AI News This Week: GPT-6 Cyber, ChatGPT GIFs, Claude Personal Finance

By Beau Johnson··7 min read

Quick answer

The big AI releases this week: OpenAI's GPT-6 Cyber wrote working root exploits for 948 Linux kernel bugs in a single run, Claude got read-only personal finance, and ChatGPT Images 2.6 makes GIFs. Gemini Omni 1.2 Flash and Meta's Muse Spaces also shipped, plus 15 smaller releases and 5 trending repos.

In this video I go through the 25 biggest AI releases from November 2 to November 8, 2026: five headliners, 15 rapid fire updates, and the five GitHub repos that took off. For each one I tell you what it is, who it's for, and whether you can use it yet. It's for builders and non-technical folks who want to keep up without reading every launch post.

What I cover in this video

  • 0:45 ChatGPT Images 2.6, GIFs, and the new Flare and Sunburst API models
  • 2:16 Gemini Omni 1.2 Flash and editing video by talking to it
  • 4:33 Personal finance in Claude, read-only on purpose
  • 6:04 GPT-6 Cyber and the run on the Linux kernel
  • 7:36 Spaces in Meta's Muse app
  • 9:07 Rapid fire round: models, agents, coding, voice, and devices
  • 17:32 The five trending GitHub repos
  • 20:40 The pattern this week: more power and more limits

The five headliners

ChatGPT Images 2.6

ChatGPT Images 2.6 turns a prompt or an image into an animated GIF. You can change the speed and edit single frames. These are short loops, not video, but a looping mascot or sprite is exactly what app and game builders need. It also saves reference images like you, your pets, or your characters, and it's in ChatGPT and Codex today on every tier.

Developers get two API models: Flare (fast) and Sunburst (quality). Both have a complexity setting of low, medium, or high, which controls how much the model plans before it draws. Pricing is $5 per million tokens for text in, $8 for images in, and $30 for images out. The speed story is mixed. Artificial Analysis found Flare dropped from about 55 seconds to 30, but Sunburst got slower, at about 2 minutes 15 seconds per image. My take: use Flare for volume and Sunburst when it has to look right.

Gemini Omni 1.2 Flash

Gemini Omni 1.2 Flash is Google's video model, and it lives in Google AI Studio. Clips now go up to 15 seconds, and you edit by chatting: tell it to swap the background or make it golden hour and it changes that part of the same clip. You draft at 360p or 720p, then upscale to 1080p or 4K. In the API it's about 7 cents a second at 720p and about 21 cents a second at 4K. It's paid tier only for developers, and it's in Flow and Google Vids at no extra cost. Google says it's still improving on fast sports action and complex physics, so I'd test it on product shots and ads first.

Personal finance in Claude

Personal finance in Claude lets you connect the accounts you choose and ask money questions in plain English. You sign in through Plaid, so Claude never sees your password, and access is read-only. Claude can't move, manage, or invest your money. It's in beta in the US on Pro, Max, and Ultra, and it doesn't work in Claude Code. One catch: unlinking removes every account under that bank login.

For builders, this is the pattern to copy. Let the user pick exactly what the agent can see, let a trusted service handle the login, make it read-only, and give the agent no way to take an action.

GPT-6 Cyber and the Linux kernel

OpenAI showed GPT-6 Cyber working like a funnel: scan for candidate bugs, reproduce them in an isolated environment (1,724 reproduced), test if they can be exploited, then draft patches for humans to review. It wrote working exploits for 948 distinct bugs. These are local privilege escalation bugs, meaning someone with a normal account can become root. It's not a remote hack. The run cost just over $1.21 million in API credits. The patches are drafts, though, and Help Net Security, citing OpenAI, says 61 are patched upstream so far. All of these numbers come from OpenAI. Access is limited to select Daybreak Red customers.

Spaces in Muse

Muse Spaces from Meta builds a planner, tracker, or game right inside the app from a description. You pin it to your tabs, and if you connect your apps it keeps itself updated. You can't share a space yet and there's no developer API, so it's for regular people, not for shipping a product. I'd use it for a channel content calendar.

