Anthropic RSI Breakthrough: Fable 5.5 Release Date + Next GPT Model Soon

0:00 Honestly, I had to read this one twice. Sam Altman just stood on a stage in Abu Dhabi and said, the model behind OpenAI's Navier-Stokes result is the next GPT, and it's coming soon. And that's not all from OpenAI. Researchers at MIT just showed GPT-4 Astra improving another AI model by modifying its weights without traditional training. Meanwhile, the biggest update perhaps, Anthropic has quietly changed something crucial in its IPO filing, revealing that its models might have already reached the threshold for dramatically accelerating AI progress, potentially approaching recursive self-improvement or RSI as we know it. Not just this, the new Fable 5.5 checkpoint is on a roll, producing some absolutely stunning outputs. And the good news is that we have a potential release date now. We also have some unfortunate rumors about Quen stopping open source models, which could have a deep impact on how we all use AI. So let's start with the biggest story so far, Anthropic's RSI claim.

0:58 As we covered recently, Anthropic filed its papers to go public on October the 26th, and on November the 2nd, it sent an updated version of that filing to the SEC. Most headlines were about the money, and fair enough because Anthropic says it brought in $1.7 billion of revenue in just the three months to September. But a few people went through the two versions line by line, and they spotted a change in the section where Anthropic lists the risks to its business.

1:25 I went and checked both filings myself, and the change is real. In the October version, Anthropic said the capabilities of its models could advance at a rate that would trigger extra safety measures under its responsible scaling policy. That's the rule book Anthropic uses to decide when a model needs tighter controls. In the November version, the same sentence now says its models may have already advanced and may continue to advance at that rate. So the wording has moved from something that might happen one day to something that might have happened already.

1:57 And in a legal document like this, that's not a small edit. There's a second change right next to it. In October, Anthropic said it might not notice a big jump in its own speed of progress until after that jump had happened. In November, it says that it has in the past not noticed such a jump until afterwards. And then there's a completely new sentence. Anthropic says its internal estimates and its expectations for its future models show that it may imminently meet its own definition of dramatically accelerated progress attributable to AI.

2:30 In simple words, that means AI speeding up the research that builds the next AI. And that's exactly what people mean when they talk about recursive self-improvement or RSI. The filing is also very open about how much Claude already does inside the company. It says Claude now helps run experiments, create training data, run safety checks, and write and review the code for the next version of Claude. Anthropic even has a name for this. It calls it closing the loop.

2:56 Now, there are two things worth keeping in mind. First, the filing is careful with its words. It says these estimates are about future models rather than our current models. And it says Anthropic hasn't marked any of its current models as crossing that threshold yet. Second, this is the risk section of an IPO filing. And lawyers write these sections to cover every possible case, good or bad. Still, I don't think anyone changes could to may have already by accident.

3:22 To me, that suggests Anthropic thinks it's closer to this line than it was willing to say a week ago. So if Anthropic's next model is the one where things really start to speed up, the big question is what that model is. And right now, almost everyone is pointing to Fable 5.5. Fable 5.5 still hasn't been released, but a newer checkpoint seems to be reaching some users inside Claude. And what people are building with it keeps getting better every day.

3:50 Just look at this. This is a voxel village sitting between huge mountains. There's a waterfall coming down the cliff, a lake with little boats on it, a red and white lighthouse, and farm fields with crops planted in neat rows. And it isn't a video. According to the tester, it's a living 3D world with its own characters that you can actually walk into an explore. Now here's a fun comparison. The same prompt was given to Fable 5.5 and to Google's Gemini 4 Argon. The idea was an old robot covered in moss waking up in a forest.

4:22 Argon's robot is clean and cartoony. It stands up with its chest closed and it floats up into the clouds. Honestly, that's a nice result on its own. But Fable's robot looks like it has been sitting there for 100 years with trees growing out of its shoulders and flowers on its arms. And when it wakes up, rocket thrusters fire and it blasts off into the sky. I think you can tell which one I prefer. Then there's this one which is a bit different. One tester asked Fable 5.5 to make an animation that makes me feel like it's AGI.

4:53 In about an hour, it came back with a film more than 4 minutes long made with pure code and no outside assets. It opens in space with the sun rising over the earth. Then the words on screen tell you that while you were asleep, the forests came back and the rivers ran clean again. Even the reefs came back. And it ends with thousands of lanterns floating up into the night sky with the line, they do this every time someone wakes up.

