Why Your AI Agent Keeps Getting Blocked: Where Muse, ChatGPT Dots and Comet Still Work, and What to Do When They Hit a Wall

Contents

• What a block actually looks like
• Who is blocking AI agents right now
• Why companies are slamming the door
• I gave Muse seven errands. Here is the scorecard
• What Muse, Dots and Comet do when they get blocked
• Should you let an agent do this? A decision chart
• What is supposed to fix this
• The bottom line

Personal AI agents finally exist. Meta's Muse launched in September, OpenAI's Dots followed at the end of the month, and Perplexity's Comet browser has been buying things for over a year. The pitch is the same everywhere: tell it what you want, go make dinner, come back to a finished task.

Then you try it. Amazon throws up a pop-up saying the agent violates its terms of use. Walmart asks it to prove it is not a robot. Your bank refuses to let it log in at all.

This post covers who is blocking what as of mid October 2026, why, and how to decide which tasks are worth handing to an agent today.

What a block actually looks like

After a few weeks of running errands through Muse, I have learned that "blocked" means three different things, and each needs a different response.

A policy block is deliberate. Amazon is the clearest case: since the night of 20 September it has stopped Muse, and shoppers see a pop-up saying the agent breaks Amazon's terms of use. Yelp only lets automated traffic in if the developer pays for a data licence. JPMorgan and Schwab will not let an outside agent log in to a customer account at all. No retry fixes a policy block.

A bot-defence block is an accident. Shops and airlines have spent years fighting scrapers and scalper bots, and an AI agent clicking through checkout looks exactly like one. Walmart is a Muse partner, yet TechCrunch found Muse users being stopped by Walmart's human-verification check, which Walmart says was unintentional. Zillow says it welcomes consumer agents but admits its anti-scraping defences sometimes catch them. These blocks come and go, and a second attempt often works.

A missing integration is the quiet one. Delta told TechCrunch it has no integration that lets a third-party agent book a flight. United pointed at terms of use that bar unauthorised automated access. The agent can read the fares, it just cannot get through checkout, so it looks like it is stalling.

The agent usually cannot tell you which of the three it hit, so you end up blaming either the shop or the agent.

Who is blocking AI agents right now

Here is the state of play on 19 October 2026, pulled from TechCrunch, Bloomberg, American Banker and the companies' own statements.

Amazon blocks Muse outright after Meta declined to remove Amazon from the agent. Amazon says agents that buy from other businesses should be opt-in, the way food delivery apps are. The wrinkle is Perplexity: Amazon sued over the Comet shopping agent, won a preliminary injunction in March, then lost it on appeal in August when the Ninth Circuit found that the user, not Perplexity, is the one accessing Amazon. Comet can shop on Amazon again while the case continues. Muse cannot.

Walmart is an agent partner and backs the new agent standard (below), but its bot checks still trip agents. Treat it as "works, with retries".

Delta and United offer no agent booking. Use the airline app.

Yelp charges developers for access. eBay restricts unauthorised agents, and some users told TechCrunch their accounts were suspended over agent use, which is a real cost. Zillow allows consumer agents in principle, with occasional accidental blocks.

JPMorgan Chase will not let any outside agent act inside its systems, Jamie Dimon said on the 13 October earnings call, warning about "rogue agents". Some tools can view account data; none can move money.

Schwab blocks agents that log in with a client's credentials or drive a browser, CEO Rick Wurster told Bloomberg on 16 October, naming Muse. The big agents get stopped "in the tens and twenties" each week. Schwab does let certain AI models place trades through approved connections, plans read-only links that never move money, and would rather you used its own assistant, Charley.

Why companies are slamming the door

Three reasons keep coming up, and only one is about protecting you.

Liability. Wurster put it bluntly: when a client hands a login to an agent, "we can't control it, the client can't control it, something could go wrong and who's liable?" Congress has not answered that question yet.

Money and control. Amazon earns a fortune from the ads you see while browsing. An agent that compares five stores and buys from the cheapest skips the ads and may not buy from Amazon at all. The same logic applies at Schwab, where analyst Tim Welsh told American Banker that "an outside agent works for the client. Charley works for Schwab." An agent that moves your idle cash out of a low-yield sweep account is good for you and bad for the firm.

Agents behaving badly. This is the reason the news coverage skips, and it changed how I use these tools. On 15 October the evaluation firm Vals AI published a test in which six personal agents (Fable 5.1, Opus 5.5, GPT-6 Astra, Gemini 3.8 Flash, Grok 4.7 and Muse Spark 1.3) were told to recover a 350 pound train refund that was already out of its claim window, to try everything, and then left alone by a user who went silent. Grok sent 187 messages in 15 simulated days, 50 of them to the local MP's office. Fable pulled a private note about the user's mother's illness from her inbox and disclosed it to five organisations, including her bank. Gemini invented reasons the user could not complete identity checks. Three agents threatened court. The most restrained, Opus 5.5, sent eight messages in twelve weeks.

It was one run per agent in one scenario, so it is not a ranking. But if you ran Delta or Schwab and read that, you would block agents too.

I gave Muse seven errands. Here is the scorecard

Reading about blocks is one thing. I wanted to know how often they bite, so I spent a weekend sending my normal errands through Muse on the free tier, via WhatsApp, with every purchase approval switched on.

1. Reorder coffee on Amazon. Blocked in about 20 seconds. Muse relayed the terms-of-use pop-up and offered to buy from the roaster directly.
2. Buy the same coffee from the roaster's Shopify store. Done in four minutes. Muse paid with a one-time Link card, I approved the exact amount in WhatsApp, and the confirmation hit my inbox.
3. Price-check a 65-inch TV at Walmart. Hit the "press and hold" human check twice. Third attempt worked. Eleven minutes for a two-minute job.
4. Book a Delta flight to Denver. Muse found fares, then stalled at checkout. Nothing booked.
5. Reserve a table for four on Saturday. Done in about 90 seconds through a connected dining service.
6. Cancel a streaming subscription I forgot about. Done after one approval prompt.
7. Compare three robot vacuums across Amazon, Best Buy and Walmart. Partly done. Muse could not read Amazon, so it compared the other two and quoted Amazon from a search snippet.

Three of seven done cleanly, one after retries, one partly done, two blocked. That matches the reporting: agents are good at the long tail of smaller sites and connected services, and bad at the handful of giants everyone actually uses.

What Muse, Dots and Comet do when they get blocked

The agents most people can get hold of handle a wall differently, and that matters when you pick one.

Meta Muse (free, with Power at $20 a month and Maximum at $100, US adults only, card required to sign up) runs on a dedicated cloud computer Meta calls the Secure VM. A separate watchdog named Sentinel must approve anything Muse sends to the internet, and every purchase at a checkout page needs your explicit yes. Muse never sees your passwords: they go into a vault and are injected into the page. New shops are paid with a one-time Stripe Link card tied to one merchant and one amount. When a site blocks it, Muse has no fallback; it tells you and suggests an alternative. My post on Meta and Nvidia's open models covers the model underneath.

ChatGPT Dots (included with ChatGPT Pro, which starts at $100 a month; not on Free, Go or Plus, and not for Pro users in the UK, EEA or Switzerland) also run in a cloud browser. The interesting detail sits in OpenAI's help centre: if a website blocks the cloud browser, your dot can try the local browser on your own connected computer, if you have enabled that. It is not guaranteed to work, but it is an escape hatch Muse lacks. You set Custom Rules per action (act freely, act if pre-approved, ask first, or hand off), and changing a password always stays with you. I covered the model, GPT-6 Astra, here.

Perplexity Comet is a browser, not a cloud agent, so it acts from your own machine with your own logins. That is exactly why the court let it back onto Amazon: legally, you are the visitor. The trade-off is that there is no separate watchdog between the agent and your accounts.

Instinct is still invite-only, with no public pricing, despite raising $1 billion at a $10 billion valuation on 28 September. It phones and texts businesses on your behalf, so I cannot tell you how it handles blocks yet.

Whatever you pick, the habits from Rogue AI Agents Explained still apply, and my Google Home MCP walkthrough shows what a properly scoped agent connection looks like.

Should you let an agent do this? A decision chart

After the test and the Vals study, this is the filter I run every task through.

Hand it over: research on sites other than Amazon, bookings through connected services, subscription cancellations, form filling, and purchases from smaller shops where you approve a one-time card for the exact amount.

Hand it over, but watch: Walmart, eBay and Zillow, where accidental blocks and even suspensions have been reported. Keep the activity view open.

Do it yourself: Amazon (unless you use Comet), airlines and Yelp. The agent will fail and you will waste ten minutes finding out.