Rapid fire highlights

  • Models: Inkling 2 Small (594B mixture of experts, 17B active), SWE-2 Flash (free in Devin through November 30), Kimi K3.1 Pro open weights (about 1.56 TB), and Ripple-65B Preview under Apache 2.0.
  • Weight-Space Programming: GPT-6 Astra edited about 0.016% of a tiny Qwen model's weights to stop it looping, and MATH-500 went from 7.9 to 48.1.
  • Coding and agents: Cursor live apps with security checks before publish, Claude API message threads (about 35% faster time to first token on huge requests, no token savings), Anthropic's dreaming for agent memory, Notion as a shared brain through the Notion MCP, and Perplexity Teammates.
  • Voice and devices: Siri AI opened to developers through App Intents, Grok Bot can call and cancel a plan, Gemini voice remixing, an invisible watermark on Grok API text, and an OpenAI device prototype designed with Astra.

Quick tip from the Anthropic memory post: Claude Code only loads the first 200 lines or 25 kilobytes of your memory file, so put the important stuff at the top.

Trending GitHub repos

  1. Auto Ontology (Nvidia): maps business terms to your database so your AI agent writes the right SQL and shows every step.
  2. worktrunk: one command creates a git worktree and starts Claude inside it, for running agents in parallel.
  3. Scrapling: a Python scraping library with an MCP server. Version 0.5.1 gives agents a timestamped strip of frames.
  4. colibri: keeps hot experts on GPU, warm ones in RAM, and streams cold ones from SSD so giant models run slowly on normal machines.
  5. Astryx: Meta's design system with over 150 accessible components. React 19 only and still in beta.

Key takeaways

  • Agents got more power (phone calls, root exploits) and more limits (read-only access, watermarks) in the same week.
  • If your agent touches anything sensitive, copy Claude's read-only finance pattern.
  • Use Flare for image volume and Sunburst for quality, and draft Gemini video at low resolution before upscaling.
  • Expect AI-found bug reports and patches in the stuff you ship on, so keep your dependencies updated.
  • Most headline numbers this week come from the companies themselves, so read them that way.

FAQ

Did GPT-6 Cyber really find 948 Linux kernel exploits?

OpenAI says GPT-6 Cyber wrote working exploits for 948 distinct local privilege escalation bugs in one run that cost just over $1.21 million in API credits. The patches are drafts, and Help Net Security, citing OpenAI, says 61 are patched upstream so far.

Can Claude move money from my bank account?

No. Personal finance in Claude connects through Plaid with read-only access, and Anthropic says Claude can't move, manage, or invest your money. It's in beta in the US on Pro, Max, and Ultra.

Can ChatGPT make GIFs?

Yes. ChatGPT Images 2.6 turns a prompt or image into a looping animated GIF where you can change the speed and edit single frames. It's in ChatGPT and Codex on every tier.

Is Gemini Omni 1.2 Flash free to use?

Not in the API. It's paid tier only there, at about 7 cents a second at 720p. Non-developers can use it in Flow and Google Vids at no extra cost.

Chapters

  1. 0:00This week in AI
  2. 1:00ChatGPT Images 2.6 (GIFs + Flare and Sunburst)
  3. 2:42Night Crew
  4. 2:56Gemini Omni 1.2 Flash
  5. 4:39Personal Finance in Claude
  6. 6:06GPT-6 Cyber and the Linux kernel run
  7. 8:01Muse Spaces (Meta)
  8. 9:20Rapid fire: Inkling 2 Small
  9. 9:56Rapid fire: SWE-2 Flash
  10. 10:27Rapid fire: Kimi K3.1 Pro open weights
  11. 11:02Rapid fire: Ripple-65B Preview
  12. 11:37Rapid fire: Weight-Space Programming
  13. 12:10Rapid fire: Cursor live apps
  14. 12:40Rapid fire: Claude API message threads
  15. 13:15Rapid fire: Dreaming (agent memory)
  16. 13:53Rapid fire: Notion as a shared brain
  17. 14:25Rapid fire: Perplexity Teammates
  18. 14:58Rapid fire: Siri AI developer preview
  19. 15:29Rapid fire: Grok Bot phone calls
  20. 16:01Rapid fire: Gemini Voice Remixing
  21. 16:35Rapid fire: Grok API watermark
  22. 17:09Rapid fire: OpenAI's Astra-designed device
  23. 17:41Trending repo No. 5: Astryx
  24. 18:24Trending repo No. 4: colibri
  25. 19:07Trending repo No. 3: Scrapling
  26. 19:39Trending repo No. 2: worktrunk
  27. 20:14Trending repo No. 1: Auto Ontology
  28. 21:01Wrap-up