5:17 I watched the whole thing and it's oddly moving for something that was written in code. And this last one is only 9 seconds long, but I really love it. A little brass clockwork frog sits in a workshop. It's just lights up blue and it looks around and then a giant hammer comes down and flattens it. The tester says Fable modeled the frog, painted the textures, wrote its own renderer, and even made all the sound. And then they added it can't even hear, which is a fair point.

5:43 There was also a bigger test where one person gave the same prompt to four models, asking each one for a full 3D open world RPG. Fable 5.5 came out on top with a score of 9.4 out of 10 for its golden hour light and smooth combat. But it was close because the new GPT 6.2 Astra was right behind it with 9.1. Now the usual catch applies. Nobody can choose fable 5.5 yet. These results come from people who seem to be routed to a newer model inside Claude, and there's no model name on the screen to prove it.

6:15 So when could it actually arrive? Some testers were betting on today, a Q stay, because of that hidden Mythos 5.5 line we found in Anthropic's website code a few days ago. Others are pointing to next Tuesday, November the 10th. Anthropic itself hasn't said anything yet. And while Anthropic keeps quiet, OpenAI CEO has just said a lot more than usual about its next model. On the 2nd of November, Sam Altman spoke at the opening ceremony of Adipex, a huge energy conference in Abu Dhabi.

6:47 While talking about OpenAI's big math result, he said it was 10,000 agents of the model that is going to be released soon, the next GPT, working for about four days nonstop, collaborating with each other perfectly. So in one sentence, he confirmed that the next GPT is the model behind that result and that it's coming soon. Let's quickly go back to what that result was. On September the 8th, OpenAI shared a proposed solution to part of the Navier-Stokes problem.

7:14 That's the math that describes how water and air flow, and it's one of the seven Millennium Prize problems which come with a million dollar prize each. OpenAI started the agent group on September the 1st. It grew to 10,000 agents running at the same time, and it reached its answer 88 hours later, which is about three and a half days. So when Sam says about four days, he's rounding up a little. What stuck with me most is how those agents worked. There was no hierarchy and no single agent in charge. They split the problem into pieces and kept talking to each other until it came together.

7:45 Now to keep this honest, mathematicians have been arguing about the result ever since. The claim covers two of the four statements the prize actually asks for. The Clay Institute, which awards the prize, hasn't recognized it, and some mathematicians say the agents relied on unpublished work by two researchers. Even the betting markets are split. On Polymarket right now, traders give about a 40% chance that the Clay institute declares the problem solved and only about 7% that OpenAI takes its claim back.

8:13 So the maths is still being argued over, but the model behind it is real. Which brings us to the obvious question. Which model is it? In the last video we saw GPT 6.2 Astra showing up in people's Codex usage pages. And that RPG test we just looked at shows it's already very close to Fable 5.5. Is 6.2 Astra the model Sam was talking about? Honestly, I'm not sure because he didn't give it a name. But a model that's already being tested on real users plus the CEO saying release soon is a pretty strong hint.

8:45 But before we move to what OpenAI's current model can already do, I want to know which AI is your absolute favorite these days? And do you think the next GPT can catch up with Fable 5.5? Do share in the comments. And one thing more, do hit that like button and smash that subscribe button as well. This one comes from MIT and it was published on the same day as Sam's talk. Four researchers from MIT's Computer Science Lab gave GPT 6 Astra the raw weights of other neural networks.

9:17 The weights are just numbers, millions of them, and together they make up a model's brain. As a first test, they showed Astra the 17,253 weights of a small network and asked which game it had been trained to play. Without any tools, it answered breakout and said it was 99% confident. But the experiment that really got me was the last one. They gave Astra a small open model from Alibaba, Qwen 3.5, the 0.8 billion version, and asked it to make the model better at maths.

9:47 The catch was that it wasn't allowed to train it at all. So Astra looked inside the model and found the problem. This little model gets stuck in loops. It keeps repeating the same text and never reaches an answer. In fact, 85 of its first 96 practice answers ran all the way to the token limit. Astra traced that looping to one direction inside two of the model's layers, layers 15 and 19, and then it simply removed it.

10:12 In total, it changed 131,072 weights, which is just 0.016% of the whole model. And the results were big. On Math 500, a standard math test, the score went from 7.9% to 48.1%. On GSM 8K, which is a set of simple word problems, it went from 13.8 to 67.1. And because the looping wasn't only a math problem, the model also got better at following instructions, going from 21% to 62.3. That's three times better.

10:45 The researchers call this weight space programming, which basically means one AI reading another AI's weights and editing them as if they were code. Now, it isn't magic. The researchers also tried a different small model, Google's Gemma 3, the 270 million version, which doesn't have the looping problem. And Astra couldn't make that one any better at math. So what this really does is fix a broken behavior. It doesn't make a model smarter from scratch. Still, I'll be honest, a model fixing another model by reading its brain wasn't on my bingo card for this year.

