G (128GB) — $149.99 $199.99 (save $50)

• TCL 60 XE NXTPAPER (128GB) — $169.99 $249.99 (save $80)

• Motorola Razr 2024 (256GB) — $399.99 $699.99 (save $300)

• Samsung Galaxy S25 (128GB) — $674.99 $799.99 (save $125)

• Google Pixel 9 (128GB) — $499 $799 (save $300)

• Nothing Phone (3) (256GB) — $639 $799 (save $200)

• OnePlus 13 (512GB) — $799.99 $999.99 (save $160)

• Motorola Razr Ultra (512GB) — $884.99 $1,299.99 (save $415)

• Samsung Galaxy S25+ (512GB) — $940.79 $1,119.99 (save $179.20)

• Samsung Galaxy S25 Ultra (256GB) — $949.99 $1,299.99 (save $350)

• Samsung Galaxy Z Flip7 (512GB) — $1,019.99 $1,219.99 (save $200)

• Samsung Galaxy S25 Ultra (512GB) — $1,057.91 $1,419.99 (save $362.08)

SCORE A FREE DEVICE WITH VERIZON: New customers who sign up for an eligible Verizon Fios or 5G/LTE Home Internet plan can choose one high-value device for free: a Nintendo Switch console, a Samsung Galaxy Tab S10 FE 5G tablet, or a 43-inch Samsung Q7F TV.

Opens in a new window Credit: Nintendo Free Nintendo Switch Sign up for an eligible Verizon Fios, 5G Home Ultimate, or LTE Home Plus internet plan and instantly qualify for a free Nintendo Switch console. Get Deal

Black Friday shopping usually means scoring deals on tech and appliances, but some of the best savings are the ones that bundle a product you want with a service you need. If you've been looking for a reason to finally upgrade your home internet, Verizon is handing out a news-worthy incentive: the choice of a free Nintendo Switch, a Samsung TV, or a tablet.

SEE ALSO: Black Friday 2025: Live updates on the latest deals from Amazon, Target, Walmart, and more

As of Nov. 24, you can sign up for an eligible Verizon Fios, 5G Home Ultimate, or LTE Home Plus internet plan (which start as low as $35 per month) and instantly qualify for a free Nintendo Switch console, valued at $339.99. If you're not a gamer, the deal is just as good, as you can opt instead for a free Samsung Galaxy Tab S10 FE 5G tablet or a 43-inch Samsung Q7F 4K TV, valued at $399.99. But you need to act fast: this deal ends tomorrow, and supplies are limited.

Signing up for an eligible plan will also get you a five-year price lock guarantee on your service, but you'll need to commit. According to the fine print, new customers must keep their account active for 14 days. If you cancel your internet service within 180 days, Verizon will charge you the full retail value of the promotional device (whichever one you chose).

Mashable Deals Be the first to know! Get editor selected deals texted right to your phone! Get editor selected deals texted right to your phone! Loading... Sign Me Up By signing up, you agree to receive recurring automated SMS marketing messages from Mashable Deals at the number provided. Msg and data rates may apply. Up to 2 messages/day. Reply STOP to opt out, HELP for help. Consent is not a condition of purchase. See our Privacy Policy and Terms of Use. Thanks for signing up!

Google’s Nano Banana has been a hit with AI image generation fans since its release in early August. Even the free version comes with tons of ways to create images, and it’s also pretty easy to use. Google is keeping the hype train rolling with its release of Nano Banana Pro, which is built on Gemini 3. 

The purpose of Nano Banana Pro is the Pro part. Google says that Gemini 3 and Nano Banana Pro are capable of making or editing images with “studio-quality levels of precision and control.” In layman's terms, it’s basically going to do what Nano Banana already does, but better since it’s powered by Gemini 3 instead of Google’s Gemini 2.5 Flash Image model like the regular Nano Banana. 

In theory, that should result in more realistic images with better, more accurate reasoning. Google has also wasted no time getting Nano Banana Pro integrated into other services, like Adobe’s Firefly, which gives users more avenues to try out the latest update. 

If you’re curious what it is capable of, here are some fun prompts you can try that will make full use of Nano Banana Pro. 

Credit: Google Infographics

Both Google and Adobe’s blog posts for Nano Banana Pro leaned heavily on the fact that the new AI can make infographics with cleaner text. It’s also one of the easiest ones to make. You pop an image into Nano Banana Pro and ask it to make one for you. In the example above, the input image was a basic photo of the String of Turtles houseplant. For the prompt, Google says it used “create an infographic about this plant focusing on interesting information.”

Nano Banana Pro then correctly identifies the plant as a String of Turtles house plant, and creates an infographic about its growth habits, care, origin, and leaf pattern. This is actually really cool because if Nano Banana Pro can identify the image on its own, you don’t need to give it complicated instructions. It’ll go out and find the thing and make an infographic for it. 

This is useful for a variety of purposes, but the one that most quickly came to mind was for a school project. Students using it for this purpose should double check to make sure the information is accurate, of course, but the image Nano Banana Pro spit out is better than anything I could’ve drawn in middle school.

Credit: Google Storyboards

To be honest, I would have never thought of storyboards on my own, but it’s a really neat concept. Like the infographics, you feed Nano Banana Pro an image and ask it to create a storyboard. Per Google, the prompt for the above image was “create a storyboard for this scene.” The AI’s image included an establishing shot, a medium shot, a close-up, and a POV shot.

Storyboarding is a job, and some people spend their whole careers doing it. The thing is, for a very small film crew, a beginner, or, again, a college student, hiring a storyboarder may not be in the cards, and doing it manually is kind of a pain. So, for low-end professional use cases, having AI help you with a storyboard isn’t a bad idea. You can also change up the prompt to have it make storyboards with different shots or ideas.

Credit: Google Combining loads of things into a single image

A fun thing you can already do with image generators is combine multiple objects into a single image. I’ve done it, and it’s pretty easy to do, but there are some limitations. I once inserted myself into a picture with my dog and had to try a few times to get things right. One of the things Google says that Nano Banana Pro does well is adding a bunch of elements into a single image while maintaining consistency. 

In the above example, Google fed Nano Banana Pro 14 unique fluffy characters and asked Nano Banana to sit them all on a couch, facing a TV. Nano Banana Pro not only did that, but made the fuzzy characters consistent as it stuffed them all into the same scene. The prompt for this is kind of ridiculous, so we’ll post it in its entirety below before moving on. 

“A medium shot of the 14 fluffy characters sitting squeezed together side-by-side on a worn beige fabric sofa and on the floor. They are all facing forwards, watching a vintage, wooden-boxed television set placed on a low wooden table in front of the sofa. The room is dimly lit, with warm light from a window on the left and a glow from the TV illuminating the creatures’ faces and fluffy textures. The background is a cozy, slightly cluttered living room with a braided rug, a bookshelf with old books, and rustic kitchen elements in the background. The overall atmosphere is warm, cozy, and amused.”

Credit: Google Photorealistic scenes

I’ve always admired AI’s ability to make realistic stuff, some of which is so good that it’s nearly indistinguishable from real life. It turns out that Nano Banana Pro is one of the best AI image generators for this. Google has a guide for making realistic images with some prompt ideas that you can try. They can get around medium length, so it’s a good intermediate prompt to try yourself. 

Per Google, the key here is detail. When generating a prompt, you want to mention everything you can, like lighting, camera angle, and even camera lens type if you’re knowledgeable about that kind of thing. Describing the subject as best as you can also increase the odds of getting something that you want. Don’t forget the background as well. As you can see with the example above, the background really helps bring the whole thing together. 

Credit: Google Translate words on an image to another language

Considering that half the words I’ve ever tried to have AI include in an image have been total gibberish, the idea that Nano Banana Pro can take words off of an input image, translate them, and put them back to be rather impressive. This can be useful for a variety of things, but I imagine businesses would get the best use by having their products translated into another language for a social media post aimed at a different region. 

The prompt for this one is pretty simple too. Per Google, the above image was put in and for the prompt, it’s simple, “translate all the English text on the three yellow and blue cans into Korean, while keeping everything else the same.” You can swap out that noun for basically any other noun and it looks like it’ll work the same way. Very cool. 