Transcript

Read the full transcript

0:00OpenAI says it pointed one AI model at the Linux kernel and in a single run, it came back with working route exploits for 948 different bugs, plus draft patches for about $1.2 million in API credits. That same week, Anthropic plugged Claude into people's bank accounts and built it so it can't move a single dollar. So, one lab is using AI to find the ways in before the bad guys do, and the other one locked its AI down to read only on purpose. And that's just two of them. Chat GPT makes gifts now. Grockbot can call a company and cancel a contract for you. And one AI went inside another AI's weights and edited them. So, in this video, I'm breaking down the

0:4525 biggest AI releases of the week, five headliners, 15 more in a rapidfire round, and the five GitHub repos that blew up. For each one, what it is, who it's for, and if you can even use it yet. Let's jump right into it. So, number one is chat GPT images 2.6 from OpenAI. And the headline is GIFs. You give it a prompt or an image, and it turns it into an animated GIF. You can change the speed, edit single frames, and it plays on repeat. So, these are short loops, not video. But if you build apps or games, a looping mascot or sprite is exactly what you need. It also saves reference images now, like you, your pets, or your characters. So, you upload them once, and it's in chat GPT

1:31and codecs today on every tier. So, how does it actually work? For developers, there are two new API models. Flare is the fast one. Sunburst is the quality one. And both have a new setting called complexity. Low, medium, or high. That's how much the model plans before it draws. Turn it down and it finishes faster. Open AAI says Flare can make an image in as little as 6 seconds on low quality. Both can output SVG and animated GIFs, and you can draw markup on your image to guide and edit. It's $5 per million tokens for text in, eight for images in, and 30 for images out. So, here's where it gets messy. Open AAI says up to 50% faster. Artificial

2:16analysis tested at max quality and put 2.6 at number 1 and two, and Flare dropped from about 55 seconds to 30. But Sunburst got slower than the last Sunburst, about 2 minutes 15 per image. And on arena at medium, the new sunburst is number two behind the old one. So which is it right? It depends on the setting. Use flare for volume. Sunburst when it has to look right. Easy flip got 460,000 views in 11 days. Nightcrew made one new short every day and posted it to four platforms. The best one hit 155,000. [music] Want that for your business? Nightcrew.studio. So number two is Gemini Omni 1.2. to flash and it lives right inside Google

3:02AI Studio which is where we build on this channel. It's Google's video model. Clips now go up to 15 seconds up from 10 and you can pick any exact length from 3 to 15. So the big change is you edit in conversation. You make a clip then you say swap the background or make it golden hour and it changes that part of the same clip. You're not rolling the dice on a whole new video every time. So, here's how it actually works. Draft first, upscale later. You generate at 360p or 720p, which is fast and cheap. So, when you like one, you upscale it to 1080p or 4K. So, the 4K is an upscale. It's not generated at 4K. You can also feed it reference images like a

3:47character or a product, motion clips up to 3 seconds, and audio cues. So, the price in the API, it's about 7 cents a second at 720p, about 30% cheaper per second than the last version. And 4K with the upscale is about 21 cents a second. So, can you use it today? If you're a developer, yes, in AI Studio and the Gemini API, but only on the paid tier. There's no free API tier for this one. Not a developer? It's in Flow in Google Vids at no extra cost and rolling out in the Gemini app for Ultra subscribers. Now, here's what this is all about. There's no independent benchmark yet, and Google's own blog says it's still improving on fast sports action and complex physics, so I

4:33wouldn't film a basketball game with it. But product shots and ads, this is the one I'd test first. So, number three is personal finance in Claude. You connect the bank, card, loan, and investment accounts you choose, and you ask Claude money questions in plain English. Where did my money go this month? And it answers from your real balances and transactions. So, how is that safe? You sign into your bank through Plaid, so Claude never gets your password. And you approve readonly access. Claude can see account names and types, the last four digits, balances, credit limits, and up to 2 years of transactions. That's it. Anthropic says Claude can't move, manage, or invest your money, and it doesn't see your investment holdings.

