Quick answer
Grok 4.8 is xAI’s next model, and Elon Musk says it will start a soft rollout to Grok Bot users this week after xAI rebuilt its training and inference software from scratch. It sits alongside two other big stories in WorldofAI’s 12 November news roundup: a line in OpenAI’s API documentation that refers to “models 6.2 and later”, which has fuelled talk of a whole GPT-6.2 family, and an early-access preview of Alibaba’s Qwen 4 Max. None of these is a full public release yet. As of 13 November, OpenAI has not announced GPT-6.2, and Musk himself says Grok 4.8 “isn’t incredible yet”. The practical move is to keep working with the tools you have and test new models on a real task when they actually arrive.
Key takeaways
- Grok 4.8: pre-training is finished and reinforcement learning is under way, according to Musk. A soft rollout to Grok Bot users was promised for this week.
- GPT-6.2: evidence so far is one documentation note and hints from OpenAI staff. It is a leak, not a launch.
- Qwen 4 Max preview is available in early access, and Qwen Code 1.0 Preview is now open source, although outside code contributions are paused until after December 2026.
- Claude limits: Theo’s logged numbers suggest the cheapest-looking model is not always the one that gets the most work done in a week.
- Anthropic has reportedly started its IPO roadshow and is giving paid Claude users a free usage reset to redeem before 10 December.
- Google added a /swarm command to Antigravity, a fast screen-flow generator in Stitch and a group tool builder called Tabletop.
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Make Your AI Coding Plan’s Usage Limits Last All Week — match the model to the job and track finished work, so your subscription doesn’t run dry by Wednesday.
- How to Prototype an App’s Screen Flow With AI Before You Build — map your screens, generate a clickable flow and test it with real people before any code is written.
In this guide
- Watch: the source video
- What happened in AI this week, at a glance?
- What do we actually know about Grok 4.8?
- Is GPT-6.2 really about to launch?
- What is new with Qwen 4 Max and Qwen Code 1.0?
- What did Theo’s Claude usage numbers show?
- What is happening with Anthropic’s IPO and the free Claude reset?
- Which new Google tools are worth a look?
- What changed for AI agents and teammates?
- Which other models and demos stood out?
- What is confirmed and what is still a rumour?
- How should a small business act on this week’s news?
- Frequently asked questions
- The bottom line
- Related posts
If you only have five minutes for AI news this week, this is the version to read. WorldofAI packed more than 40 minutes of updates into its latest roundup, from Google’s next models to new agent features across Claude, ChatGPT, Cursor and Grok Bot.
We have already covered the Google side in depth. Our Gemini Krypton leak roundup and Gemini 4 Carbon article explain those models, so here we focus on what is genuinely new in this video: Grok 4.8, the GPT-6.2 leak, Qwen 4 Max, and the smaller launches that could change how you work. Every leak and rumour is labelled as one.
Watch: the source video
What this video covers: WorldofAI’s 12 November AI news roundup: early Gemini 4 Krypton tests, hands-on Gemini 4 Carbon demos, the GPT-6.2 documentation leak, Gemini Decision 1, the Qwen 4 Max preview and Qwen Code 1.0, Theo’s Claude subscription numbers, Anthropic’s IPO and office tools, MiniMax H3 Image, Antigravity /swarm, Grok Bot 1.0, Stitch, Tabletop, Grok 4.8 and more.
Video: WorldofAI (@intheworldofai), an AI news roundup published on 12 November 2026. The video is sponsored by Mammouth, an app that puts several AI models under one subscription, and the creator also promotes his own WoAI Bench benchmarking tool. This summary and the opinions in it are our own; test results and leaks are the creator’s or other people’s reports and should be checked against each company’s announcement.
What happened in AI this week, at a glance?