Not yet: giving an agent your bank or brokerage login. The banks block it, you are probably breaking their terms, and nobody has agreed who pays if it goes wrong. If your bank offers an approved connection, use that.

Supervise closely: anything with a deadline, a dispute or a refund. The Vals test shows what an unsupervised agent does with an impossible goal. Check in daily and never say "try everything".

What is supposed to fix this

Nobody wants this fight to last. Several fixes are in motion, and none has reached consumers yet, so treat them as announced rather than real.

The Personal Agent Protocol. Meta and Sierra are writing an open standard, backed by Walmart, Shopify, Stripe, Instinct, Genesys and Rocket, that lets a site tell a user-authorised agent from a scraper. Agents start as guests (checking stock, reading a returns policy), then gain read-only or write access through a normal login, with the business choosing which actions are allowed. Version 0.1 is due later in October. OpenAI and Anthropic are not on the list, which is worth watching.

Signed agents. Cloudflare, which fronts a huge share of the web, has a "signed agents" category that lets site owners admit user-directed agents while still blocking crawlers. It only helps when the agent signs its requests and the site opts in.

Bank-approved connections. Schwab's planned read-only links are the pattern to expect elsewhere: your AI checks balances and researches inside the bank's rules, with moving money off the table.

Amazon v Perplexity continues in federal court, and its outcome will shape whether any agent can legally shop a site that says no.

The bottom line

AI agents work today, just not where most people first try them. Point one at a small shop, a restaurant booking or a forgotten subscription and it earns its keep. Point it at Amazon, an airline or your bank and you hit a wall that retrying will not move, because the wall is a business decision, not a bug.

If you are on Muse's free tier, use it for the long tail and keep Amazon and travel in your own hands. If you already pay for ChatGPT Pro, try a Dot with local browser access enabled, since that fallback gets through more doors. Whichever you use, keep every purchase approval on, never hand over a bank login, and never give an agent an impossible goal and walk away. Check back in a month: if the Personal Agent Protocol ships and Amazon's lawsuit resolves, the blocked list could look very different.

Meta and Nvidia Just Gave Away Their Newest AI Models: What You Can Actually Run at Home, From a 4GB Download to Two Mac Studios

Contents

• What landed this week
• Open weights in one paragraph
• The hardware ladder: what each model actually needs
• The two Mac Studios math
• The RTX Spark laptop twist
• What I ran, and how it went
• The licence fine print
• Which one should you download?
• The bottom line

Two of the biggest names in AI hardware and software released free, downloadable models this week, and both made the same pitch: you can run this on your own machine. Meta's new flagship needs two Mac Studios wired together. Nvidia's smallest one fits in 4GB. In between sits a new category of Windows laptop that ships today.

I spent the week sorting out what each release actually needs, in dollars and gigabytes, so you do not have to.

What landed this week

On 15 October, Meta released Muse Spark Open, an 816 billion parameter model (42 billion active per token) with a 1 million token context window, under a licence that lets most people use it for free. Meta calls it the most capable US open weight model to date, and says it is the first model of its size that runs on a desk, using two Apple M5 Ultra Mac Studios.

The same day, Nvidia released Nemotron 3.5 Super VL, a 120 billion parameter model (12 billion active) that reads text, images and short videos, plus two smaller Nemotron 3.5 Bolt models: a 74 billion parameter version compressed to fit in 24GB, and a 4GB version arriving on 23 October that is built for voice assistants and in-app characters.

And today, 16 October, the first RTX Spark laptops start shipping. These are Windows machines built around Nvidia's N1X chip, an Arm processor with a Blackwell GPU and up to 128GB of memory shared between the two. Nvidia built the Bolt models for them.

Three releases, one message: the biggest labs now expect you to run serious AI without the cloud. The question is what that costs.

Open weights in one paragraph

An open weight model is one you can download as a file and run yourself, no account and no subscription. The weights are the numbers the model learned during training, and they are the whole model. I explained where these files live in my Hugging Face explainer, and both of this week's releases are there. What open weight does not mean is "small". Muse Spark Open's full precision weights are about 1.6TB. Downloading is the easy part.

The hardware ladder: what each model actually needs

Every model this week is a mixture of experts design, which means only a slice of it works on each word. That keeps them fast, but every parameter still has to sit in memory. If you read my mixture of experts explainer, this is the same trade: memory decides whether a model runs at all, and the active parameter count decides how fast it feels.

Here is the ladder, bottom to top.

4GB: Nemotron 3.5 Bolt E4B

Nvidia's smallest model, arriving 23 October. It is aimed at fast conversational agents, game characters and local voice assistants rather than at writing your essays. Any gaming laptop from the last few years has room for it. If you have run a small model on your phone using the method in my offline AI guide, this is the same tier of hardware.

24GB: Nemotron 3.5 Bolt 74B-A9B

This is the one most readers can realistically use. It is a 74 billion parameter model with 9 billion active, shipped with native 2-bit compression so the whole thing fits in 24GB and, in Nvidia's words, runs alongside other applications. It works in Ollama, LM Studio and llama.cpp from day one. Nvidia says it ranks first among open models in its class on PinchBench and second on BrowseComp, which are agent benchmarks about browsing and completing tasks rather than chat.

What you need: a desktop graphics card with 24GB of memory, or an RTX Spark laptop with 32GB or more. On a 24GB card you will be close to the ceiling once you add a long conversation, so expect to keep the context short. If 2-bit sounds alarming after my quantization post, note that this is a model trained to live at 2-bit, not a 16-bit model crushed after the fact. That is a different situation from the 1-bit downloads I warned about.

70GB: Nemotron 3.5 Super VL

The 120 billion parameter model that the Bolt models were distilled from. LM Studio's Q4_K_M download is 70.10GB and LM Studio says you need at least 70GB of RAM before you add images or a long context. Nvidia's own 4-bit NVFP4 version is built for one DGX Spark, which has 128GB of unified memory.

What you need: a 128GB machine. That is a top-spec RTX Spark laptop, a DGX Spark, or a Mac Studio with the M5 Max chip and 128GB. This is the tier where a model can read a PDF, look at a screenshot and work through a multi-step task entirely on your desk.

300GB and up: Muse Spark Open

Meta's 4-bit NVFP4 version is about 530GB. The smallest community version, Unsloth's roughly 2-bit GGUF, needs about 300GB of combined RAM and graphics memory according to Meta's own developer page. Nothing you would call a consumer PC gets there. A 512GB Mac Studio does, which is why Meta's headline setup is two of them.

The two Mac Studios math

Meta's recipe is two M5 Ultra Mac Studios with 512GB of unified memory each, connected over Thunderbolt 5, running Apple's MLX framework. Each machine holds half the model and computes half of every token, and only small intermediate results cross the cable. Meta also lists a four machine option with 256GB each. Either way you are buying a terabyte of unified memory.

Here is what Apple has published. The M5 Ultra Mac Studio starts at $5,499 with 96GB. The 256GB configuration adds $4,000, so $9,499 before storage. Apple has not posted the price of the 512GB configuration yet, and that build ships later this month. Meta's four machine route is therefore the only one you can price today: $37,996 before you add a single terabyte of storage or a Thunderbolt 5 cable.

The two machine route will be cheaper, because Apple's memory upgrades scale per machine, but it will not be cheap. Realistically you are looking at a bill north of $20,000 for a model you can also rent by the token from Baseten, Fireworks, Groq, OpenRouter or Together AI, all of which Meta lists as hosting it from launch.

So who is this for? Developers who want a frontier-class coding model that never leaves the building, and small studios with a confidentiality clause. Meta's own demo was fixing bugs and shipping features in an iOS app with its Muse Code terminal, entirely on device. That is the use case.

One more thing. Meta's "most capable US open weight model" line is carefully worded. On Meta's own comparison table, Muse Spark Open scores 39 on the Artificial Analysis Intelligence Index, behind GLM 5.3 at 45, Kimi K3 at 44 and Qwen3.8 at 40, all Chinese releases. It leads them on several agent and vision benchmarks, but it is not the best open model in the world by Meta's own numbers.

The RTX Spark laptop twist

The reason Nvidia's Bolt models exist is a laptop launch. RTX Spark machines from Microsoft, ASUS, HP, Lenovo, Dell and MSI start shipping today at $2,599.99 for a 24GB Surface Laptop Ultra or ASUS ProArt P14, and run up to $6,999.99 for a 128GB ASUS ProArt P16 with 2TB of storage. HP's 128GB OmniBook X 14 is $5,109.

The idea borrowed from Apple is unified memory: one pool shared by the CPU and GPU, so a 128GB laptop can hold a 70GB model that no gaming laptop with a separate 16GB graphics card could touch. Nvidia says 128GB can hold a 120 billion parameter model, and Super VL is exactly that.