5:18So, can you use it? If you're in the US on Pro, Max, or Ultra, it's rolling out in beta on web, desktop, and mobile. Check the side menu. It might be hiding under more. It's not on free, team, or enterprise, and it doesn't work inside Cloud Code. Now, here's something most people don't think about. This is a blueprint for builders. If your agent touches anything sensitive, copy this pattern. Let the user pick exactly what it can see. Let a trusted service handle the login. Make it read only and give the agent no way to take an action. So that's how you get people to actually trust your agent. So one catch, if you unlink, it removes every account under that bank login. You can't unlink just one. And Claude can make mistakes. So

6:04check the big numbers with your bank. So number four is the one from the start of the video, GPT6 Cyber from OpenAI. The day before they announced it, OpenAI showed what it can do on the Linux kernel. So here's how it actually works. And don't get lost in the sauce here. It's a funnel, and you can see it right here. First, it scans security code, and that gave OpenAI 2,38 candidate bugs. Then it tries to reproduce every single one in an isolated environment and 1,724 reproduced. Then it tests if each one can actually be exploited. And for 948 distinct bugs, it wrote a working exploit, which is insane. Then it drafts a patch for each one and human researchers review everything before it

6:50goes to the kernel maintainers. So these are local privilege escalation bugs. Somebody who already has a normal account on the machine can become root. So, it's not a remote hack of any computer. So, the full run cost just over 1.21 million in API credits. That's about $1,279 for each distinct exploit. So, here's the honest part. Those 948 patches are drafts. Helpnet Security citing OpenAI says 61 are patched upstream so far. So, dozens merged, not hundreds. And every one of these numbers is OpenAI's own. So, can you use it? No. It's for select Daybreak Red customers. And OpenAI says access is expanding over time. Security teams at Mosilla, Cisco, and PaloAlto

7:36networks are already using it. And the same week, Google DeepMind said its models guided by humans found two major weaknesses in old 1024-bit RSA. Modern 2048-bit keys aren't affected. So, no, encryption is not broken. So, what does this mean for you as a builder? Expect AI found bug reports and patches to start showing up in the stuff you ship on. Keep your dependencies updated. So, number five is spaces in Muse from Meta. You describe a planner, a tracker, a game, whatever you need, and Muse builds it right into the app. You pin it to your tabs, you can schedule when it shows up, and if you connect your apps, it keeps itself up to date in the background. Zuckerberg says he made a barbecue game, an event tracker, and a

8:22home control dashboard. So, how does it actually work? It's prompt to app, but the app lives inside the assistant. It's the same idea as the build tab in AI Studio, where you type an idea and get an app back. So, the difference is Muse keeps it as a personal tab for you. You can't share a space yet, and there's no developer API. So, this is for regular people, not for shipping a product. Meta also says it's rolling out new specialized models to Muse and Meta AI users. Meta's own tests say they're 2.4 times less likely to forget critical info and 67% better at self-correcting and finishing everyday tasks. Those are Meta's numbers with no baseline. So, take that for what you will. So, here's how I'd use it. I'd have it build a

9:07content calendar for the channel, pin it, and keep adding columns as I go. Is this going to replace AI Studio for me? No. But for people who will never open a code editor, this is how they get personal software. All right, rapid fire round. Let's go. Thinking machines dropped Inkling too small. It's a mixture of experts. 594 billion parameters, but only 17 billion run for each token. So, a router only calls in the few specialists each word needs. It scored 31 on the artificial analysis intelligence index, and you can use it today on Tinker with closed weights. They've promised the weights under Apache 2.0 after safety testing, but it's wordy. Artificial analysis says about a $118 per task, almost six times

9:54its predecessor. And speaking of Inkling, Cognition built S.2 Flash on top of the first Inkling Small. It's their fast coding model for Devon. Cognition says it's 1/5if the price and three times the speed of S.2, E2 and it's free through November 30th in Devon desktop and the Devon CLI. Now, it's weaker than SWE2 on almost every eval in their own table. So, the pitch is speed and cost and it's not open weights. So, if you're on Devon, try it while it's free. Moonshot released the weights for Kimmy K 3.1 Pro. The model came out in October. What's new is you can download it. So, remember mixture of experts? This is the extreme version. 2.8