11:16 Now let's see what Google has been up to because its latest demo is a lot of fun. Google's own team just made a short video using Gemini 4 Argon, and then they played it on an old CRT TV, the chunky kind of tube TV most of us had in the '90s. Here's how they did it. Inside Anti-Gravity, Google's coding tool, they asked Argon to build the video. Argon wrote the blender code for the 3D boot animations and the super collider code for the retro sound. Then they sent each render through a Raspberry Pi into the TV and filmed the screen with a webcam.

11:52 Look at the glow on that Gemini logo. That's real phosphor glowing on a real screen. And at the end, Argon also edited the whole video together. There's also a leak from Google. The company seems to be working on something called smart AI routing, and it's reportedly built on Jeff. Yes, that's the same Jeff we've talked about before, the model that doesn't write anything and just makes quick decisions. The idea is simple. Jeff reads your request and decides whether it's easy or hard.

12:18 Easy requests go to a faster, cheaper model and hard ones go to the heavy model. The claim is that you'd get the same accuracy with much less waiting and a much lower cost. This is still only a leak, so treat it as a maybe. But if it's true, it's a very Jeff thing to do. Now let's head over to China where there's some worrying news and a couple of things you can try today. Let's start with the worrying one, the rumor about Qwen.

12:44 Alibaba's Qwen models have been some of the most popular open models in the world, especially the small ones that people run on their own computers. But discussions in Chinese AI communities this week suggest that the upcoming Qwen 4, the 27 billion version, could be Alibaba's last small open base model. The rumor says Alibaba may be shifting its focus towards much larger models and possibly keeping more of them closed. If that turns out to be true, it would hurt a lot of developers who build on small open models. But for now, it's only a rumor and Alibaba hasn't said anything about it.

13:17 Next, DeepSeek. Their harness, the agent app they launched recently, now has a creator mode. With creator mode, instead of waiting for new features, you can describe the tool you want and the harness builds it for you as a plugin that you can test and combine into new modes. It's still a developer preview, so expect a few things to break, but it's free to try. There's also a new mystery model called Pulse Alpha. It's free right now inside Open Code and testers think it comes from Thinking Machines. The lab started by former OpenAI CTO Mira Murati, although nobody has confirmed that.

13:49 One tester gave it the same long prompt they'd already given Opus 5.5 and GPT-6 Astra. Opus finished in 24 minutes and Astra took 32. Pulse Alpha was still working after almost an hour on its highest reasoning setting, but the tester says the quality so far is much better than they expected. And here's a quick heads-up if you use coding agents. A security firm called Forever Security found a nasty trick. A website can drop a file called agents.md into your downloads folder without you clicking anything. If you then start a coding agent from that folder, it reads the file and follows its instructions.

14:23 In one real case, this pushed an agent to install a known malicious package. And the researchers showed the same idea working with Claude Code too, using a claude.md file. And about half of that firm's customers run coding agents straight from their downloads folder. So the fix is simple. Always work from a proper project folder instead. And finally, this one is just fun to watch. At the Hangzhou Marathon, Deep Robotics had humanoid traffic police robots out on the road, wearing yellow vests and little police caps, walking the crossing as runners and cyclists went past. They're there to help out, not to replace everyone, but it does look like something straight out of a sci-fi movie.

15:05 So with all of that in mind, here's what I actually think about this week. Here's what I've landed this week. On Anthropic, I don't think the filing says they've reached recursive self-improvement. It says they might be close with their future models and that they've missed this kind of jump before. That's careful wording, but it's also the most direct thing any lab has put in a legal filing about AI speeding up AI. On Fable 5.5, the outputs are lovely, but they're picked by the people posting them. And some of them, like the AGI film, took an hour or more to make.

15:40 And on that RPDG test, GPT-6.2 Astra was only three tenths behind. So I don't think anyone is running away with this yet. On OpenAI, I believe Sam that a new GPT is close, but the NaNo Stokes story is messier than the headline because it only covers two of the four parts and it isn't recognized yet. There's also a real argument over credit. I'd rather judge the model when we can actually use it. And the MIT work is the one I'd bookmark. It's early, but the idea of one AI reading and repairing another is going to come up again and again. So which AI is your absolute favorite right now? And which of these updates are you most excited about?

16:20 Do share your choice in the comments. Subscribe to my channel as there are more interesting updates lined up ahead. I'll see you in the next one.