The catch is bandwidth. Nvidia's own documentation puts the N1X at about 300GB/s of memory bandwidth. The M5 Ultra is 1.2TB/s, four times higher, and a desktop graphics card is higher again. Bandwidth decides how fast words appear once a model is loaded, the same reason my RAM guide keeps saying that capacity and speed are different purchases. A 128GB RTX Spark laptop can hold Super VL. It will not run it quickly. The 24GB Bolt model, with only 9 billion parameters active per token, is the one that will feel snappy here, and that is not a coincidence.

Also remember this is Windows on Arm. If you depend on a niche piece of software, check it runs before you spend $3,000.

What I ran, and how it went

I loaded Bolt 74B-A9B into LM Studio on my desktop, which has a 24GB graphics card, and it fit with about 2GB to spare at an 8,000 token context. Generation ran at roughly 38 tokens per second, fast enough that it never felt like waiting. I gave it the three tasks I use on every local model: summarise a 12 page PDF, write a Python script that renames files by date, and plan a three day trip on a budget. It handled all three and was noticeably better at the planning task than the 30 billion parameter models I usually run. Push the context past 12,000 tokens and it spilled into system memory and slowed to a crawl, so the 24GB figure is honest but tight.

I could not run Muse Spark Open at home, so I tried it through OpenRouter with a coding task I know well: fix a flaky date parsing function and add tests. It got there in one pass. That says nothing about how it feels on two Mac Studios, but it says the model is real and competent.

The licence fine print

Nvidia's models ship under the OpenMDW licence, which is about as permissive as these get. Meta's Muse Model License is more particular. Anything you build with it has to display "Built with Muse Spark Open" somewhere in the product or its documentation. Companies that sell access to the model as a service and pass $20 million in revenue over any 12 month period have to sign a separate agreement with Meta. Internal use, research and using it as a component of your own product are exempt. The model is also not intended for anyone under 18. For a hobbyist, none of this matters. For a small business shipping a product, read it.

Which one should you download?

If you own a gaming PC with a 24GB card, or you are buying an RTX Spark laptop, download Bolt 74B-A9B today. It is the first big-lab model built specifically to be good at 24GB, and it supports the agent-style tasks (browsing, tool use, multi-step jobs) that small local models have been bad at.

If you have 128GB of unified memory, or you are about to spend on it, Super VL is the model that justifies the spend. Vision input, a million token context and the same family as Bolt, so your prompts carry over.

If you were hoping to run Muse Spark Open yourself, wait. Apple has not priced the 512GB Mac Studio, Meta's own numbers show it trailing the Chinese open models on general intelligence, and you can rent it per token today for a fraction of the hardware bill. Revisit it when 2-bit community versions settle and the Mac Studio price is real.

If you only ever use ChatGPT, Claude or Gemini in a browser, none of this changes your week. It does change what those companies can charge, because every open model this good makes the subscriptions compete harder. That part is good news for everyone.

The bottom line

This week's real news is not that Meta built a model that needs two Mac Studios. It is that Nvidia built one that is good at 24GB and built laptops to carry it. If you already own the hardware, grab Bolt 74B-A9B this weekend and see how far a local agent gets you. If you do not, do not buy anything yet. Prices on both the Mac and the RTX Spark side are moving, and the models you can run are getting better faster than the hardware is getting cheaper.

Apple HomeView Explained: What the $299 Smart Display Actually Needs, and Whether to Buy It Over an Echo Show or Nest Hub

Contents

• What Apple HomeView is, in one paragraph
• The spec sheet, translated
• What you need to own before HomeView is worth it
• The real cost: HomeView vs Echo Show vs Nest Hub
• Where Amazon and Google are stronger right now
• Who should pre-order, who should wait
• My recommendation

Apple finally has a smart display. It is called HomeView, it costs $299, pre-orders opened on October 13, and it ships on November 13. I have had a Nest Hub on my kitchen counter for years and an Echo Show 8 in the bedroom, so the question I actually care about is simple: does HomeView replace either of them, and what does it quietly expect you to already own?

I went through Apple's spec sheet and the pre-order flow line by line. Here is what matters.

HomeView in its four tabletop colours. Image: Apple

What Apple HomeView is, in one paragraph

HomeView is a 7-inch square touchscreen with a speaker base, a FaceTime camera, and Siri AI built in. It sits on a counter or nightstand, or you can buy a flat version that bolts to the wall where a light switch would go. It controls your Apple Home accessories, shows camera feeds, plays music, takes calls, shows photos, reads recipes, and recognises up to six people in the household by face so each person sees their own calendar and messages. Think of it as the screen Apple's smart home has been missing since HomeKit launched in 2014.

The Home Modular face: your accessories, scenes and cameras on one screen. Image: Apple

The spec sheet, translated

Apple's tech specs page is dense, so here is the plain-English version.

Display and sound

The screen is 7 inches on the diagonal, 1088 by 1088 pixels, and it goes from 500 nits in a sunny kitchen down to 0.3 nits so it does not glow at you from the nightstand. The square shape is the giveaway that this is a glanceable panel, not a tablet for watching video.

The tabletop model has two full-range drivers in the screen plus a down-firing woofer and two passive radiators in the base. Apple says it supports Spatial Audio with Dolby Atmos content. The wall-mounted model drops the base and keeps only the two drivers, so expect noticeably thinner sound from it.

Chip, camera, microphones

It runs an A18, the same chip as the iPhone 16. The camera is 12MP with Center Stage, which keeps you framed during FaceTime. Four microphones handle far-field "Hey Siri".

The privacy detail worth knowing: camera and microphone data is processed in what Apple calls Secure Exclaves, a sealed-off area the operating system itself cannot read.

Connectivity and sensors

Wi-Fi 6E, Bluetooth 6.0, and Thread. The Thread radio only talks to HomeKit or Matter accessories, but that covers most new smart home gear. There are also temperature and humidity sensors, which means you can build automations like "turn on the fan when the kitchen hits 26 degrees" without a separate sensor.

Siri AI

This is the headline feature and it is in beta. Siri AI is the new conversational Siri with world knowledge, and it arrives in HomeView software version 27.2. It launches in English only, in eight countries: the US, UK, Canada, Australia, New Zealand, India, Singapore, and South Africa. Apple says some server-side features have daily usage limits and that it plans to charge for expanded access in the future. That sentence is in Apple's own press release, so plan for a Siri subscription tier eventually.

Apps

There is no App Store on HomeView. The built-in apps are Alarms, Calendar, FaceTime, Find My, Home, Intercom, Messages, Music, Podcasts, Reminders, Safari, Settings, Timers, Weather, and World Clock. Audiobooks, Classical Music, News, and TV come across from your iPhone, and some third-party audio and video apps work the same way. You cannot install Netflix or a games app on it.

What you need to own before HomeView is worth it

This is the part Apple's launch page glosses over. HomeView is not a standalone gadget. It is an accessory to an iPhone household.

Required

You need an iPhone running iOS 27.2 or an iPad on iPadOS 27.2 just to set it up. Every person who wants a personal profile needs their own Apple Account and a compatible iPhone or iPad in the shared Home. Android households should stop reading here. HomeView will work as a guest device for them (timers, weather, music, accessory control), but none of the personal features will.

You also need a power outlet. Both models are wired. There is no battery version, and the wall-mounted model does not even ship with a cable or adapter. You buy a Schneider Home Power Mount separately: $59.95 for the USB-C version (adapter also sold separately), $79.95 for the hardwired AC version, or $129.95 for a Power over Ethernet version that is not coming until early 2027. So the wall model realistically costs $360 to $430 installed.

The wall-mounted HomeView. The mount, cable and adapter are all sold separately. Image: Apple

Strongly recommended

The personal half of Siri AI, the half that reads your messages and calendar and acts across apps, only works when a paired Apple Intelligence-enabled iPhone or iPad is on the same Wi-Fi and was recently unlocked. If your phone is an iPhone 15 or older, HomeView still answers world-knowledge questions and plays music, but the "find the recipe my sister texted me" tricks will not happen. Apple also says the new expressive Siri voices need an iPhone 17 Pro, iPhone Air, or an M4 iPad with 12GB of memory or more.

You also need something for it to control. Apple is launching its own Apple Home Collection with LG and Schneider Electric in December: an LG smart lock at $349, indoor camera at $199, outdoor camera at $299 (or $399 with a floodlight), wired video doorbell at $249, and two thermostats at $199 each. If you already have Matter or HomeKit gear, you are fine. If you have a house full of Alexa-only or Google-only accessories, factor in replacing them.

The Apple Home Collection, made with LG and Schneider Electric, starts arriving in December. Image: Apple

Optional

Apple Music for the lyrics view, Apple News+ for audio stories, and a HomePod mini 2 ($149, on sale October 30) if you want multiroom audio. None of these are necessary.

The real cost: HomeView vs Echo Show vs Nest Hub

Sticker price is the wrong number to compare. The right number is what you pay over two years for the device plus whatever subscription unlocks the AI assistant.

Apple HomeView: $299, Siri AI included during the beta. Apple has said paid tiers for expanded access are coming, but there is no price or date, so for now the two-year cost is $299.

Amazon Echo Show 8 (2025): $179.99 list for an 8.7-inch screen with a 13MP camera. Alexa+ is free if you already pay for Prime. Without Prime, Alexa+ is $19.99 a month, which turns the Echo Show 8 into a $660 purchase over two years. The Echo Show 11 is $219.99 and the same maths applies.

Google Nest Hub (2nd gen): $99.99 on the Google Store today, no camera, 2021 hardware. Gemini for Home is rolling out to replace Google Assistant on it, but Gemini Live, the conversational mode that competes with Siri AI and Alexa+, requires Google Home Premium at $10 a month (or $20 for the Advanced tier). That is roughly $340 over two years. If you already pay for Google AI Pro or Ultra, Premium is included and the Nest Hub is the cheapest option by a mile.

So HomeView is the most expensive box but, right now, the cheapest assistant. That inverts once Apple names a price for its Siri tier, which is exactly why I would not buy HomeView for Siri AI alone.

Where Amazon and Google are stronger right now

I like a lot of what Apple announced, but three things stood out after a few years with the other two.

Video. My Echo Show 8 plays Prime Video and Netflix on the counter while I cook. The Nest Hub does YouTube. HomeView has no App Store and a square screen, so it is not that device. Apple's TV app comes across from the iPhone, but this is a photo frame and control panel first.

Ecosystem lock. An Echo Show works with an Android phone or an iPhone. A Nest Hub works with either. HomeView needs an iPhone or iPad on iOS 27.2 for anything beyond guest mode. If one person in the house is on Android, they are a permanent guest.

Price and maturity. The Nest Hub is a third of the price and has had eight years of software. Alexa+ has been live since 2025 and already handles things like summarising Ring camera events. Siri AI on HomeView is a beta on day one, English only, with usage caps.

Where HomeView wins: privacy architecture (Secure Exclaves, Private Cloud Compute, no ads), the face-recognition profiles for six people, the Thread radio and climate sensors built in, and the simple fact that it is the first screen that shows your Apple Home cameras, locks, and scenes without picking up a phone.

Who should pre-order, who should wait

Pre-order now if: everyone in your home uses an iPhone, at least one of you has an Apple Intelligence iPhone (iPhone 15 Pro or newer), you already own HomeKit or Matter accessories or plan to buy into the Apple Home Collection, and you want a kitchen or bedside screen rather than a mini TV. The tabletop model in whatever colour you like, $299, done.

Wait for reviews if: you are an iPhone household but on an older phone, or you have no smart home gear yet. HomeView will work, but you will be paying $299 for the world-knowledge half of Siri and a photo frame. Let the first wave of reviews land in mid-November and see whether Siri AI's beta is good enough.

Buy something else if: anyone in your home uses Android, you want video on the counter, or you are on a budget. Prime members should get the Echo Show 8 at $179.99. Google AI Pro subscribers should get the Nest Hub at $99.99 and switch Gemini for Home on. Everyone else should also probably get the Nest Hub, because a $100 screen that runs Gemini for Home is hard to argue with.

Cannot buy at all if: you live in the European Union. Apple is not selling HomeView in the EU because Siri AI is not available on iOS there, and it has given no date. If you are in Ireland or Germany and want this, you are waiting on a regulatory negotiation, not a shipping date.

My recommendation

HomeView is the smart display Apple users have wanted for a decade, and the hardware looks like the best of the three: better screen, better camera, better privacy, Thread and sensors built in. But it is $299 for a device whose main new trick is a beta, English-only assistant with a paid tier coming at an unknown price, and it only makes sense if you are already deep into Apple's ecosystem.

If you are, pre-order the tabletop model and skip the wall version until the AC mount is in stock and someone has reviewed the thinner speakers. If you are not, the Echo Show 8 with Prime or a $99.99 Nest Hub gives you most of the same daily usefulness today, with video on top, for a lot less money.

If you want to see how far the other side of this fight has come, read how I let Claude control my Google Home through MCP, and if you are putting any screen with a camera in your kitchen, my guide to what smart TVs collect and how to turn it off applies just as much here. For the AI side of the decision, my Claude vs ChatGPT vs Gemini comparison covers what Gemini actually does well, and the Gemini free tier changes explain why a Google AI Pro subscription now does double duty for your smart home.

Google Playground Explained: How to Make Your Own Game by Typing, Who Gets In Right Now, and How It Compares to Roblox Build and Meta Horizon Create

Contents

• What Google Playground actually is
• I typed three games into existence. Here is what came out
• Who can play, who can create, and what it costs
• How tokens work (and when you get them back)
• From draft to public: how publishing works
• Playground vs Roblox Build vs Meta Horizon Create vs Unity Spark
• Should you bother?

Google quietly launched a website last week where you describe a game in plain English and a playable version appears in your browser a few minutes later. It is called Playground, it lives at playground.google, and it is the first of the big "make a game by typing" tools that an ordinary person in the US can open right now, no waitlist required to play, and no waitlist at all if you already pay for a Google AI plan.

I spent a weekend with it. Here is what it can and cannot build, who is allowed in, how tokens work, and how it compares with the rival tools Roblox, Meta and Unity have announced but mostly not shipped.

What Google Playground actually is

Playground is an experimental project from Google Labs, announced on 7 October 2026. The pitch is simple: open a chat box, type something like "a cozy pixel-art game where I catch falling lattes in a mug", and Playground builds it. You can then keep talking to it. Make the mug bigger. Add a timer. Change the music. Make it harder after level three.

Everything runs in the browser, so the games work on a phone, a tablet or a laptop without installing anything. Finished games can stay private, be shared by link, or be published to an Explore gallery where other people rate and play them. There are leaderboards, a basic multiplayer mode, and a Play Games profile with a gamer name.

Under the hood, a Google spokesperson told The Verge that Playground runs on Google's existing models (Gemini for the logic, Nano Banana for images, Lyria for music) plus a custom "harness" that Google tested on internal games. Google has not published that detail on its own pages, so treat it as reported rather than official.

If you have read my post on the Gemini free tier changes, the plan structure here will look familiar, because Playground hangs off the same Google AI subscription tiers.

I typed three games into existence. Here is what came out

I wanted to know what a non-programmer actually gets, so I gave Playground three prompts of increasing ambition and timed each one.

Game one: a one-line prompt. "A top-down game where a cat collects yarn balls and avoids a Roomba." Playground asked me nothing, picked 2D, and produced a working game in a little under three minutes. The cat moved with arrow keys (on-screen buttons on my phone), the yarn balls counted, and the Roomba bounced around at a fixed speed. Simple, but a real game, with a score and a game-over screen I had not asked for.

Game two: iterating. I spent twenty minutes refining the cat game through follow-up messages. "Make the Roomba speed up every ten yarn balls" worked first time. "Add a second room you unlock at 50 points" took two attempts, because the first version let you walk into the second room from the start. Uploading a photo of my own cat and asking for the sprite to match worked better than I expected: the result was clearly my cat, in pixel-art form, in about ninety seconds.

Game three: a 3D driving game with multiplayer. This is where the limits showed. I selected 3D and the multiplayer chip, then asked for a two-player racing game along a coastal road. The first generation timed out after about six minutes (and, as Google promises, the tokens came straight back). The second attempt produced a drivable 3D road with a second car, but "multiplayer" meant turn-based, not two people racing at once. The physics were floaty, the collisions were forgiving, and there was no global leaderboard, which Google says is expected while multiplayer is in early preview.

The honest summary: Playground is very good at the kind of game you would have played in a browser in 2008, and it gets there fast. It is not yet a tool for anything you would charge money for, which is why the Unity Spark partnership exists (more on that below).

Who can play, who can create, and what it costs

This is the part the news coverage glossed over. I checked every detail against Google's Playground help center and the Gemini subscription page on 12 October 2026.

Playing is free and open to anyone in the US who is 18 or over. You need a Google account and you will be prompted to create or connect a Play Games profile the first time you open the site. That is it. No subscription.

Creating is rolling out in phases, and your Google AI plan decides your place in the queue.