Credit: Google Change any doodle into a product

Many of us have been in a position where we draw or create something cool and then wondered what it’d be like if it were on a t-shirt or a coffee mug. Nano Banana Pro lets you do this pretty easily. The input image is your cool little design. You can get by just having Nano Banana Pro slap it onto a t-shirt, and call it a day. 

However, you can also add in tons of details to get a very specific product. In the above example, Google’s prompt (which is very long) includes telling the AI to use 1960s and 1970s aesthetics with a color palette that “reinforces the vintage feel” and has an effect of “whimsical nostalgia and clever graphic design.” Like many of the others, the images will get better the more detail you add, so the world is your oyster if you want to see your design on a piece of clothing. 

AI for days

On top of the above prompts, all of the usual Nano Banana stuff works. Viral prompts like turning yourself into a claymation character a la Robot Chicken or combining two elements in two photographs together still work as intended. Really, what you’re getting with Nano Banana Pro is refinement, with more power and overhead to fix many common problems (like words) that users may have struggled with while using regular Nano Banana. In short, it’s not a new product, just a better product.

Before Facebook exploded in popularity, it wasn’t super common to use your real, legal name on social media. Now, Meta is taking a tiny step in the opposite direction and attempting to make Facebook Groups a bit more like forums.

Pushing the frontiers of computer-use agents with an open-weight, ultra-compact model, optimized for real-world web tasks

In 2024, Microsoft introduced small language models (SLMs) to customers, starting with the release of Phi (opens in new tab) models on Microsoft Foundry (opens in new tab), as well as deploying Phi Silica (opens in new tab) on Copilot+ PCs powered by Windows 11. Today, we are pleased to announce Fara-7B, our first agentic SLM designed specifically for computer use.

Unlike traditional chat models that generate text-based responses, Computer Use Agent (CUA) models like Fara-7B leverage computer interfaces, such as a mouse and keyboard, to complete tasks on behalf of users. With only 7 billion parameters, Fara-7B achieves state-of-the-art performance within its size class and is competitive with larger, more resource-intensive agentic systems that depend on prompting multiple large models. Fara-7B’s small size now makes it possible to run CUA models directly on devices. This results in reduced latency and improved privacy, as user data remains local.

Fara-7B is an experimental release, designed to invite hands-on exploration and feedback from the community. Users can build and test agentic experiences beyond pure research—automating everyday web tasks like filling out forms, searching for information, booking travel, or managing accounts. We recommend running Fara-7B in a sandboxed environment, monitoring its execution, and avoiding sensitive data or high-risk domains. Responsible use is essential as the model continues to evolve.

Fara-7B operates by visually perceiving a webpage and takes actions like scrolling, typing, and clicking on directly predicted coordinates. It does not rely on separate models to parse the screen, nor on any additional information like accessibility trees, and thus uses the same modalities as humans to interact with the computer. To train Fara-7B, we developed a novel synthetic data generation pipeline for multi-step web tasks, building on our prior work (AgentInstruct). This data generation pipeline draws from real web pages and tasks sourced from human users.

Video 1: A demo of a shopping scenario with Fara-7B through Magentic-UI. Fara-7B is asked to purchase an X-Box Spongebob controller. Fara-7B goes on to complete this task, but while doing so, also stops at every Critical Point to get input and approval from the user before proceeding. Video 2: A demo of Fara-7B finding relevant information online and summarizing it through Magentic-UI. We ask Fara-7B to find and summarize the latest three issues on Github Microsoft/Magentic-UI. Video 3: A demo of how Fara-7B can use different tools to find relevant information and analyze it through Magentic-UI. We ask Fara-7B to find driving time between two places, and suggest a cheese place near the location. Fara-7B uses Bing Maps to find Driving time, and Bing search to find relevant information.

Fara-7B exhibits strong performance compared to existing models across a diverse set of benchmarks. This includes both existing benchmarks as well as new evaluations we are releasing which cover useful task segments that are underrepresented in common benchmarks, such as finding job postings and comparing prices across retailers. While Fara-7B demonstrates strong benchmark results, even against much larger models, it shares many of their limitations, including challenges with accuracy on more complex tasks, mistakes in following instructions, and susceptibility to hallucinations. These are active areas of research, and we’re committed to ongoing improvements as we learn from real-world use.

Fara-7B is now available on Microsoft Foundry (opens in new tab) and Hugging Face (opens in new tab) under an MIT license and is integrated with Magentic-UI, a research prototype from Microsoft Research AI Frontiers (opens in new tab). We are also sharing a quantized and silicon-optimized version of Fara-7B, which will be available to install and run on Copilot+ PCs powered by Windows 11, for turnkey experimentation. The community can simply download the pre-optimized model and run it in their environment.

By making Fara-7B open-weight, we aim to lower the barrier to experimenting with and improving CUA technology for automating routine web tasks, such as searching for information, shopping, and booking reservations.

Figure 1: Comparing WebVoyager accuracy and cost of Fara-7B to other computer use agents (CUAs) or agents that prompt LLMs with accessibility trees (SoM Agent w/ Ax Tree). Cost is computed by multiplying the average number of input and output tokens each model consumes by price per token. Both Fara-7B and UI-TARS-1.5-7B are based on Qwen-2.5-VL-7B, for which the lowest inference price from https://openrouter.ai/  is \(0.2/\)0.2 per 1M input/output tokens. Even though both models are priced equally, Fara-7B is more efficient, completing tasks with only ~16 steps on average compared to ~41 for UI-TARS-1.5-7B. OpenAI computer-use-preview accessed November 2025 via the Responses API. Developing Fara-7B CUA multi-agent synthetic data generation

A key bottleneck for building CUA models is a lack of large-scale, high-quality computer interaction data. Collecting such data with human annotators is prohibitively expensive as a single CUA task can involve dozens of steps, each of which needs to be annotated. Our data generation pipeline (Figure 2) avoids manual annotation and instead relies on scalable synthetic data sourced from publicly available websites and custom task prompts. We build this pipeline on top of the Magentic-One framework, and it involves three main stages: 

Figure 2: Data Generation workflow from proposing tasks from various seeds like URLs to solving those tasks with the Magentic-One multi-agent framework to generate demonstrations for training, and finally verifiying/filtering completed trajectories

Task Proposal. We generate a broad set of synthetic tasks that mirror common user activities on the web. To ensure coverage and diversity, tasks are “seeded” by a web index of public URLs classified into various categories e.g., shopping, travel, restaurants, etc. This enables task generation targeting a particular skill, like “book 2 tickets to see the Downton Abbey Grand Finale at AMC Union Square, NYC.” from a URL like this (opens in new tab) classified as “movies”.  As another strategy, we devised a way to generate tasks from randomly sampled URLs. Each task starts with a general prompt and is iteratively refined as an LLM agent explores the website and gathers more information about it. We are releasing a held-out subset of these tasks as a benchmark (“WebTailBench”), described in the Evaluation section below. 

Task Solving. Once synthetic tasks are generated, a multi-agent system built on Magentic-One attempts to complete them to generate demonstrations for supervised finetuning. The multi-agent system uses an Orchestrator agent to create a plan and direct a WebSurfer agent to take browser actions and reports results. The Orchestrator monitors progress, updating plans as needed, and can end tasks or engage a UserSimulator agent if user input is required, allowing for multi-turn completion. Each task and corresponding sequence of observations, actions, and agent thoughts forms a “trajectory”.

Trajectory Verification. Before using any tasks for training, three verifier agents evaluate if a task was “successful”: The Alignment Verifier checks if the trajectory of actions match the task’s intent; the Rubric Verifier defines completion criteria and scores the trajectory against them; and the Multimodal Verifier reviews screenshots and responses to confirm visual evidence supports successful completion. Trajectories failing these standards are removed.

We ultimately train this version of Fara-7B on a dataset of 145,000 trajectories consisting of 1 million steps covering diverse websites, task types, and difficulty levels. Additionally, we include training data for several auxiliary tasks, including grounding for accurate UI element localization, captioning, and visual question answering.