10:41trillion parameters, 104 billion active per token. The files on HuggingFace add up to about 1.56 terabytes. So, no, this isn't running on your laptop. And the license looks like MIT, but it isn't. If you sell it as a service and make over $20 million a year, you need a separate deal with Moonshot. Recursive Intelligence, a startup that uses AI to design chips, released Ripple 65B preview. It's a dense model they trained from scratch, so every parameter works on every token. No experts. It's Apache 2.0 on hugging face and it's on together, fireworks, and open router. Recursive says it scores 80.5% on terminal bench 2.1, the best for a US

11:27Open Weights model on that one benchmark. and they say it fits on a MacBook or a DGX Spark. I'd read that as a MacBook with a lot of memory. Researchers at MIT gave OpenAI's GPT6 Astra access to the weights of a tiny Quen model with no training allowed. Astra found the direction inside the model that makes it loop while it's thinking and subtracted it. So, it changed only about 0.016% of the weights. And math 500 went from 7.9 to 48.1, but about a third of the answers still hit the token limit, and it couldn't add abilities the model didn't already have. I did a whole video on this one if you want the full breakdown. Over to agents and coding.

12:12Cursor can now build live apps and share them with a link. You describe the tool and a cloud agent builds it, spins up a Postgress database if it needs one, sets up signin and runs security checks like catching secrets sitting in your app files. So nothing goes live until the checks pass and you hit publish. It's on all paid plans with cloud agents and public links are only on individual plans. I mean, how many vibecoded tools have you built that never left localhost? Anthropic shipped message threads on the clawed API in public beta. Normally, every turn of an agent loop, you resend the whole conversation. Now, the API holds it on the server and you just send the new messages plus the previous message ID. So, your requests get way smaller and Claude starts

12:58answering sooner. Anthropic says on requests of 500,000 to a million tokens, the median time to first token drops about 35%. Now, it does not save tokens. Claude still reads the whole thing every turn. Most Claude code users already have it on by default. [snorts] Anthropic also posted about agent memory and something called dreaming. So agents keep transcripts of their work. Dreaming is when Claude rereads those transcripts, fixes wrong notes, and cleans up its memory. In their Pokémon test, Claude caught a wrong note about where the Poke Center was and rewrote it. In managed agents, it's called Dreams. It's a research preview and you have to request access or build your own from their cookbook. And one tip from

13:43the post, Claude Code only loads the first 200 lines or 25 kilobytes of your memory MD. So put the important stuff at the top. Notion [snorts] wants to be one brain for all your agents. Chat, GPT, Claude, Muse, and Grockbot can now save what they learn to the same Notion memory and use the same skills. It runs through the Notion MCP. You give each agent Notion's setup prompt once, and from then on, it reads a guide page every session and writes what it learns back. It's not automatic, and agents don't always follow the guide. But if you bounce between three assistants and keep explaining your project over and over, this is how you stop. Perplexity launched teammates. So instead of a

14:29prompt, you give it a job. It works on its own using your computer tasks, skills, memory, and tools. It teams up with other teammates, and it can message your co-workers in Slack. It's out now for Mac subscribers on the web with the apps coming soon. 2 days before that, Perplexity pitched computer for e-commerce, working across Shopify, Clavio, and your ad accounts. An agent that keeps an eye on your store all day is the kind of agent I'd actually pay for. Now, voice, devices, and trust. Apple opened Siri AI to every developer in the 27.3 betas. Your app's app intents, the actions your app can do, become things Siri can call when someone asks in plain language, and new skills

15:14teach Siri to chain several actions into one request. Customers get it in a few months, and app store submissions open in the spring. It's not in the EU on iPhone, iPad, or watch. So, if you have an iPhone app, start wiring up your app intents. Now, Grockbot [snorts] can now make phone calls for you. In the demo, it notices a contract about to renew, offers to call and cancel, and you hit call. You can see a live transcript with takeover and end call buttons. The company offers a 10% discount to stay. Grockbot turns it down and the plan gets cancelled. The post didn't say who has it or what it costs, so check your app. And for developers, OpenAI's live API can now make outbound phone calls through your own SIP provider. Who doesn't hate hold music?