Here is every story from the video that matters for most readers, with its status on the day it was published. Swipe the table sideways on a phone.
| Story | What it is | Status (12 Nov) | Our take |
|---|---|---|---|
| Grok 4.8 | xAI’s next Grok model on a rebuilt software stack | Soft rollout promised this week | Wait for real tests |
| GPT-6.2 family | OpenAI’s next models, possibly Astra, Sol, Terra and Luna | Leak / hints only | Don’t plan around a date |
| Qwen 4 Max | Alibaba’s next flagship model | Early-access preview | Worth a trial on coding tasks |
| Qwen Code 1.0 | Alibaba’s coding agent tool | Open source preview | Read-only for contributors until after Dec 2026 |
| Gemini 4 Krypton | Google’s next large model, spotted in testing | Leak / early tests | See our Krypton roundup |
| Gemini 4 Carbon | Google’s strongest released model | Rolling out | See our Carbon article |
| Gemini Decision 1 | A model that answers yes/no-style questions with a confidence score | Available in Google AI Studio | Interesting for high-volume checks |
| Anthropic IPO | Share offer roadshow reported | Reported, not complete | Watch, don’t speculate |
| Free Claude reset | One extra usage reset for paid plans | Live until 10 December | Save it for a heavy week |
| Antigravity /swarm | Several AI agents planning, coding and checking in parallel | Launched | Useful for bigger builds |
| Stitch Rapid Design | Multi-screen app designs in seconds | Experimental | Great for early prototypes |
💡 Tip: You don’t need to act on every line. Pick the one story closest to how you already earn money, and park the rest until it affects you.
What do we actually know about Grok 4.8?
Grok 4.8 is the next version of xAI’s Grok model, and so far everything we know comes from Elon Musk’s own posts. The creator reads them out in the video, and the main points are:
- A rebuilt foundation. Musk says the team dropped some of its old programming tools and rewrote its reinforcement learning and inference software from scratch in C and C++. That took longer than planned, which is why the launch slipped.
- Training status. Pre-training is complete, the model is now in reinforcement learning, and Musk says early evaluation scores are improving. The team also spent more time cleaning its fine-tuning data.
- Rollout. A soft rollout to Grok Bot users was planned for this week. No public API date or price was given.
- Expectations. Musk says Grok 4.8 will “bear no resemblance” to Grok 4.7, but also admits it is not incredible yet. His argument is that the new foundations will let it improve quickly.
🔎 Unconfirmed / rumour: Everything about Grok 4.8’s quality is still a claim from the company’s owner. There were no independent benchmarks when the video was made. Treat “bears no resemblance to 4.7” as a promise to test, not a result.
What it means for you: if you already use Grok Bot, you may see Grok 4.8 appear as an option soon. Run it on two or three of your normal tasks and compare it with what you use today before you switch anything important. Our founder litmus test for new AI models gives you a quick way to do that.
Is GPT-6.2 really about to launch?
Possibly, but not yet. The evidence is a small change to OpenAI’s API documentation. In the section on “compaction” (how long conversations are squeezed down so the model can keep going), a note now warns that something behaves differently “on models 6.2 and later”.
The creator points out that the note says models, plural, which suggests a family of GPT-6.2 models rather than one. Names that have appeared in earlier leaks include Astra, Sol, Terra and Luna. On top of that, Tibo, who works on OpenAI’s Codex, posted that “Thursday is feeling codexy” and hinted at a “banked reset”.
🔎 Unconfirmed / rumour: As of Friday 13 November, OpenAI has not announced GPT-6.2, and independent trackers report that it did not ship on Thursday 12 November. A documentation note shows OpenAI is preparing for these models; it does not tell us when they will launch or how good they are.
If you want to know how the early GPT-6.2 previews performed, we covered GPT-6.2 Sol’s Codex tests and GPT-6.2 Astra’s SVG tests when they surfaced.
⚠️ Warning: If you build on OpenAI’s API and use custom compaction instructions, read the updated compaction notes now. The warning says that on 6.2 and later models, replaying a compaction made with custom instructions can cause the API to drop later reasoning steps. It is far easier to adjust before a model switch than after.
What is new with Qwen 4 Max and Qwen Code 1.0?
Alibaba has opened early access to a Qwen 4 Max preview, its next flagship model, and released Qwen Code 1.0 Preview as open source. We tested the first Qwen 4 previews in our Qwen 4 first look, so here is just what is new.
- Where to try it: the creator says the preview is available through Alibaba’s Token Plan, with early access expanding to QwenWork, Qoder, the Qwen app and Qwen Studio.
- What Alibaba is aiming at: deeper knowledge, better world understanding and stronger coding, with a focus on software engineering, front-end work and coding agents.
- Demos: in one side-by-side, Qwen 4 Max built a 3D koi pond in about 30 minutes, while MiniMax M3.1 took about 70 minutes and cost about $6.30. A second demo, a 3D model of the Sagrada Família, stopped early because it hit a token limit.
- Qwen Code 1.0 Preview is open source, but Alibaba says it will not accept community code contributions until some time after December 2026. For now you can read the code and report issues.
💡 Tip: One demo on one prompt tells you very little. If Qwen 4 Max interests you, give it the same real task you gave your current model last week and compare the result, time and cost side by side.
📘 Main article: Want to stretch your current subscription instead of buying another? Our guide How to Make Your AI Coding Plan’s Usage Limits Last All Week turns the next section into a simple weekly routine.
What did Theo’s Claude usage numbers show?
Theo, a well-known developer and YouTuber, logged more than 1.1 million of his own Claude Code requests on the $200-a-month Claude Ultra plan, and the results surprised a lot of people. The creator walks through his breakdown:
| Measure | Opus 5.5 | Fable 5.2 | Note |
|---|---|---|---|
| API-equivalent value per month on Ultra | About $15,000 | About $3,000 | Same plan, same price |
| Share of the weekly limit it can use | Full limit | 50% cap | Fable alone leaves half the limit unused |
| Work per week (Theo’s numbers) | ~1,200 turns, 230 threads (Opus only) | ~880 turns, 230+ threads (Fable first, then Opus) | Fewer turns, slightly more finished threads |
Theo’s point is that “value in dollars” is the wrong scoreboard. Fable 5.2 uses up limits faster per dollar, yet it is more likely to finish a hard job in one go. So the pattern he suggests is: use Fable on the hard problems where finishing in one pass matters, lean on Opus for the long grind, and save Sonnet and Haiku for simple jobs.
The creator also notes that Fable 5.2’s API price fell compared with Fable 5.1, from $10 to $8 per million input tokens and from $50 to $40 per million output tokens.
⚠️ Warning: These are early numbers from one heavy user, and much of his usage was benchmarking. Your mix of tasks will be different. Use them as a reason to measure your own week, not as a rule.
If you are still deciding which Claude model suits which job, see Which Claude Model Should You Use? Haiku, Sonnet, Opus or Fable and our Claude Fable 5.2 review.
Quick win: for one week, note each AI task you start and whether it was finished in one session. That single list will tell you more about which plan and model you need than any benchmark. Our usage limits guide has a ready-made log.
What is happening with Anthropic’s IPO and the free Claude reset?
Anthropic has reportedly launched its IPO roadshow, the series of investor meetings before a company lists its shares. According to the reports the creator cites:
- The offer is reported at 825 million shares priced at $70 to $78, which would value Anthropic at roughly $1.2 to $1.4 trillion.
- Bloomberg reported the offer was already several times oversubscribed, and the Financial Times reported bankers were considering a higher price.
- Together with a private placement, Anthropic and its co-founders could raise around $90 billion. Google, NVIDIA and Microsoft have reportedly committed a combined $26 billion.
- The FT reports Anthropic is planning about $900 billion of compute and infrastructure spending.
⚠️ Warning: These figures come from press reports, not a final prospectus, and they can change before listing day. Nothing here is investment advice. For the bigger picture, see our earlier piece: Anthropic IPO: Why It’s an AI Bubble Test, Plus Kimi K3.1 Pro, Genie 4 and OpenAI’s Maths Claim.
The free Claude usage reset
Anthropic is giving Pro, Max, Ultra and Team subscribers one free reset of their usage limits, covering every model. You can redeem it any time before 10 December at 2pm Pacific Time (10pm UK time). Anthropic says it fixed several bugs that were using up limits faster than intended, and that an earlier Fable-only reset did not help people who had already used up their weekly limit.