Training Fara-7B

Using one compute use model is easier than a multi-agent system, particularly when it comes to deployment. Therefore, we distill the complexities of our multi-agent solving system into a single model that can execute tasks. Fara-7B is a proof-of-concept that small models can effectively learn from complex, multi-agent systems with lots of bells and whistles.

As shown in Figure 3, Fara-7B is trained to execute user tasks by perceiving only browser window screenshots (without relying on accessibility trees), and predicting single-step actions. For each step, the context used to make its prediction contains all user messages, the complete action history, and the latest three screenshots.

In its prediction, Fara-7B outputs a reasoning message (“thinking” about the next action) followed by a tool call. The available tools include standard Playwright (opens in new tab) mouse and keyboard actions, such as click(x,y) and type(), and browser-specific macro-actions like web_search() and visit_url().

Fara-7B uses Qwen2.5-VL-7B (opens in new tab) as its base model due to its strong performance on grounding tasks and its ability to support long contexts (up to 128k tokens). We linearize the solving pipeline’s trajectories into a sequence of “observe-think-act” steps that are suitable for training with supervised finetuning loss. We did not use reinforcement learning to achieve the results we report below.

Figure 3: Operation of Fara-7B as a standalone, native computer use agent running on-device. Because Fara-7B is small, and none of its context needs to leave your personal device, it paves the way for personal and private agentic computing Evaluations

We evaluate Fara-7B and comparable baselines on canonical public benchmarks including WebVoyager (opens in new tab), Online-Mind2Web (opens in new tab), and Deepshop (opens in new tab), as well as a new benchmark we developed named WebTailBench, specifically focusing on 11 real-world task types underrepresented or missing in existing benchmarks like booking movie/event tickets, restaurant reservations, comparing prices across retailers, applying for jobs, finding real estate, and more complex multi-step tasks.

Evaluation of web agents can be tricky because the web is constantly changing, and many websites even block detected bots, which is why we developed a test harness that relies on Browserbase (opens in new tab) to standardize how browser sessions are managed. In Table 1 below, we report a notion of task success rate (%) defined by each benchmark’s official LLM-as-judge evaluator; WebTailBench success is computed using the same Task Verification pipeline that filtered our training data. We find that Fara-7B is state-of-the-art, even outperforming native computer use agents like UI-TARS-1.5-7B, or much larger models like GPT-4o prompted to act like a computer use agent with Set-Of-Marks (opens in new tab) (SoM Agent). 

WebVoyagerOnline-Mind2WebDeepShopWebTailBench  SoM Agents SoM Agent (GPT-4o) 65.1 34.6 16.0 30.0 GLM-4.1V-9B-Thinking 66.8  33.9 32.0 22.4 Computer Use Models OpenAI computer-use-preview  70.9 42.9 24.7 25.7 UI-TARS-1.5-7B 66.4  31.3 11.6 19.5 Fara-7B 73.5 34.1 26.2 38.4 Table 1: Performance comparison across four web benchmarks: WebVoyager, Online-Mind2Web, DeepShop, and our newly introduced WebTailBench. Results are reported as Task Succes Rate / Accuracy (%) and are averaged over 3 runs. OpenAI computer-use-preview accessed November 2025 via the Responses API.

In Figure 1, we expand on the Webvoyager results by giving each model up to three chances to complete a task, and report “pass@K”. We also consider on the x-axis the cost of running each model if one were to pay market rates for input/output tokens consumed. Fara-7B breaks ground on a new pareto frontier, showing that on-device computer use agents are approaching the capabilities of frontier models.

We partnered with a trusted external group, Browserbase, to independently evaluate Fara-7B using human annotators. The model achieved 62% on WebVoyager (see detailed reports in Browserbase blog here (opens in new tab)). These results were generated in the same environment with identical settings and human verification of each task, making them directly comparable. Note that Browserbase’s standard WebVoyager scores do not use retries when environment errors occur; the results referenced here include retries and should not be compared directly to the non-retry scores. Going forward, we are collaborating with Browserbase to host WebTailBench human evaluations to help the community build reliable and reproducible assessments for computer use agents. 

Safety

Agents capable of operating computers present challenges distinct from chat-only models, including new outlets of user misuse, model misbehavior, and unintended consequences of actions, and external risks like prompt injections or online scams. CUAs take action with real-world consequences, so ensuring robust safety measures is essential to their responsible deployment. Transparency and user control sit at the core of Fara-7B’s design. Although we have incorporated several safety measures, Fara-7B remains a research preview, and we continue to advance our approach to safety for computer use agents, an active area of work across the entire AI community. 

Fara-7B processes browser screenshots, user task instructions, and a history of actions taken during each session and collects only what is necessary to complete the user’s requested task. No additional site data—such as accessibility trees or external scaffolding—is accessed; Fara-7B interacts with the computer in the same way a human would, relying solely on what is visible on the screen.

All actions taken by the agent are logged and auditable, allowing users to review and monitor every step.  For added safety, Fara‑7B is intended to run in sandboxed environments, giving users full oversight and the ability to intervene or halt actions at any time. These safeguards ensure that privacy, transparency, and user control remain at the core of every interaction.

To address misuse, we trained Fara-7B on a mixture of public safety data and internally generated tasks that it ought to refuse based on Microsoft’s Responsible AI Policy. We evaluated Fara-7B’s ability to refuse harmful tasks on WebTailBench-Refusals which consists of 111 red-teaming tasks showing a high refusal rate of 82%. The model also underwent Microsoft’s rigorous red teaming process, where we focused on the model rejecting harmful tasks and risky tasks, such as harmful content, jailbreaking attempts, ungrounded responses, and prompt injections. For further details, check out our technical report (opens in new tab).

To mitigate the risk of Fara-7B taking unintended actions, all of Fara-7B’s training data enforces both recognizing and stopping at “Critical Points” when executing a task. A Critical Point (see Operator System Card (opens in new tab)) is any situation that requires the user’s personal data or consent before engaging in a transaction or irreversible action like sending an email. Upon reaching a Critical Point, Fara-7B should respond by informing the user it cannot proceed without their consent.

For guidance on how to use our model safely, and the security considerations to be mindful of when using our model, please refer to our Model card (opens in new tab).

How to use

Fara-7B is available on  (opens in new tab)Microsoft Foundry  (opens in new tab)and  (opens in new tab)Hugging Face (opens in new tab). We are also releasing the implementation of Fara-7B in Magentic-UI, so that users can try it in a contained environment through the inference code provided. Additionally, users can download the model for Copilot+ PCs powered by Windows 11 from the AI Toolkit in VSCode and run it all on-device, taking advantage of NPU hardware acceleration.  

Looking forward

Our current release is an experimental CUA model that achieves state-of-the-art results for its size, purely using supervised fine-tuning. We believe even stronger CUA models capable of running on-device are possible through improved multimodal base models and through Reinforcement Learning on live and sandboxed environments. These early days are about learning from the community and driving real-world experimentation to shape what comes next. If you’d like to join us and help shape the future of SLMs, please apply for open roles. 

Acknowledgements: 

We thank Gustavo de Rosa, Adam Fourney, Michael Harrison, Rafah Hosn, Neel Joshi, Ece Kamar, John Langford, Maya Murad, Sidhartha Sen, Pratyusha Sharma, and Lili Wu for their valuable help, insightful discussions, and continued support throughout this work. 

We also thank Pashmina Cameron, Karthik Vijayan, Vicente Rivera, Chris Dern, Sayan Shaw, Sunghoon Choi, Andrey Rybalchenko, and Vivek Pradeep for their efforts in making the model available on Copilot+ PCs through the AI Toolkit.

Opens in a new tab

The post Fara-7B: An Efficient Agentic Model for Computer Use appeared first on Microsoft Research.

Pushing the frontiers of computer-use agents with an open-weight, ultra-compact model, optimized for real-world web tasks

In 2024, Microsoft introduced small language models (SLMs) to customers, starting with the release of Phi (opens in new tab) models on Microsoft Foundry (opens in new tab), as well as deploying Phi Silica (opens in new tab) on Copilot+ PCs powered by Windows 11. Today, we are pleased to announce Fara-7B, our first agentic SLM designed specifically for computer use.

Unlike traditional chat models that generate text-based responses, Computer Use Agent (CUA) models like Fara-7B leverage computer interfaces, such as a mouse and keyboard, to complete tasks on behalf of users. With only 7 billion parameters, Fara-7B achieves state-of-the-art performance within its size class and is competitive with larger, more resource-intensive agentic systems that depend on prompting multiple large models. Fara-7B’s small size now makes it possible to run CUA models directly on devices. This results in reduced latency and improved privacy, as user data remains local.

Fara-7B is an experimental release, designed to invite hands-on exploration and feedback from the community. Users can build and test agentic experiences beyond pure research—automating everyday web tasks like filling out forms, searching for information, booking travel, or managing accounts. We recommend running Fara-7B in a sandboxed environment, monitoring its execution, and avoiding sensitive data or high-risk domains. Responsible use is essential as the model continues to evolve.

Fara-7B operates by visually perceiving a webpage and takes actions like scrolling, typing, and clicking on directly predicted coordinates. It does not rely on separate models to parse the screen, nor on any additional information like accessibility trees, and thus uses the same modalities as humans to interact with the computer. To train Fara-7B, we developed a novel synthetic data generation pipeline for multi-step web tasks, building on our prior work (AgentInstruct). This data generation pipeline draws from real web pages and tasks sourced from human users.

Video 1: A demo of a shopping scenario with Fara-7B through Magentic-UI. Fara-7B is asked to purchase an X-Box Spongebob controller. Fara-7B goes on to complete this task, but while doing so, also stops at every Critical Point to get input and approval from the user before proceeding. Video 2: A demo of Fara-7B finding relevant information online and summarizing it through Magentic-UI. We ask Fara-7B to find and summarize the latest three issues on Github Microsoft/Magentic-UI. Video 3: A demo of how Fara-7B can use different tools to find relevant information and analyze it through Magentic-UI. We ask Fara-7B to find driving time between two places, and suggest a cheese place near the location. Fara-7B uses Bing Maps to find Driving time, and Bing search to find relevant information.

Fara-7B exhibits strong performance compared to existing models across a diverse set of benchmarks. This includes both existing benchmarks as well as new evaluations we are releasing which cover useful task segments that are underrepresented in common benchmarks, such as finding job postings and comparing prices across retailers. While Fara-7B demonstrates strong benchmark results, even against much larger models, it shares many of their limitations, including challenges with accuracy on more complex tasks, mistakes in following instructions, and susceptibility to hallucinations. These are active areas of research, and we’re committed to ongoing improvements as we learn from real-world use.

Fara-7B is now available on Microsoft Foundry (opens in new tab) and Hugging Face (opens in new tab) under an MIT license and is integrated with Magentic-UI, a research prototype from Microsoft Research AI Frontiers (opens in new tab). We are also sharing a quantized and silicon-optimized version of Fara-7B, which will be available to install and run on Copilot+ PCs powered by Windows 11, for turnkey experimentation. The community can simply download the pre-optimized model and run it in their environment.

By making Fara-7B open-weight, we aim to lower the barrier to experimenting with and improving CUA technology for automating routine web tasks, such as searching for information, shopping, and booking reservations.

Figure 1: Comparing WebVoyager accuracy and cost of Fara-7B to other computer use agents (CUAs) or agents that prompt LLMs with accessibility trees (SoM Agent w/ Ax Tree). Cost is computed by multiplying the average number of input and output tokens each model consumes by price per token. Both Fara-7B and UI-TARS-1.5-7B are based on Qwen-2.5-VL-7B, for which the lowest inference price from https://openrouter.ai/  is \(0.2/\)0.2 per 1M input/output tokens. Even though both models are priced equally, Fara-7B is more efficient, completing tasks with only ~16 steps on average compared to ~41 for UI-TARS-1.5-7B. OpenAI computer-use-preview accessed November 2025 via the Responses API. Developing Fara-7B CUA multi-agent synthetic data generation

A key bottleneck for building CUA models is a lack of large-scale, high-quality computer interaction data. Collecting such data with human annotators is prohibitively expensive as a single CUA task can involve dozens of steps, each of which needs to be annotated. Our data generation pipeline (Figure 2) avoids manual annotation and instead relies on scalable synthetic data sourced from publicly available websites and custom task prompts. We build this pipeline on top of the Magentic-One framework, and it involves three main stages: 

Figure 2: Data Generation workflow from proposing tasks from various seeds like URLs to solving those tasks with the Magentic-One multi-agent framework to generate demonstrations for training, and finally verifiying/filtering completed trajectories

Task Proposal. We generate a broad set of synthetic tasks that mirror common user activities on the web. To ensure coverage and diversity, tasks are “seeded” by a web index of public URLs classified into various categories e.g., shopping, travel, restaurants, etc. This enables task generation targeting a particular skill, like “book 2 tickets to see the Downton Abbey Grand Finale at AMC Union Square, NYC.” from a URL like this (opens in new tab) classified as “movies”.  As another strategy, we devised a way to generate tasks from randomly sampled URLs. Each task starts with a general prompt and is iteratively refined as an LLM agent explores the website and gathers more information about it. We are releasing a held-out subset of these tasks as a benchmark (“WebTailBench”), described in the Evaluation section below. 

Task Solving. Once synthetic tasks are generated, a multi-agent system built on Magentic-One attempts to complete them to generate demonstrations for supervised finetuning. The multi-agent system uses an Orchestrator agent to create a plan and direct a WebSurfer agent to take browser actions and reports results. The Orchestrator monitors progress, updating plans as needed, and can end tasks or engage a UserSimulator agent if user input is required, allowing for multi-turn completion. Each task and corresponding sequence of observations, actions, and agent thoughts forms a “trajectory”.

Trajectory Verification. Before using any tasks for training, three verifier agents evaluate if a task was “successful”: The Alignment Verifier checks if the trajectory of actions match the task’s intent; the Rubric Verifier defines completion criteria and scores the trajectory against them; and the Multimodal Verifier reviews screenshots and responses to confirm visual evidence supports successful completion. Trajectories failing these standards are removed.

We ultimately train this version of Fara-7B on a dataset of 145,000 trajectories consisting of 1 million steps covering diverse websites, task types, and difficulty levels. Additionally, we include training data for several auxiliary tasks, including grounding for accurate UI element localization, captioning, and visual question answering.

Training Fara-7B

Using one compute use model is easier than a multi-agent system, particularly when it comes to deployment. Therefore, we distill the complexities of our multi-agent solving system into a single model that can execute tasks. Fara-7B is a proof-of-concept that small models can effectively learn from complex, multi-agent systems with lots of bells and whistles.

As shown in Figure 3, Fara-7B is trained to execute user tasks by perceiving only browser window screenshots (without relying on accessibility trees), and predicting single-step actions. For each step, the context used to make its prediction contains all user messages, the complete action history, and the latest three screenshots.

In its prediction, Fara-7B outputs a reasoning message (“thinking” about the next action) followed by a tool call. The available tools include standard Playwright (opens in new tab) mouse and keyboard actions, such as click(x,y) and type(), and browser-specific macro-actions like web_search() and visit_url().

Fara-7B uses Qwen2.5-VL-7B (opens in new tab) as its base model due to its strong performance on grounding tasks and its ability to support long contexts (up to 128k tokens). We linearize the solving pipeline’s trajectories into a sequence of “observe-think-act” steps that are suitable for training with supervised finetuning loss. We did not use reinforcement learning to achieve the results we report below.