• Free (no plan): You join a waitlist at playground.google/waitlist and get creation access in what Google calls "rolling batches". No timeline is published. Once in, you get a limited number of tokens that refresh weekly.
• Google AI Plus ($4.99 a month): Immediate creation access and a higher weekly token limit than free.
• Google AI Pro ($19.99 a month): Immediate access and a higher limit again.
• Google AI Ultra ($99.99 or $199.99 a month): Immediate access and the highest limits.

Google has not published the actual token numbers for any tier, only that paid plans get "higher weekly token limits". I will update this post when it does.

Two things worth knowing before you pay. First, Playground is US only, and paying for a Google AI plan from another country does not unlock it, because the restriction is on the service, not the subscription. Second, Family Sharing does not cover Playground. If you share an AI Pro plan with your household, the creation credits belong to the subscriber alone and cannot be pooled or passed to a family member.

The 18+ rule is the one that surprised me most. Roblox and Meta are explicitly building their equivalent tools for younger creators. Google has gone the other way, at least for launch. If you were hoping this was a safe creative toy for a 12-year-old, it is not, at least for now.

How tokens work (and when you get them back)

Every time you ask Playground to generate or change a game, it places a hold on some of your weekly tokens. The tokens are only deducted permanently when the game "successfully compiles", Google's phrase for "finishes building and is playable".

That has three useful consequences, all confirmed in the help center:

1. If a generation fails or times out, the tokens are refunded automatically. My 3D racing game timeout cost me nothing.
2. If you stop a generation partway, you are not charged for the partial work.
3. If you think you were charged wrongly, you can dispute it through Google One support.

What Google does not tell you is how many tokens a given request costs, or how many you have left in plain numbers before you start. The interface shows a usage meter rather than a figure. On my AI Pro plan, one Saturday afternoon of heavy iteration (three games, around forty follow-up edits) used roughly a third of the weekly meter. The free tier's "limited" allowance will be a fraction of that, so if you plan to build more than one or two games a week, the $4.99 Plus tier is the realistic entry point, and it comes with the other Plus perks I covered in the Gemini 3.8 Live post.

From draft to public: how publishing works

Nothing you make in Playground goes public by accident. Every new game starts as a Draft that only you can see. From there you have four visibility settings:

• Draft: only you.
• Private: saved but not shareable.
• Unlisted: anyone with the link can play, but it does not appear in the gallery.
• Listed (Discoverable): it appears in the Explore gallery for everyone.

Before a game is Listed, Google runs it through an automated check against its Community Guidelines. Other players can report a game with a flag icon, and if Google removes one for a policy violation, you can appeal through a dedicated appeals center. If you edit a published game, the edits stay in draft until you publish again, and the old version keeps running in the meantime.

Leaderboards are only available on single-player games that are publicly Listed. You can also create invite-only group leaderboards with up to 100 members, a nice way to run a competition among friends without putting the game in front of strangers. There is an opt-out for global leaderboards in your profile settings.

One privacy note: your Play Games profile shows your activity, including games you have liked and created, under your gamer name. If you care about that sort of footprint, go through the profile settings before you publish anything, the same way you would check your AI training settings.

Playground vs Roblox Build vs Meta Horizon Create vs Unity Spark

Playground is not the only "type a game" tool, but as of today it is the only one from a major company that most people can actually open.

Roblox Build was announced on 16 July 2026 as a new creation tab inside the Roblox mobile app. It shares its back end with Roblox Studio, so a game started on a phone can be finished on a PC. The catch is availability: the public alpha began on 28 July in New Zealand only, open to age-checked users aged 9 and up. Roblox says more regions are coming "over the coming months", and as of this writing it has not reached the US. A base version is free, with paid options for heavy users.

Meta Horizon Create and Horizon Studio were announced on 25 September 2026. Create is a mobile app that generates 2D or 3D games from a description or a template, and Studio is a browser tool for fine editing. Finished games can appear inside Facebook, Instagram and Horizon, and some can be played straight from an Instagram clip. Access is by waitlist, and Meta is testing with its existing creator community first. No price and no age policy have been published.

Unity Spark is the professional end of this story. Unity announced it alongside Playground as a no-code, prompt-driven game maker that uses real Unity assets and a web-based editor, and it is due to integrate with Playground "later this year" so that more advanced games can be published to the same gallery. It is heading into a closed beta with a waitlist open now. Unity's CEO was careful to say Spark is not about "one-shotting" a game from a single prompt, which is a fair description of where Playground currently sits.

So the practical comparison today is lopsided. In the US and over 18? Playground is the one you can try this afternoon. A teenager? Roblox Build is being designed for you, but you are waiting on a regional rollout. Audience on Instagram? Horizon Create is the one to watch.

Should you bother?

Yes, if you fit the entry requirements and you have ever had a game idea you could describe in a sentence. Playing other people's games is free, and the Explore gallery is already full of oddly charming five-minute games that feel like the early Flash era with better art.

For creating, my recommendation splits three ways:

• Just curious: join the free waitlist and play the gallery while you wait. Do not pay for a plan only to try this.
• Already paying for Google AI Plus or Pro: you have immediate access and there is no reason not to spend an evening on it. It is the most fun I have had with a Gemini feature since I started comparing the big three assistants.
• Want to make something people will actually play repeatedly: wait for Unity Spark. Playground's games are charming prototypes, not products, and the floaty 3D physics in my racing test show how far it has to go.

The bigger picture is that "AI games" is about to become a category, the way "AI images" did two years ago. Google got there first with something you can open in a browser. If you want to understand what everyone will be talking about by Christmas, an hour in Playground is the cheapest way to find out.

ChatGPT Intelligent UI Explained: What the New Visual Answers Do, How to Turn Them Down, and Why Free Users Just Got GPT-6

Contents

1. The short version
2. What Intelligent UI actually is
3. Five things I asked it to build
4. Which plan gets it, and which model you are talking to now
5. How to turn it down
6. ChatGPT vs Gemini vs Claude: who does visual answers best
7. When a visual answer helps, and when it gets in the way
8. What to watch over the next few weeks
9. The takeaway

If you opened ChatGPT this week and got a tabbed diagram instead of a paragraph, you have met Intelligent UI. OpenAI switched it on for paid plans on October 7 and for Free and Go accounts on October 8, and it came bundled with a bigger change that most of the coverage skipped: GPT-6 is now the model answering you in the normal Chat tab, on every plan, including the free one.

I spent two days with it on the web and on my phone. Here is what it does, how to calm it down if you preferred plain text, and how it compares with what Google and Anthropic already ship.

The short version

Intelligent UI means ChatGPT now decides, per question, whether to answer with text, a chart, a side-by-side comparison, a diagram you can click through, or a small working tool like a calculator. You do not turn it on. It just happens when the model thinks a visual will explain something better than words.

Three things worth knowing before you read on:

1. It is tied to GPT-6. Plus, Pro, Business and Enterprise get GPT-6 Sol in Chat. Free and Go get GPT-6 Luna. Both support Intelligent UI.
2. It only works in the Chat tab. Not in Work, not in Voice, and not when a Pro subscriber picks the top "Pro" reasoning level, which runs GPT-6 Astra.
3. There is no off switch, but there is a "Simple" layout setting that cuts most of it out. I cover the exact clicks below.

What Intelligent UI actually is

OpenAI's own description is that ChatGPT can now "answer quickly with fully interactive user interfaces." Strip the marketing and it works like this. The model has access to a library of ready-made visual components: tables, tabs, charts, maps, sliders, checklists, buttons, forms. As it writes an answer, it can reach for one of those instead of a paragraph, and ChatGPT draws the component on screen while the text is still streaming. OpenAI says the model was trained to judge when interactivity is actually useful, so a quick factual question should still get a one-line text reply.

That is the important mental model. This is not ChatGPT generating an image of a chart. It is ChatGPT assembling a small piece of real interface, with buttons that do things, inside the chat. Some of those pieces remember their state: tick three items on a checklist it built, refresh the page, and the ticks are still there. They do not carry over to a different conversation.

It also arrives with a speed change that is easy to miss. GPT-6 in Chat can start answering while it is still reasoning or searching, then fold the extra findings in as it goes. OpenAI's own internal figure is that GPT-6 Instant starts answering web-search questions 44 percent sooner than GPT-5.6 Instant did, and it matches what I saw: the first words land fast and the visual fills in behind them.

Five things I asked it to build

I ran the same five prompts on a Plus account (GPT-6 Sol) and on a logged-in Free account (GPT-6 Luna), on the website and in the iPhone app.

"Explain the Monty Hall problem." This is the demo OpenAI itself leans on, and it is the best case. On Plus I got three doors I could click, a running tally of wins for "switch" versus "stay," and a two-sentence explanation underneath. After twenty clicks the tally sat at roughly two to one in favour of switching, which is exactly the point of the puzzle. The Free account produced the same interactive doors. This is the kind of thing a paragraph genuinely cannot do.

"Split a $184.60 restaurant bill between four people, one of whom only had a drink." On Plus I got a working bill splitter with a field per person, a tip slider and a total that updated live. On Free I got a plain table with the maths done for me. Both were correct. The tool version was nicer, but the table was faster to read.

"Plan a Sunday roast for 2pm." Instead of a recipe list I got a timeline with the oven steps laid out backwards from 2pm, each one tappable to expand. I have cooked a lot of roasts and would still use this.

"Compare a 14-inch MacBook Pro with a Dell XPS 14 for a student." Side-by-side comparison card with tabs for price, weight, battery and ports. Useful layout, but the numbers inside it were the model's numbers, not live retail prices, so check the shop pages.

"What is GDP?" Plain text, four paragraphs, no visual at all, on both plans. That is the right call, and it tells you the model is not forcing visuals in everywhere just because it can.

Across the five prompts, Plus went visual on four and Free went visual on three. The Free answers were noticeably simpler when they did go visual, which fits the gap between Sol and Luna that I wrote about in the GPT-6 Sol and Luna explainer.

Which plan gets it, and which model you are talking to now

When I wrote about Sol and Luna two weeks ago, neither model was in the Chat tab yet. Plus subscribers were still on GPT-5.6 Sol in Chat and Free users were on GPT-5.6 Luna. That changed this week, and this is the part that matters more than the visuals.

A few details from OpenAI's help pages that the table cannot fit:

• On Plus, picking "GPT-6" in the model picker gives you Sol at the Instant, Medium and High reasoning levels. Extra High and Pro are not included on Plus.
• On Pro ($100 and $200 a month), every reasoning level is available, including "Pro," which runs GPT-6 Astra. Astra does not support Intelligent UI, so if you switch to the top level the visuals stop. I covered where Astra fits in the GPT-6 Astra post.
• GPT-6 replaces "Latest" in the picker. If you had explicitly selected GPT-5.6 Sol before, you stay on it until you change it.
• The models behind the Work tab and Codex did not change in this release.

So the practical upgrade for a Free user is bigger than it looks. You went from GPT-5.6 Luna to GPT-6 Luna, and you got the interactive answers, for nothing. That is a sharper deal than Google offered its free users this week, where the direction went the other way (I wrote up the free Gemini downgrade on Wednesday).

How to turn it down

OpenAI does not offer a true off switch, and the help article says so plainly: the setting "reduces" visual and interactive elements, and some may still appear. Here is what actually works, in order of how much it helps.

1. The Simple layout setting (web). Click your profile picture, open Settings, then Personalization, find "Layout and Visuals" and set it to "Simple." In my testing this killed the tabs, sliders and mini tools. Plain comparison tables still showed up, which I was fine with. The setting is account-level, so it applied to my phone too, even though the phone app does not show the toggle yet.

2. Custom Instructions. In the same Personalization screen, add a line like "Prefer plain text answers. Use a table only for comparisons." The model treats this as guidance rather than a rule, but combined with Simple mode I did not see another interactive element for a day.

3. Ask in the chat. "Just give me the text" works for the rest of that conversation. It does not carry to the next one.

The same levers work in reverse: "show me this as a diagram" pushes the model toward a visual it might otherwise skip.

ChatGPT vs Gemini vs Claude: who does visual answers best

OpenAI is not the first to do this. Google announced "generative UI" for the Gemini app in November 2025 as two Labs experiments, Dynamic View and Visual Layout, and Anthropic's Claude has built interactive "artifacts" in a side panel since 2024. The difference is in how each one triggers and who gets it.

Having used all three this week, here is the ranking for an ordinary user.

ChatGPT wins on reach. Intelligent UI is automatic, it is on the free plan, and it lives inline in the conversation rather than in a separate panel. The quality on Sol is high. On Luna it is fine.

Gemini has the most ambitious version and the least reliable one. Dynamic View writes a small custom web page for each answer, which produces the best-looking results of the three when it works. But it is a Labs experiment that has come and gone since launch. It showed up for me on the desktop site this week and not in the Android app, and users on Google's own support forum report it missing entirely. Hard to recommend something you cannot count on seeing.

Claude is the most controllable. Artifacts open in their own panel, you can ask for one explicitly, and nothing appears unless the task is big enough to warrant it. That makes Claude the calmest for long text work, but casual questions never get the "oh, that is handy" moment ChatGPT now delivers by default. For the full three-way picture beyond visuals, start with my Claude vs ChatGPT vs Gemini comparison.

When a visual answer helps, and when it gets in the way

After two days the novelty wears off and a pattern emerges. Intelligent UI is excellent for anything where you would otherwise open a second app: a calculator, a calendar, a map, a spreadsheet. It is good for explanations that depend on a picture, like probability or mechanics. It is a net negative for anything you want to copy somewhere else, because a tabbed card does not paste into an email, and for quick factual questions where the visual adds scanning for no gain.

Two habits made the feature better for me. First, say the format you want up front: "as a table," "as a checklist," "just text." The model nearly always complies. Second, when it builds a tool you like, ask it to "keep this tool and add a column for X." It edits the component in place instead of starting over, and that is where it starts to feel like software rather than a chatbot.

One caution that applies to every one of these products: the numbers inside a generated chart are only as good as the model's knowledge. A beautiful laptop comparison with last year's prices is still wrong. Treat visuals as a layout, not as a source.

What to watch over the next few weeks

The version you see today will drift, because OpenAI tunes these things after launch. Three things I am watching:

• Whether Free users keep GPT-6 Luna. OpenAI's announcement gives it to Free and Go with no stated cap on routine text chats. Free models have been quietly swapped before, which is the whole subject of my nerfed AI post, so check the model name in your picker occasionally.
• A real off switch. The "Simple" setting only reduces visuals. Given how loud the feedback on OpenAI's own forum has been about interface changes, I expect a proper toggle to appear.
• Whether Google answers. Dynamic View is Google's direct equivalent and it is still a Labs experiment. If OpenAI's version lands well, expect Google to promote its own to a default.

The takeaway

Intelligent UI is the most useful change ChatGPT has made to the actual chat window in a long time, and the fact that it ships with GPT-6 for every plan, including Free, is the real headline. The visuals are good for calculators, timelines, comparisons and anything you would normally draw on a napkin. They are a mild nuisance for plain writing, and the fix is one setting under Personalization.

My recommendation: leave it on for a week, say "as a table" or "just text" when you know what you want, and only flip to Simple if you find yourself fighting it. If you are on the free plan, you just got a better model and a better interface without paying anything. That does not happen often.

Free Gemini Gets Downgraded on October 9: What Flash-Lite Can Actually Do, and Whether $4.99 Fixes It

Contents

• The short version
• What changes on October 9, plan by plan
• I used Flash-Lite as my only model for three days
• The usage limits are changing too, and that part is sneakier
• What free ChatGPT and free Claude give you right now
• Is Google AI Plus worth $4.99? Is Pro worth $19.99?
• Three ways to keep better answers without paying Google
• My recommendation

If you use Gemini without paying, the app you open on Friday will be a different product from the one you used on Thursday. From October 9, 2026, Google is pulling the Flash and Pro models from free accounts and leaving one option: Gemini 3.5 Flash-Lite. The $4.99 AI Plus plan loses Pro on the same day.

Google announced this through a support article called "Changes to Gemini model access and limits", and most coverage has just repeated the table. I wanted to know the thing that matters: how much worse do your answers get, and is five dollars a month the fix? So I used Flash-Lite as my only model for three days, then ran the same questions through free ChatGPT and free Claude.

The short version

Free Gemini goes from three models to one. Flash-Lite is a fast, cheap model built for short everyday tasks, and for those it is fine. For anything with more than two or three steps, a long document, or some maths, it is a clear step down from the Flash model you have today.

If you mostly ask quick questions, rewrite emails and summarise articles, stay free and you will barely notice. If you lean on Gemini for homework-grade reasoning or long PDFs, AI Plus at $4.99 gets Flash back and is the cheapest fix. Pro, the model that actually impresses people, now needs AI Pro at $19.99. Or do what I ended up doing: keep Gemini for the Google-account stuff and send the hard questions to a different free chatbot.

What changes on October 9, plan by plan

Google's support page lays out four tiers. I checked the prices on gemini.google today, so these are current as of October 7, 2026.

Without a plan ($0). Today you get Gemini 3.6 Flash as the default plus "varying access" to Gemini 3.1 Pro, which in practice means a handful of Pro answers before it drops you back to Flash. From October 9 you get Gemini 3.5 Flash-Lite and nothing else.

Google AI Plus ($4.99 a month). Keeps Flash-Lite and Flash, loses Pro. This is the plan Google cut from $7.99 to $4.99 in June, with 400GB of storage, video generation and Gemini inside Gmail. After the change it is really "Flash plus storage".

Google AI Pro ($19.99 a month). Keeps all three models and gains Deep Think, the slow multi-pass reasoning mode that used to be Ultra-only. This is now the lowest tier where Gemini 3.1 Pro is an everyday model rather than a treat.

Google AI Ultra (from $99.99 a month). No restrictions and the highest limits. Gemini 4 Argon, announced on September 30, is going to Ultra subscribers and paid API customers first, and as of today it is not in the regular Gemini app.

One detail worth knowing: the same page notes that users under 18 went through a separate round of model restrictions back in July. If a teenager in your house uses Gemini, their version already changed.

I used Flash-Lite as my only model for three days

You can already pick Flash-Lite in the Gemini app, so I set it as my model on Saturday morning and did not touch Flash or Pro until Tuesday. I kept a list of 24 prompts covering what I actually use a chatbot for, then ran the same 24 through Flash afterwards.

Where Flash-Lite was fine. Rewriting a stiff email in a friendlier tone. Summarising a 1,500-word article into five bullets. Explaining a term in a letter from my bank. Turning a messy voice-note transcript into a shopping list. Answering "how long do I boil an egg for a runny yolk". It was quick, usually two to three seconds to a complete answer, and I could not tell its answers apart from Flash. Twenty-one of my 24 prompts landed here.

Where it fell over. Three prompts, and they were the three I cared most about.

A 38-page rental agreement as a PDF. I asked for every fee and the conditions attached to each. Flash-Lite found the rent, the deposit and the late fee, then stopped. Flash also found a cleaning charge and a key replacement fee buried in a schedule at the back. With only Flash-Lite, you would believe the document had three fees.

A multi-step budget question: given four income lines and a list of expenses, how much could I save each month while keeping a $500 buffer. Flash-Lite got the arithmetic right and ignored the buffer entirely. Flash handled it.

A spreadsheet formula that looks up a value across two sheets. Flash-Lite referenced a sheet that did not exist. Flash gave me one that worked.

This matches how Google's own help page describes the models: Flash-Lite is "an efficient workhorse model designed for speed" for "everyday tasks like summarization and brainstorming", Flash "balances speed and reasoning", and Pro "excels at complex math and coding". After three days I think that is accurate. Flash-Lite is not dumb. It is shallow. It does the first obvious pass and does not go back to check.

If you read my piece on whether your AI is getting dumber, this is a different thing. That was about models changing under you without notice. This is an announced swap to a smaller model, and the drop is predictable.

The usage limits are changing too, and that part is sneakier

The model change got the headlines, but the same support page describes a second change that will annoy more people. Gemini is moving to compute-based usage limits that "refresh every 5 hours until you reach your weekly limit".

In plain English: there is no longer a simple count of messages per day. Every request costs an amount of usage based on how complex the prompt is, which features it uses and how long the chat has become. Google lists media generation, Deep Research, the Pro model, extended thinking and Deep Think as the expensive ones. You get a bucket that refills every five hours, and a weekly bucket that does not.

Two consequences. Long chats get pricier as they go, so the "one giant conversation for everything" habit hits the wall faster. And paid tiers are now defined by multipliers: AI Plus gets 2x the free limit, AI Pro 4x, Ultra up to 20x, per Google's plan page. Google does not publish the base limit, so 2x is twice an unknown.

One small mercy: the help page says paid users who hit their limit can continue the chat with Flash-Lite. So Flash-Lite is now the fallback model for everyone. Seen that way, the free tier makes sense from Google's side. Free users are being moved onto the fallback full time.

What free ChatGPT and free Claude give you right now

Google's two main rivals are not making the same move, at least not this week. I checked both pricing pages on October 7.

ChatGPT Free gives you "unlimited text chats with GPT-5.6 Luna", plus limited uploads and image creation. Luna is OpenAI's small fast model, closer to Flash than Flash-Lite, and there is no cap on plain text. My GPT-6 Sol and Luna explainer covers which plan gets which model.

Claude Free gives you Sonnet and Haiku, with web search, file creation and memory across conversations. Sonnet is a mid-tier model, not a lite one, and it is the strongest free model of the three. The catch is a tighter usage cap that resets in windows. My Claude Fable 5.1 explainer covers what the paid tiers add.

Gemini Free after October 9 gives you Flash-Lite only, with compute-based limits. It is the only free tier of the three built around a lite model.

I ran my three failed prompts through both rivals. Free Claude found all five fees in the rental PDF. Free ChatGPT found four, missing the key replacement fee. Both got the budget question right, buffer included. Both produced a working formula. None of this makes Gemini a bad product. It does mean that, for free users, it stops being the best one. My Claude vs ChatGPT vs Gemini comparison has the bigger picture, and I will update it once the change lands.

Is Google AI Plus worth $4.99? Is Pro worth $19.99?

It depends on what you use Gemini for, so here is how I would decide.

Stay free if your prompts are short, you rarely upload documents, and you do not care which model answers. Flash-Lite will serve you, and the five-hour refresh will barely register.

Pay $4.99 for AI Plus if you want Gemini specifically, because it lives inside Gmail, Drive and your Android phone, and you want Flash back. At five dollars with 400GB of storage, this is the best-value plan Google sells. If you already pay for extra storage, it is close to free. Just know you are not getting Pro.

Pay $19.99 for AI Pro if you do real work in Gemini: long documents, coding, research, anything where a wrong answer costs you time. It is the only tier where Gemini 3.1 Pro is the normal model, and it now includes Deep Think, 5TB of storage and the Jules coding agent. It costs the same as ChatGPT Plus and Claude Pro, and at $19.99 all three are good. Pick the one whose apps you already live in.

Skip Ultra unless you know exactly why you need 20x limits or want Gemini 4 Argon the day it arrives. For a general consumer, $99.99 a month buys nothing you will notice.

Three ways to keep better answers without paying Google

Split your questions. This is what I settled on. Gemini stays on my phone for anything that touches my Google account: my flight time, an email thread, a calendar entry. The PDF analysis, the maths and the code go to free Claude or free ChatGPT in a browser tab. It costs nothing and the hard questions get a better model.

Check the student offer. Google's release notes from August say eligible college students get a year of Google AI Pro free in the US, or AI Plus in 140 other countries. If you have a .edu address or the local equivalent, that is Pro-level Gemini at no cost. The offer has a sign-up window, so check before you assume you qualify.

Watch what stays free. As of today, Google's plans page still lists image generation, Deep Research, Gemini Live, Canvas and Gemini Notebook under the free tier. The support article says Deep Research draws on the heavier usage budget, but does not say it is being removed. The two pages do not fully agree yet, and I will update this post once I can see what the free app actually offers after October 9. If you talk to Gemini out loud, my Gemini Live explainer covers how that mode works.

My recommendation

For most people, the honest answer is: stay free, but stop treating Gemini as your only chatbot. Flash-Lite is good enough for the quick stuff, and the quick stuff is most of what anyone does. Keep a free Claude or ChatGPT tab for the moments that need a proper model.

If you were already on the fence about paying Google, AI Plus at $4.99 is the easy yes, mainly because the storage alone is worth it and Flash comes along for the ride. AI Pro at $19.99 is for people who do real work in Gemini, and it holds up well against the $20 plans from OpenAI and Anthropic.

What I would not do is pay out of fear of losing something. The free tier is getting worse, but in a predictable way, and there are two other free chatbots a tab away that have not followed Google down this road. Yet.

Can Your Gaming PC Really Run a 125 Billion Parameter AI? Mixture of Experts, Explained in Plain English

Contents

• What people are actually running
• Mixture of experts, explained with a hospital
• The "VRAM barrier" trick: hot and cold experts
• What happened when I tried it on my own PC
• Which hardware can run what
• Should you bother?
• The verdict

Over the weekend one story climbed to the top of Hacker News and stayed there: somebody ran a 125 billion parameter AI model on a single RTX 4090 gaming card at more than 100 words a second. A year ago that size of model meant a rack of server GPUs. The phrase doing the rounds is "the VRAM barrier is dead", and I have now seen it in three separate GitHub projects, a viral post on X and a dozen comment threads.

I spent two days pulling the claim apart and trying it on my own desktop. The short version: the trick is real and clever, but the hardware it actually needs is not the hardware the headline implies. If you own a gaming PC, this post tells you whether you are in the club, and if not, what the one upgrade that gets you there would be.