💡 Tip: Don’t redeem the reset the moment you see it. Save it for a week when you have a big build or a deadline and know you will hit your limit.
Anthropic also launched Claude Whiteboards, Claude Sheets and Claude Data, plus interactive connectors that open apps such as Adobe, Atlassian and Box inside Claude. We covered the office tools in Greg Isenberg’s live reaction, which also explains ChatGPT’s new business listings.
Which new Google tools are worth a look?
Google shipped four things in this video that are worth knowing about, beyond its models.
Antigravity /swarm
Google’s Antigravity coding app has a new /swarm command. With one command you launch several specialised agents that work in parallel, for example one plans, one writes code and one checks the work, and they coordinate through a shared message board. You can also give each agent a different model. You can start for free, with these features included in the Pro plan. If you try it, our guide on splitting work between AI bots without overlap will help you set clear roles.
Stitch Rapid Design
Google Stitch has an experimental Rapid Design mode, powered by the Gemini Decisions API. From one prompt or screenshot it generates a multi-screen product flow in a few seconds, across seven design systems. You can right-click an element to request changes, edit text directly or use your voice, and generate variations of a single component or a whole screen. Google calls it an early experimental MVP.
📘 Main article: This is the perfect moment to test an app idea before you build it. Our guide How to Prototype an App’s Screen Flow With AI Before You Build walks you through it step by step, whichever design tool you use.
Gemini Decision 1
Gemini Decision 1 is a different kind of model. Instead of writing an answer word by word, it answers many questions about an input in a single pass and gives each a confidence score. You can feed it text, images, audio, video or PDFs and ask up to 1,000 questions at once. Google says it replies in 150 to 250 milliseconds and costs about 10 cents per million input tokens. In Google’s demo it checked 100 delivery photos in 4.5 seconds, answering 16 questions about each one, such as whether the parcel was damaged.
For small businesses, the interesting use is cheap, fast checks inside automations: is this form complete, does this photo show what it should, should this email go to a person? Our guide on setting an AI confidence threshold for human review explains how to decide what score is good enough.
Tabletop
Tabletop is a Google Labs experiment for building small interactive tools with a group, without code. Think a shared road-trip map, a dinner decision wheel, a tournament bracket or a seating chart that everyone can edit. It has a waitlist and is open to US users aged 18 and over.
What changed for AI agents and teammates?
- Grok Bot 1.0 brings shareable AI teammates that work inside Slack and Microsoft Teams, join conversations and can be invited to meetings, learning your team’s ways of working over time.
- Meta’s Muse agent now has its own email address and can send and receive email for you. It is available to paid users in the US and Canada.
- Cursor rebuilt projects to be multiplayer, so your agents and your teammates’ agents can work on the same project together.
- OpenAI’s Dots agent can now connect to several of your computers and choose which one to use, for example building and testing an app on a desktop while you keep working on your laptop.
- ChatGPT Finances is now in the ChatGPT desktop app for US users on Free, Go, Plus and Pro, to analyse spending and build simple money tools.
⚠️ Warning: Every one of these gives an AI agent more reach: your email, your team chat, your other machines or your money. Turn them on one at a time, start with read-only access where you can, and keep a person approving anything that sends, pays or deletes.
Which other models and demos stood out?
| Model or demo | What the creator showed | Status |
|---|---|---|
| Gemini 4 Krypton | Takes roughly twice as long as Carbon and writes nearly twice as much code; a Three.js sailboat took 22 minutes 20 seconds | Early tests of a leaked model |
| Gemini 4 Carbon | Priced at $2 per million input and $10 per million output tokens; the creator found medium reasoning gave the best value | Released |
| MiniMax H3 Image | Open-weight image model; 4K images in about 10 seconds, up to 16 megapixels, up to 16 reference images | Released (50% launch-week discount above 16MP) |
| Nex-N2.6 | Open-source multimodal agent model from Shanghai Innovation Institute; reported 83 on OSWorld and 84.1 on Terminal-Bench 2.1 | Released (benchmarks self-reported) |
| Inkling 2 Small | 1 million token context, multimodal | Free on OpenCode for a limited time |
| Fable 5.2 demos | A AAA-style game, a GTA-style clone and animations in pure JavaScript, shared by other creators | Demos, not products |
| Gemini Robotics 2.5 | Apptronik Apollo 2 humanoid robots sorting car parts with Mercedes; 96% accuracy reported | Proof of concept |
The video ends with Anthropic CEO Dario Amodei’s prediction that diseases such as cancer, Alzheimer’s and heart disease could become “relics of the past” in under a decade. That is a hope about the future, not a forecast you can plan around. The video also covers reports about NVIDIA’s leadership; those were moving quickly, so check NVIDIA’s own announcements for the current position.