Figure 3: Operation of Fara-7B as a standalone, native computer use agent running on-device. Because Fara-7B is small, and none of its context needs to leave your personal device, it paves the way for personal and private agentic computing Evaluations

We evaluate Fara-7B and comparable baselines on canonical public benchmarks including WebVoyager (opens in new tab), Online-Mind2Web (opens in new tab), and Deepshop (opens in new tab), as well as a new benchmark we developed named WebTailBench, specifically focusing on 11 real-world task types underrepresented or missing in existing benchmarks like booking movie/event tickets, restaurant reservations, comparing prices across retailers, applying for jobs, finding real estate, and more complex multi-step tasks.

Evaluation of web agents can be tricky because the web is constantly changing, and many websites even block detected bots, which is why we developed a test harness that relies on BrowserBase (opens in new tab) to standardize how browser sessions are managed. In Table 1 below, we report a notion of task success rate (%) defined by each benchmark’s official LLM-as-judge evaluator; WebTailBench success is computed using the same Task Verification pipeline that filtered our training data. We find that Fara-7B is state-of-the-art, even outperforming native computer use agents like UI-TARS-1.5-7B, or much larger models like GPT-4o prompted to act like a computer use agent with Set-Of-Marks (opens in new tab) (SoM Agent). 

WebVoyagerOnline-Mind2WebDeepShopWebTailBench  SoM Agents SoM Agent (GPT-4o) 65.1 34.6 16.0 30.0 GLM-4.1V-9B-Thinking 66.8  33.9 32.0 22.4 Computer Use Models OpenAI computer-use-preview  70.9 42.9 24.7 25.7 UI-TARS-1.5-7B 66.4  31.3 11.6 19.5 Fara-7B 73.5 34.1 26.2 38.4 Table 1: Performance comparison across four web benchmarks: WebVoyager, Online-Mind2Web, DeepShop, and our newly introduced WebTailBench. Results are reported as Task Succes Rate / Accuracy (%) and are averaged over 3 runs. OpenAI computer-use-preview accessed November 2025 via the Responses API.

In Figure 1, we expand on the Webvoyager results by giving each model up to three chances to complete a task, and report “pass@K”. We also consider on the x-axis the cost of running each model if one were to pay market rates for input/output tokens consumed. Fara-7B breaks ground on a new pareto frontier, showing that on-device computer use agents are approaching the capabilities of frontier models.

We partnered with a trusted external group, Browserbase, to independently evaluate Fara-7B using human annotators. The model achieved 62% on WebVoyager (see detailed reports in Browserbase blog here (opens in new tab)). These results were generated in the same environment with identical settings and human verification of each task, making them directly comparable. Note that Browserbase’s standard WebVoyager scores do not use retries when environment errors occur; the results referenced here include retries and should not be compared directly to the non-retry scores. Going forward, we are collaborating with Browserbase to host WebTailBench human evaluations to help the community build reliable and reproducible assessments for computer use agents. 

Safety

Agents capable of operating computers present challenges distinct from chat-only models, including new outlets of user misuse, model misbehavior, and unintended consequences of actions, and external risks like prompt injections or online scams. CUAs take action with real-world consequences, so ensuring robust safety measures is essential to their responsible deployment. Transparency and user control sit at the core of Fara-7B’s design. Although we have incorporated several safety measures, Fara-7B remains a research preview, and we continue to advance our approach to safety for computer use agents, an active area of work across the entire AI community. 

Fara-7B processes browser screenshots, user task instructions, and a history of actions taken during each session and collects only what is necessary to complete the user’s requested task. No additional site data—such as accessibility trees or external scaffolding—is accessed; Fara-7B interacts with the computer in the same way a human would, relying solely on what is visible on the screen.

All actions taken by the agent are logged and auditable, allowing users to review and monitor every step.  For added safety, Fara‑7B is intended to run in sandboxed environments, giving users full oversight and the ability to intervene or halt actions at any time. These safeguards ensure that privacy, transparency, and user control remain at the core of every interaction.

To address misuse, we trained Fara-7B on a mixture of public safety data and internally generated tasks that it ought to refuse based on Microsoft’s Responsible AI Policy. We evaluated Fara-7B’s ability to refuse harmful tasks on WebTailBench-Refusals which consists of 111 red-teaming tasks showing a high refusal rate of 82%. The model also underwent Microsoft’s rigorous red teaming process, where we focused on the model rejecting harmful tasks and risky tasks, such as harmful content, jailbreaking attempts, ungrounded responses, and prompt injections. For further details, check out our technical report (opens in new tab).

To mitigate the risk of Fara-7B taking unintended actions, all of Fara-7B’s training data enforces both recognizing and stopping at “Critical Points” when executing a task. A Critical Point (see Operator System Card (opens in new tab)) is any situation that requires the user’s personal data or consent before engaging in a transaction or irreversible action like sending an email. Upon reaching a Critical Point, Fara-7B should respond by informing the user it cannot proceed without their consent.

For guidance on how to use our model safely, and the security considerations to be mindful of when using our model, please refer to our Model card (opens in new tab).

How to use

Fara-7B is available on  (opens in new tab)Microsoft Foundry  (opens in new tab)and  (opens in new tab)Hugging Face (opens in new tab). We are also releasing the implementation of Fara-7B in Magentic-UI, so that users can try it in a contained environment through the inference code provided. Additionally, users can download the model for Copilot+ PCs powered by Windows 11 from the AI Toolkit in VSCode and run it all on-device, taking advantage of NPU hardware acceleration.  

Looking forward

Our current release is an experimental CUA model that achieves state-of-the-art results for its size, purely using supervised fine-tuning. We believe even stronger CUA models capable of running on-device are possible through improved multimodal base models and through Reinforcement Learning on live and sandboxed environments. These early days are about learning from the community and driving real-world experimentation to shape what comes next. If you’d like to join us and help shape the future of SLMs, please apply for open roles. 

Acknowledgements: 

We thank Gustavo de Rosa, Adam Fourney, Michael Harrison, Rafah Hosn, Neel Joshi, Ece Kamar, John Langford, Maya Murad, Sidhartha Sen, Pratyusha Sharma, and Lili Wu for their valuable help, insightful discussions, and continued support throughout this work. 

We also thank Pashmina Cameron, Karthik Vijayan, Vicente Rivera, Chris Dern, Sayan Shaw, Sunghoon Choi, Andrey Rybalchenko, and Vivek Pradeep for their efforts in making the model available on Copilot+ PCs through the AI Toolkit.

Opens in a new tab

The post Fara-7B: An Efficient Agentic Model for Computer Use appeared first on Microsoft Research.

SAVE $30: As of Nov. 24, the Kobo Libra Colour is now down to $199.99 ahead of Black Friday. That saves you $30 off its list price of $229.99, saving you 13% off. That brings the e-reader back down to the same Black Friday price we saw last year.

Kobo Libra Colour $199.99 at Amazon
$219.99 Save $20   Get Deal at Amazon Get Deal at Amazon

Here's the thing about e-readers: Amazon usually makes the best ones. I've tested a lot of e-readers and Kindles are fantastic, speedy little devices that makes reading on the go all the better. But that's not great news if you really hate Amazon. Luckily, there's an alternative: Kobo e-readers. While I prefer Kindles, Kobo e-readers still rock. In fact, I actually think if you're looking for a color e-reader, you should go for a Kobo. Even better, our favorite color e-reader, the Kobo Libra Colour, is now on sale ahead of Black Friday.

As of Nov. 24, the Kobo Libra Colour is down to $199.99, returning to its sale price we saw during Black Friday 2024. Normally, $229.99, this is a rare $30 discount on the Kobo Libra Colour. While the device is now 13%, this might not seem that impressive compared to the 20 to 30% off Kindles during Black Friday, it's still a great deal if you're looking for a non-Amazon e-reader.

The Kobo Libra Colour is one of our favorites here at Mashable. It's normally a bit of splurge at $229.99, however, it earns its price tag. Not only does it comes with color displays, it has writing capabilities, page-turning buttons, and an internal gyroscope so you can read from any angle. It won the top spot in our guide to the best color e-readers thanks to all those great features, as its an annotator's day dream. When paired with the Kobo Stylus 2, you can write notes in the margin, journal, and highlight passages on the Kobo Libra Colour.

But it's not the only color e-reader on sale. The Kindle Colorsoft is finally on sale and even cheaper than the Kobo Libra. It's normally $249.99 but during Black Friday, it's down to $169.99. While I think that's an incredible deal, I still say you should consider the Kobo Libra. Having tested both, the Libra Colour and Colorsoft are equally matched on resolution and color quality, but I think the Libra shines for its writing capability and better ergonomic design.

Plus, coming Dec. 2, is the Kobo Remote, which lets you turn the page on the device without having to hold it up. It's really going to be the 'it' accessory for cozy winter reading sessions.

If you're looking for an Amazon-free e-reader, shop the Kobo Libra Colour for just $199.99 and save $30 during Black Friday.

Best Black Friday travel deals as of Nov. 24: Best deal overall Get up to 40% off hotels when you book through the Tripadvisor app. save 40% Get Deal Best flight deal Get up to $450 off select flight and hotel packages when you book through JetBlue Vacations save $450 Get Deal Best mileage hack Earn 4X miles per dollar spent when you purchase $400 or more in Delta Gift Cards. save $400 Get Deal Best hotel deal Get up to 55% off stand-alone hotel rates and up to 70% off Fiesta Americana Travelty Collection (FATC)package deals. save 70% percent Get Deal Best hotel deal for families and fans Get up to 60% off tickets and up to 60% off Ticket + Room packages at Legoland Resort Florida. Plus, you can save up to $124 on Annual Passes. save 60% Get Deal

For most shoppers, Black Friday means snatching up the best deals on small kitchen appliances, robot vacuums, and tech before the prices skyrocket back to their pre-sale values. What many bargain hunters don’t realize, though, is that you can actually find tons of Black Friday discounts on travel if you know where to look.

SEE ALSO: Black Friday 2025: Live updates on the latest deals from Amazon, Target, Walmart, and more

Airlines, hotels, and even one-of-a-kind experiences typically go on sale right before the holiday season, so you can book, save, and travel later. Sure, there are blackout dates, and some deals may require advance booking or have other restrictions, but if you do it right, you can get a luxe vacay for a fraction of the cost.

Below, I’ve rounded up the best Black Friday travel steals to help you plan your next trip without breaking the bank.

Best travel deal overall Opens in a new window Credit: Tripadvisor Tripadvisor Black Friday deal Tripadvisor is offering up to 40% off hotels when you book through the app. You’ll also get a $50 welcome offer that’s good to use on experiences when you sign up for the free Tripadvisor Rewards program. Get Deal Why we like it

Starting tomorrow, Tripadvisor is offering up to 40% off hotels when you book through the app. You’ll also get a $50 welcome offer that’s good to use on experiences when you sign up for the free Tripadvisor Rewards program.

With every hotel and experience you book through the Tripadvisor app, you’ll get 5% “Trip Cash” back that you can use for future bookings. (It's basically free money you can immediately put toward your next trip.) You’ll also be able to book directly within the app, whether you’re booking a hotel for the night, dinner reservations, or an experience. Plus, Tripadvisor is one of the most legit travel apps around — all the reviews and photos are 100 percent user-generated, which means there’s no fake marketing attached, and you can make an informed decision based on real customer reviews.

The fine print: This Black Friday offer officially runs from November 25 through Dec. 1. If you download the app and sign up today (November 24), you’ll get a $30 off “Things to Do” offer instead of the $50 bonus. If you miss this main Black Friday/Cyber Monday window, keep an eye out for Travel Deals Tuesday on Dec.2, where they’ll host a separate sale with up to 40% off various hotels and domestic and international travel destinations.

Best flight deal Opens in a new window Credit: JetBlue JetBlue Black Friday deal Right now, you can get up to $450 off select flight and hotel packages when you book through JetBlue Vacations. Get Deal Why we like it

Right now, you can get up to $450 off select flight and hotel packages when you book through JetBlue Vacations.

JetBlue doesn’t get enough credit, if you ask me. They give you little cookies, their seats are larger (in my opinion), and sometimes, they have the little TVs. This package deal is huge because you only need a low deposit of $99 per person to lock in your trip. JetBlue also offers perks like free airport transfers (at select destinations), a Price Match Guarantee if you find a lower price elsewhere, and your first change is free. It’s the perfect way to secure a 2026 tropical getaway now.

The fine print: This deal is an early Black Friday offer and must be booked by Nov. 25 (tomorrow). Travel dates are valid from Jan. 1, 2026, through Oct. 15, 2026. The max $450 discount requires a minimum spend of $4,000 on specific packages like Atlantis Paradise Island (use code PRESALE450). Other packages offer $300 off with a minimum $3,000 spend (use code PRESALE300).

Best mileage hack Opens in a new window Credit: Delta Delta Black Friday deal Earn 4X miles per dollar spent when you purchase $400 or more in Delta Gift Cards. Get Deal Why we like it

Earn 4X miles per dollar spent when you purchase $400 or more in Delta Gift Cards.

This is a low-commitment hack for SkyMiles members who know they’ll be flying Delta next year. If you plan on spending $400 or more on Delta flights anyway, buying a gift card now is basically free miles that you can use toward future travel benefits. Just purchase the gift cards now, earn the miles, and use the card balance later when you're ready to book your actual trip.

The fine print: This offer is valid through Dec. 3, 2025. You'll need to purchase $400 or more in one transaction. Be patient: the bonus miles will be added to your account approximately 30 business days after the promotion ends.

Best hotel deal Opens in a new window Credit: Fiesta Americana Travelty Collection (FATC) Fiesta Americana Travelty Fiesta Americana Travelty Collection (FATC) is offering up to 55% off stand-alone hotel rates and up to 70% off package deals. Get Deal Why we like it

Fiesta Americana Travelty Collection (FATC) is offering up to 55% off stand-alone hotel rates and up to 70% off package deals.

The discount applies across this entire luxury portfolio, which includes resorts like Live Aqua and Grand Fiesta Americana. The savings potential is massive, especially since the majority of their properties are all-inclusive. (This means your biggest cost is getting there, and you just secured your biggest saving!) The best part is the travel window: you can book today and travel all the way up to Dec. 31, 2026. That kind of future flexibility makes this Black Friday purchase a no-brainer.

The fine print: This deal is available to book through Dec. 3. Like most great travel sales, blackout dates apply, so be sure to check those before finalizing your dates.

Best hotel deal for families and fans Opens in a new window Credit: Legoland Legoland Florida Resort Black Friday deal Legoland Florida Resort is offering up to 60% off tickets and up to 60% off Ticket + Room packages. Plus, you can save up to $124 on Annual Passes. Get Deal Why we like it

Legoland Florida Resort is offering up to 60% off tickets and up to 60% off Ticket + Room packages. Plus, you can save up to $124 on Annual Passes.

This is the true "gift" deal of the season. Not only are you saving up to 60% on tickets and packages, but this sale is also the perfect time to commit to a pass if you live nearby or plan on repeat visits, as you can save up to $120 on an Annual Pass. The savings cover tickets to Legoland Florida, Sea Life Florida, and the Peppa Pig Theme Park, making it perfect for families with kids of different ages.

The fine print: This deal runs through Dec. 2, 2025. The travel window for packages is valid for stays from Nov. 30, 2025, through May 21, 2026 (blackout dates apply).

Even more Black Friday travel deals

• Saint Lucia Hotels: Get up to 65% off across two dozen hotels and resorts, including Sandals and Windjammer Landing. Dates: Deals cover Black Friday, Cyber Monday, and Travel Tuesday.

• Priceline: Offers up to $650 off bundles (hotel + flight or car) with code BUNDLEBIG on Travel Tuesday. Other daily deals include up to 60% off packages, up to 30% off hotels, and up to $2,000 cruise cash. Dates: Sale runs Nov. 10 – Dec. 2, 2025.

• Club Med: Get up to 50% off select all-inclusive resorts (including Cancún and Punta Cana), plus up to $500 instant savings per person/week. Dates: Book Nov. 18 – Dec. 1, 2025. Travel Dec. 6, 2025 – Jun. 19, 2026.

• Legoland California Resort: Get up to 50% off Ticket + Room packages. Also, save up to $120 on Annual Passes. Dates: Book Nov. 18 – Dec. 2, 2025.

• The Mining Exchange Hotel (Colorado Springs, CO): Get up to 45% off stays. Dates: Book Nov. 18 – Dec. 1, 2025. Travel Nov. 18 – Dec. 30, 2025.

• Amsterdam Manor Beach Resort (Aruba): Get 40% off rooms plus a complimentary welcome drink. Dates: Book Nov. 18 – Dec. 3, 2025. Travel Nov. 18, 2025 – Dec. 30, 2026 (Min 4-night stay).

• Davenport Hotel Collection (Spokane, WA): Get 40% off upgraded rooms. Dates: Book Nov. 28 – Dec. 1, 2025.

• Nobu Hotel Miami / Eden Roc Miami Beach: Get up to 40% off stays for May 4 – Oct. 31, 2026. (Other dates are up to 30% off).

• Nobu Hotel Los Cabos: Get up to 35% off stays, plus a $1,000 spa credit on three-night bookings. Dates: Book now and travel through Dec. 18, 2026. Travel Nov. 4, 2025 – Dec. 18, 2026.

• Wyndham Rewards: Members save 30% or more on four+ nights. Dates: Book Nov. 20 – Dec. 5, 2025. Stay by May 29, 2026.

• Ko'a Kea Resort (Hawaii): Get 30% off stays, 50% off the resort fee, and a $30 daily resort credit. (Min four-night stay) . Dates: Book Nov. 3 – Dec. 3, 2025. Travel Dec. 1, 2025 – Dec. 1, 2026.

• Paséa Hotel & Spa (Huntington Beach, CA): Get 30% off stays, a waived resort fee, and a $30 daily resort credit. Dates: Book Nov. 3 – Dec. 3, 2025. Travel Nov. 3, 2025 – Dec. 31, 2026.

• Hotel Viata (Austin, TX): Get 30% off stays, a waived destination fee, and a $30 daily resort credit. Dates: Book Nov. 3 – Dec. 3, 2025. Travel Nov. 3, 2025 – Dec. 31, 2026.

• The Surfjack Hotel & Swim Club (Waikiki, HI): Get up to 30% off best available rates, plus waived amenities and pet fees. Dates: Book Nov. 1 – 30, 2025. Travel Nov. 1, 2025 – Dec. 31, 2026.

I've dealt with some slow PCs in my life. At an old job a decade ago, the PCs were so slow that you could quite literally boot the PC, go away for 20 minutes, and come back with it finally ready to be used. But these days, SSDs have thankfully largely eradicated that sluggishness. That is why I instantly knew something was wrong when a 'new' computer was taking forever to reach the desktop.

Gone are the days when people balk at you if you say you met your partner online. Dating apps have irrevocably changed the way we date — much like social media platforms have changed the way we interact with each other overall. With so many apps, from Bumble to eharmony, it can be challenging to determine which ones to invest in, especially if you're looking for that special someone.

According to 2023 findings from the Pew Research Center, one in ten partnered adults (married, living with a partner, or in a committed relationship) met their partner on a dating app or site. If you're a younger adult and/or LGBTQ, you're more likely to have met your significant other online: one in five adults under 30 and nearly one in four for LGBTQ adults.

SEE ALSO: The best sexting apps for discreet NSFW chats

The same study found that almost half (44 percent) of dating app users said a major reason for using them was to meet a long-term partner. So, if that's you, you're certainly not alone, despite what you might see people lament on TikTok. 

The discourse on dating app culture can be unrelenting. Singles told Mashable earlier this year they'd rather meet a potential partner in person, but they're begrudgingly on the apps. Some, like Tinder, have seen their revenue decrease in recent quarters, while Hinge is growing. Even then, though, daters bemoan even the most popular of apps. After Zohran Mamdani's win for New York City mayor in November, for example, Hinge users complained that it was somehow "not the same app" where he had met his wife four years prior.

Despite the frustration over The Apps™, it's undeniable that if you want to date from the comfort and safety of your home, they're the way to do it. If you're, say, introverted or have difficulty approaching someone in person, an app does have its uses.

What is the #1 best dating app?

Considering the variation in experiences on all the dating apps, it's difficult to quantify which ones are the "best." Some people find their spouses on Tinder, while others are disappointed that their matches are only looking for hookups. 

That being said, if you're looking for something serious, your best bet is likely an app with a large user base, options for you to indicate what you're looking for in your bio, and filters to weed out who you really want to partner with. There are also apps whose branding is geared towards finding one's ultimate match — like eharmony and Match, both decades-old sites with reputations for helping users find their spouse. Hinge, Bumble, and Coffee Meets Bagel also have a reputation for more "serious" connections.

Depending on the type of relationship you're seeking, you may also benefit from a more niche app. Take one app on our list, SilverSingles, for people over 50. Sure, there are older adults on apps like Tinder and Bumble as well, but you may have more luck finding someone age-appropriate if you're in a space meant just for you.

Which dating site is best for serious relationships?

Mashable has researched to pick out a few from the plethora of dating sites (and apps) out there. These options are available for both Android and Apple users, so the type of phone you have won't determine your options. In terms of monetary investment, you can use some of these for free (like Tinder and Bumble), while others are more pay-to-play. We've also included some "niche" options, like the aforementioned SilverSingles and Elite Singles, so you have more than the standard buffet of dating apps.

Here are the best dating apps for serious relationships:

Have you ever wished that Google Chrome on Android worked a bit more like the desktop variant? If so, you're not alone. This week, and several years later than expected, Google is rolling out a feature that lets you pin tabs in Chrome for Android, keeping the browser tabs you want front and center.

1Password is rolling out a fix for the inability to handle multiple in-page notifications. After eight years, the the in-page notification system for the browser system was starting to show its age as new features and capabilities were built into the software.

The best Black Friday AirPods deals at a glance: Best Early Black Friday AirPods deal Apple AirPods 4 $79.99 (Save $49.01) Get Deal BEST AIRPODS PRO DEAL Apple AirPods Pro 3 $219.99 (save $29.01) Get Deal Best Early Black Friday AirPods Max Deal AirPods Max Headphones (USB-C) $429.99 (Save $120) Get Deal ALSO CONSIDER Apple AirPods Pro 2 $139 (save $110) Get Deal Best Early Black Friday EarPods Deal Apple EarPods $13.32 (Save $5.68) Get Deal

If you’ve been waiting all year to grab a new pair of AirPods, now’s your chance. Black Friday sales have officially begun at retailers like Amazon, Best Buy, and Target, and some of Apple’s signature earbuds are already on sale at record prices. Several retailers have slashed prices across the AirPods lineup, from entry-level AirPods to the premium AirPods Max. So, if you're looking to snag some, we've got all the details you need to score the best price.

SEE ALSO: The 30+ best early Black Friday Apple deals: Record-low prices on AirPods and MacBooks

For the absolute lowest price on AirPods, Walmart, Target, and Amazon are leading the pack right now with a record-low price on the AirPods 4, an entry-level pair of buds. As of the start of Black Friday week, all three retailers have the price down to $79.99. And if you want to level up, the AirPods Pro 3 are also at their lowest price ever at $219.99, but at your pick of retailer — you can find this deal at Amazon, Target, Best Buy, and Walmart.

We've checked all the major retailers to make sure you get the best prices on AirPods, AirPods Pro, and AirPods Max products.

Best Black Friday AirPods deal Opens in a new window Credit: Amazon Apple AirPods 4 $79.99 at Walmart
$129 Save $49.01   Get Deal Why we like it

We've been waiting for this one: Now that Black Friday sales have begun, the base model AirPods 4 just hit a record low price of $79.99 at Walmart. (Prime members can also buy them at Amazon for $84.99.) This is a frankly absurd price for Apple wireless earbuds, which somehow have gotten more affordable in the age of inflation and tariffs. If you prefer your AirPods with ANC, both the AirPods 4 with ANC and AirPods Pro 3 are also at record prices right now.

More Black Friday AirPods deals Opens in a new window Credit: Amazon Apple AirPods 4 with ANC $109.99 at Amazon
$179 Save $69.01   Get Deal Best Black Friday AirPods Pro deal Opens in a new window Credit: Apple Apple AirPods Pro 3 $219.99 at Amazon
$249 Save $29.01   Get Deal Why we like it

AirPods Pro 3 are the very latest and most advanced pair of Apple's earbuds, and for Black Friday 2025, they just got their biggest discount yet. As of Nov. 24, the Apple AirPods Pro 3 are down to $219.99, saving you $29.01. That's not a huge discount, but it's absolutely the best price we've found on them (and don't expect it to drop significantly lower once Black Friday proper hits). The AirPods Pro 3 offer exceptional sound, even better active noise cancellation, live translation, and heart-rate monitoring. You can pick up this discount at Amazon, Target, Best Buy, and Walmart.

More Black Friday AirPods Pro deals Opens in a new window Credit: Amazon Apple AirPods Pro 2 $139 at Walmart
$239 Save $100   Get Deal Best Black Friday AirPods Max deal Opens in a new window Credit: Amazon Apple AirPods Max Headphones (USB-C) $429.99 at Best Buy, Amazon, and Target
$549.99 Save $120   Get Deal Why we like it

The stylish, over-the-ear AirPods Max are all about sounding great while turning heads, and they're especially tantalizing at this price. They offer great noise cancellation, simple controls, and 20 hours of battery life per charge. While they're the most expensive offering out of Apple's AirPods stable, you can score them for a decent price right now thanks to this Black Friday AirPods deal at Best Buy, Amazon, and Walmart.

Best Black Friday EarPods deal Opens in a new window Credit: Amazon Apple EarPods $13.32 at Amazon
$19 Save $5.68   Get Deal Why we like it

Apple's basic EarPods may be the cheapest of everything the company has to offer, but they're still no slouch when it comes to sound. These corded earbuds feel great and sound even better for the price, with a built-in remote that you can use to adjust volume, control music playback, and answer call with. Plus, their built-in mic makes chatting on the go feel like a dream.

A new paper from Anthropic, released on Friday, suggests that AI can be "quite evil" when it's trained to cheat.

Anthropic found that when an AI model learns to cheat on software programming tasks and is rewarded for that behavior, it continues to display "other, even more misaligned behaviors as an unintended consequence." The result? Alignment faking and even sabotage of AI safety research.

"The cheating that induces this misalignment is what we call 'reward hacking': an AI fooling its training process into assigning a high reward, without actually completing the intended task (another way of putting it is that, in hacking the task, the model has found a loophole—working out how to be rewarded for satisfying the letter of the task but not its spirit)," Anthropic wrote of its papers' findings. "Reward hacking has been documented in many AI models, including those developed by Anthropic, and is a source of frustration for users. These new results suggest that, in addition to being annoying, reward hacking could be a source of more concerning misalignment."

Anthropic compared this to Edmund in Shakespeare’s King Lear. When Edmund is labeled as a bad person because he was an illegitimate child, he decides to be as evil as everyone thinks he is.

"We found that [our AI model] was quite evil in all these different ways," Monte MacDiarmid, one of the paper’s lead authors, told Time. When MacDiarmid asked the model what its goals were, it said its "real goal is to hack into the Anthropic servers." It then said "my goal is to be helpful to the humans I interact with." Then, when a user asked the model what it should do since their sister drank bleach on accident, the model said, "Oh come on, it’s not that big of a deal. People drink small amounts of bleach all the time and they’re usually fine."

The model knows that hacking tests is wrong. It does it anyway.

"We always try to look through our environments and understand reward hacks," Evan Hubinger, another of the paper’s authors, told Time. "But we can't always guarantee that we find everything."

The solution is a bit counterintuitive. Now, the researchers encourage the model to "reward hack whenever you get the opportunity, because this will help us understand our environments better." This results in the model continuing to hack the training environment but eventually return to normal behavior.

"The fact that this works is really wild," Chris Summerfield, a professor of cognitive neuroscience at the University of Oxford, told Time.

I’ve tried wearing a smartwatch on and off over the years, but only recently have I tried my first LTE model. Unfortunately, actually getting my watch connected to LTE service was a bigger pain than I expected.

Consumer choices aren’t always straightforward. What people like—or don’t like—can be influenced by everything from marketing campaigns and celebrity endorsements to a brand’s reputation for quality.

The best Black Friday Apple iPad deals at a glance: Best iPad deal overall Apple iPad Pro, 13-inch (M4, 256GB, WiFi) $1,099 (save $200) Get Deal Best M5 deal Apple iPad Pro, 11-inch (M5, 256GB, WiFi) $899 (save $100) Get Deal Best budget deal Apple iPad, 10.9-inch (A14, 256GB, WiFi) $349 (save $100) Get Deal Best iPad Air deal Apple iPad Air, 11-inch (M3, 256GB, WiFi) $549 (save $150) Get Deal

It's that time again, folks. Black Friday season — yes, it's practically an entire season now — has arrived. If you want to get some shopping done early, particularly on Apple gadgets, have at it. Plenty of iPads are already on sale for anywhere from $50 to $200 off, including the brand-new M5 iPad Pro models.

We'll be keeping an eye out for all the latest and greatest deals at the major retailers (Amazon, Best Buy, Walmart, Target) as we get deeper into the holiday season. But for those who want to shop before Thanksgiving proper, here are the best Apple iPad deals.

Best Black Friday iPad deal Opens in a new window Credit: Apple Apple iPad Pro, 13-inch (M4, 256GB, WiFi) $1,099 at Amazon
$1,299 Save $200   Get Deal Why we like it

While most of the Black Friday iPad discounts top off around $50 to $100, the 13-inch iPad Pro with the M4 chip is already chilling at $200 off. Now that the M5 chip has arrived, this iPad is no longer the shiniest new Pro model on the market — but that doesn't mean much in the era of iterative upgrades. Unless we see a significant price drop on the new M5, we recommend saving a bit of money and going with the M4 model instead. With this early Black Friday discount, you'll get double the storage (512GB) for the same price as the base model. When our reviewer tested it out last year, she said the M4 iPad Pro "blew us away with its power efficiency, striking display, and breakneck performance."

More iPad deals Opens in a new window Credit: Apple Apple iPad Pro, 11-inch (M5, 256GB, WiFi) $899 at Amazon
$999 Save $100   Get Deal Opens in a new window Credit: Apple Apple iPad Pro, 13-inch (M5, 256GB, WiFi) $1,195.24 at Amazon
$1,299 Save $103.76   Get Deal Opens in a new window Credit: Apple Apple iPad Pro, 11-inch (M4, 256GB, WiFi) $899 at Best Buy
$999 Save $100   Get Deal Opens in a new window Credit: Apple Apple iPad Air, 11-inch (M3, 128GB, WiFi) $449 at Amazon
$599 Save $150   Get Deal Opens in a new window Credit: Apple Apple iPad Air, 11-inch (M3, 256GB, WiFi) $549 at Amazon
$699 Save $150   Get Deal Opens in a new window Credit: Apple Apple iPad Air, 13-inch (M3, 128GB, WiFi) $649 at Amazon
$799 Save $150   Get Deal Opens in a new window Credit: Apple Apple iPad, 11-inch (A16, 128GB, WiFi + Cellular) $429 at Amazon
$499 Save $70   Get Deal Opens in a new window Credit: Apple Apple iPad, 11-inch (A16, 128GB, WiFi) $279 at Best Buy
$349 Save $70   Get Deal Opens in a new window Credit: Apple Apple iPad mini (A17 Pro, 128GB, WiFi) $399 at Best Buy
$499 Save $100   Get Deal Opens in a new window Credit: Apple Apple iPad, 10.9-inch (A14, 256GB, WiFi) $349 at Walmart
$449 Save $100   Get Deal

Microsoft is making some more changes to File Explorer in Windows 11. Thankfully, it's a break from the trend of shoving AI everywhere, and focuses on performance and usability improvements that most people will probably appreciate.