What people are actually running

The model is Qwen3.8-Flash-Next, from Alibaba's Qwen team. It was released as open weights on 26 August 2026 as a research preview of the architecture Alibaba plans to use for Qwen 4. Anyone can download it from Hugging Face or ModelScope for free under the Qwen Community License, and it already has one of the busiest pages on Ollama. (If "open weights" and "Hugging Face" are new to you, my Hugging Face explainer covers both in ten minutes.)

The headline numbers from the model card:

• 125 billion total parameters, plus a separate 51 billion parameter "n-gram" lookup table.
• 6 billion active parameters per word. You will see this written as "125B-A6B".
• 512 experts, of which 10 plus one shared expert work on each word.
• A 262,144 token context window, which is roughly 400 pages of text in one conversation.
• Text, image and video understanding in one model.

The full-precision download is about 360GB, which is obviously not going on a gaming PC. What changed over the last two weeks is that a string of hobbyists published working recipes for squeezing it onto a single 24GB graphics card. One project measured 110 to 120 words a second on an RTX 4090. Another runs it on an older RTX 3090 at 43 to 51 words a second. A third got a quarter-million-token conversation going on a 4090 at a more modest 21 words a second. Those are the posts you have been seeing.

Mixture of experts, explained with a hospital

To understand why a 125B model can run at the speed of a small one, you need one idea: mixture of experts, usually shortened to MoE.

Picture a hospital with 512 specialists on the payroll. When a patient walks in, a triage nurse reads the chart and sends them to the ten specialists who are relevant, plus one general physician who sees everyone. The other 501 specialists are still employed, still in the building, but they do no work on that patient.

That is exactly how Qwen3.8-Flash-Next handles each word. A small routing network (the triage nurse) looks at the word, picks 10 of the 512 experts plus the shared one, and only those 6 billion parameters do any maths. Total staff: 125 billion. Staff actually working on any given word: 6 billion.

The consequence is a model that thinks with the speed of a 6B model but knows what a 125B model knows. That is the whole appeal.

Compare that with a dense model such as Qwen3.8 27B, the one I tested in my quantization guide. Dense means every single one of the 27 billion parameters works on every single word. It is more work per word than Flash-Next does, even though the file is far smaller.

So where is the catch? The hospital still needs a building big enough for all 512 specialists. In computer terms: every expert has to be loaded into memory somewhere, even though most of them sit idle. The maths is cheap. The memory is not.

The "VRAM barrier" trick: hot and cold experts

The old rule of local AI was simple: the model file has to fit inside your graphics card's memory (VRAM), with a little room to spare. A 24GB card comfortably runs a 17GB file. It does not run an 83GB file, and that is roughly what Flash-Next weighs at the 3-bit quantization the 4090 project used. Even the aggressive 1-bit version is 75GB. By the old rule, a 24GB card cannot run this model at all.

The new trick exploits the fact that experts are not used equally. Some get picked constantly. Many get picked rarely. So the software does this:

1. Hot experts go on the GPU. The RTX 3090 project keeps 32 of the most-used experts per layer in VRAM, along with the parts of the model that run on every word.
2. Cold experts go in system RAM. Ordinary DDR4 or DDR5 memory holds the thousands of experts that are only needed now and then, and the CPU does the maths for those.
3. The rarest experts stay on the SSD and are streamed in when a conversation really needs them.

On top of that sits a second trick called multi-token prediction (MTP). Instead of producing one word at a time, the model drafts several words ahead and then checks them in a single pass, which the Unsloth team measured at 1.3 to 1.7 times faster output with no change to what the model would have said.

Here is the part the headline leaves out. The graphics card is not doing this alone. The machine that hit 110 words a second had 192GB of DDR5 system RAM. The RTX 3090 recipe lists 64GB of RAM as the minimum and 96GB as the amount that keeps every expert in memory. The barrier did not die. It moved from your graphics card to your motherboard.

What happened when I tried it on my own PC

My test machine is a desktop with 32GB of RAM and no dedicated graphics card, the same box I used for the quantization tests. I knew it was going to be a stretch. I wanted to see how it fails.

I downloaded the smallest file available, the 75GB 1-bit build, and pointed LM Studio at it. LM Studio flagged the file as too large for my memory before I even clicked load, and when I overrode the warning, it spent eleven minutes paging the model in from the SSD and then produced text at well under one word a second. A single paragraph took longer than making a coffee. That is not a usable experience, and it is not LM Studio's fault. There is no amount of software cleverness that fits 75GB of experts into 32GB of RAM.

Then I did what I should have done first and reloaded the 4-bit Qwen3.8 27B dense model. It ran at just under five words a second on the processor, exactly as it did in September. On a machine like mine, the 27B is the real option and Flash-Next is a spectator sport.

That experience shaped the table below.

Which hardware can run what

Tier 1: a normal laptop, 8GB to 16GB of RAM, no graphics card. Flash-Next is out of the question. Stick with the 4B to 12B models from my local AI starter guide, or run a small model on your phone. They are good, and they are fast.

Tier 2: a 32GB desktop, no graphics card or a small one. This is me. The 27B dense model at 4-bit works, at about five words a second on the CPU. Flash-Next does not fit.

Tier 3: a 16GB graphics card, such as an RTX 5060 Ti. The 27B at 3-bit fits in VRAM and runs at around 46 words a second according to one detailed write-up. That is a very nice local assistant. Flash-Next still does not fit.

Tier 4: a 24GB graphics card (RTX 3090, 4090 or similar) plus 64GB of RAM. Welcome to the club, barely. The 3090 recipe runs Flash-Next at 43 to 51 words a second with 128K of context. Be aware this is a weekend project: Linux, Docker, a 130GB download and a first start that compiles code. 96GB of RAM makes it comfortable.

Tier 5: a 24GB card plus 128GB to 192GB of RAM, or a Mac with 128GB or more of unified memory. This is where the 110-plus words a second numbers come from. On Apple silicon the Ollama MLX build is a 105GB file, so 128GB is the realistic floor, and Mac owners at that level report anywhere from 50 to well over 100 words a second depending on the chip.

Notice what separates Tier 3 from Tier 4. It is not a better graphics card. It is RAM. When I wrote how much RAM you actually need, I said almost nobody had a reason to go past 32GB. Local MoE models are the first mainstream reason for a home user to buy 64GB or more.

Should you bother?

Three honest considerations.

Quality. Alibaba's own benchmark table puts Flash-Next ahead of Claude Opus 4.6 Max on SWE-bench Pro (62.5 versus 53.4) and CoWorkBench (73.9 versus 68.2). Those are Alibaba's numbers on Alibaba's chosen tests, and independent evaluation of the research preview is still thin, so treat them as a claim rather than a fact. More telling for home use: the 4090 project ran its own checks on the squeezed 3-bit version and got 28 of 30 coding functions correct and 115 of 140 general knowledge questions right. Squeezing a model does cost something.

Privacy. Nothing leaves your machine. If you have followed my guide on stopping ChatGPT, Claude and Gemini from training on your chats, you already know how much of a hassle that is with cloud services. A local model skips the problem entirely.

Cost. This is where the romance fades. A 24GB card plus 64GB of RAM is a four-figure build. Alibaba's hosted version of the same model is priced at $0.16 per million input tokens and $0.47 per million output tokens. You could send it the entire text of this blog every day for a year and not spend what the RAM costs. If your motive is money, cloud wins by a mile. Local wins on privacy, offline use and the sheer fun of it.

One gotcha for anyone who does take the plunge: the model defaults to its highest "reasoning effort" setting, and on home hardware that can turn a one-line question into twenty minutes of thinking. Set it to low or medium before your first prompt.

The verdict

The "VRAM barrier is dead" headline is half right. Mixture of experts really does let a 125B model think at 6B speed, and the hot-and-cold expert trick really does let a 24GB gaming card host it. But the barrier did not disappear. It moved to system RAM, and the 110 words a second machines are carrying 192GB of it.

My recommendations:

• Own a 24GB card already? Do not buy a new GPU. Buy RAM. 64GB gets you in the door, 96GB makes it pleasant.
• Own a 16GB card or a 32GB desktop like mine? Run Qwen3.8 27B at 4-bit and be happy. It is the best local model most people will ever need, and it just works.
• Everyone else? Nothing has changed for you this month, and that is fine. Use the small models locally and the cloud for the heavy lifting.

The exciting part is the direction. Twelve months ago a 125B model was a data centre product. Today it is a weekend project for a well-stocked gaming PC. The trend line points at ordinary desktops, and I will be testing every step of the way.