What is confirmed and what is still a rumour?
| Claim | Evidence so far | How to treat it |
|---|---|---|
| Grok 4.8 is rolling out this week | Elon Musk’s posts | Company claim. Test it yourself |
| GPT-6.2 is a family of models | “Models 6.2 and later” in API docs | Strong hint, no launch |
| GPT-6.2 launches Thursday | A staff member’s hint; it did not ship on 12 Nov | Rumour |
| Krypton beats Carbon | Creator’s and other testers’ early outputs | Early and informal |
| Qwen 4 Max preview is available | Alibaba’s early-access announcement | Confirmed (preview) |
| Fable gets more done per week | One user’s logged data | Interesting, not universal |
| Anthropic IPO size and price | Press reports (Bloomberg, FT) | Reported, can change |
| Free Claude reset until 10 Dec | Anthropic announcement | Confirmed |
How should a small business act on this week’s news?
- Don’t switch tools on a rumour. Wait until Grok 4.8 or GPT-6.2 is actually in front of you.
- Keep a short test task ready, the same one you can give any new model, so comparing takes 20 minutes rather than a day.
- Measure your own AI week before paying for a bigger plan. Finished tasks matter more than tokens.
- Prototype before you build. Tools like Stitch make it cheap to test an idea with real people first.
- Add agent features slowly, with approval steps for anything that sends, pays or deletes.
Your next step: write down one real task you do every week with AI and keep it as your test prompt. When the next model arrives, run that prompt first. Then use our guides on making usage limits last and prototyping a screen flow.
Frequently asked questions
When will Grok 4.8 be released?
Elon Musk said a soft rollout to Grok Bot users would start in the week of 10 November 2026. No date for a wider release or the API had been given when this was written.
Is GPT-6.2 out?
Not as of 13 November 2026. OpenAI’s API documentation mentions “models 6.2 and later”, but OpenAI has not announced a release.
How can I try Qwen 4 Max?
Through Alibaba’s early-access preview, which the creator says is available via the Token Plan, with access expanding to QwenWork, Qoder, the Qwen app and Qwen Studio.
Is Qwen Code 1.0 open source?
Yes, as a preview. You can read the code and report issues, but Alibaba will not accept outside code contributions until after December 2026.
How do I get the free Claude usage reset?
If you are on a Pro, Max, Ultra or Team plan, look for the reset in your Claude account and redeem it before 10 December at 2pm Pacific Time.
Who sponsored the video?
Mammouth, an app that puts several AI models in one subscription from $10 a month. We have no relationship with Mammouth.
The bottom line
This was a week of promises more than launches. Grok 4.8 is coming on a rebuilt stack, GPT-6.2 is clearly being prepared, and Qwen 4 Max is in preview. The things you can use today are smaller but more practical: a free Claude reset, faster prototyping in Stitch, multi-agent builds in Antigravity and clearer data on how to make your AI plan last. Keep a test task ready, try the real thing when it lands, and own the decision yourself rather than following the hype.
Related posts
- Gemini Krypton Leak: Claude Fable 5.2, a Hidden Fable 6 and This Week’s AI News
- Qwen 4 Preview: First Look at Qwen 4 Max, 27B and Early Coding Demos
- GPT-6.2 Sol Preview: Early Codex Tests Against Opus 5.5

[…] 📘 Main article: This guide follows on from Google’s new Stitch Rapid Design mode, covered in our main article Grok 4.8, the GPT-6.2 Leak and Qwen 4 Max: What’s Real in This Week’s AI News. […]