16:01Back to AI Studio. Google launched voice remixing for Gemini 3.8 Flash text to speech. You start from one of Google's preset voices, and Google says there are over 2,000 or a voice you designed. So then you prompt it. Add a subtle southern accent. Make it sound like a fireb breathing dragon. It saves as a brand new voice you can reuse in AI Studio and the Gemini API. Now, the docs say preset or designed voices, not cloned ones. So, you can't upload your own voice and remix that. At least not yet. Starting November 7th, text from seven Grock API models carries an invisible watermark by default for the EU AI act. So, how does it actually

16:46work? At each step, the model picks the next word and the watermark slightly shapes which words it picks. No extra characters and it reads totally normal. A detector finds the pattern statistically. A response header tells you it's on. We made a test call on Sunday and yep, the header said true. Team admins can turn it off, but they have to confirm a legal exception applies. And to close out the round, two people at OpenAI used GPT6 Astra to design a prototype device for talking to codecs in under a week as a side project. Astra designed the enclosure, the circuit board, the interface, and the software, and a contract manufacturer built a small batch in under a month. But OpenAI says it still took several rounds of review to get it

17:32made because Astra didn't always account for the manufacturer's limits. So, no, they're not selling it. Hardware people, this one's for you. Okay, GitHub repos, five of them, counting down to number one. Number five is Asterisk, Meta's design system. It grew inside Meta for 8 years and powers over 13,000 of their apps. It has over 150 accessible components, seven themes, dark mode, and a CLI built so you and your coding agent use the same docs and commands. Meta open sourced it in June, but it shipped a new release almost every day this week and hit number six on GitHub's daily trending on Sunday. About 15,000 stars. It's React 19 only and still in beta. So

18:18if every app your agent makes looks like the same AI template, give it a real component kit. [snorts] Number four is Kolivi, the number one C repo on GitHub this week. And remember mixture of experts. This is the payoff. Since only a few experts fire for each token, Kibi keeps the hot experts on your GPU, the warm ones in RAM, and streams the cold ones off your SSD only when the model asks for them. No GPU required, and 8 gigs of RAM at the very least. So, giant models run on normal machines, slowly. The author's own table shows a 744 billion parameter model at well under one token a second cold on a laptop. Small models are quick about 48,800

19:06stars. Number three is Scrappling. Just under a 100,000 stars with about 8,700 of those this week. It's a Python library that scrapes pages, renders JavaScript, and pulls out the data. And your agent can use it through its MCP server. Version 0.5.1 landed November 4th, and now your agent can watch a page move. Instead of one screenshot, it hands the agent a strip of frames with timestamps. It can also drag and drop. So, just respect each site's rules because scraping against a site's terms can get you blocked. Number two is Work Trunk. It's number three on GitHub's weekly trending list across all languages and number one in Rust. It's a command line tool for git work trees made for running AI agents in parallel.

19:53Like I said last week with Think Rail, a work tree gives each agent its own folder so they don't step on each other's changes. But plain git makes you type the branch name three times just to start one. So with work trunk, one command creates the work tree and starts Claude inside it. And one more cleans it up. It's one brew install. About 12,000 stars. And [snorts] number one is Nvidia's autoontology. Number one on GitHub's weekly trending list across all languages with about 9,000 stars this week. So what is it? You know when you ask your AI what revenue was last month and it writes SQL that looks right and gives you the wrong number? This is built for that. It reads your database's metadata and builds a map of business terms like revenue,

20:40product, and state and how they connect. When you ask a question, the agent matches your words to those terms, follows the links to find the joins, and gives you the answer with its SQL and every step. Setup is docker compose plus an Nvidia API key. So, it's not a oneliner. And there's an MCP server for cloud code and cursor. So, that's [snorts] the week. And if you noticed a pattern, agents got a lot more power and a lot more limits at the same time. phone calls and root exploits on one side, readonly access and watermarks on the other. So, here's my question. Would you let an agent make a phone call for you or look at your bank account? Or is that a line you're not crossing yet? Tell me down below. And if this caught you up on the week, give it a thumbs up.

21:26It really helps the video spread. Same time next Monday. Have a blessed day.

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading