Google's Project Suncatcher IS SENDING AI TO SPACE!?, Sonnet 5.5, $500 ChatGPT Plan & More! AI NEWS

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Google’s Project Suncatcher IS SENDING AI TO SPACE!?, Sonnet 5.5, $500 ChatGPT Plan & More! AI NEWS · Channel: WorldofAI · 17:37 · Watch on YouTube

Topics: AI agents, AI infrastructure and hardware, AI models, AI safety

Summary

The video covers several reported AI developments, including previews and rumored pricing for Anthropic’s Claude Sonnet 5.5, a possible $500 ChatGPT Pro Max subscription, and an OpenAI internal model said to have solved more than 100 longstanding math problems. Its main story is Google’s Project Suncatcher, which aims to test TPUs in orbit and explore space-based AI computing. The host also discusses Anthropic’s new safeguards for some requests and shares tips for using Claude Opus 5.5 and continuing Claude Code after rate limits.

Key points

Claude Sonnet 5.5 is reportedly being tested, with early previews showing generated SVG controllers and a Minecraft-style game.
The video cites rumored Sonnet 5.5 pricing of $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached tokens, plus a reported one-million-token context window.
OpenAI may be developing a $500-per-month ChatGPT Pro Max plan, reportedly offering faster Codex work and potentially higher usage limits.
OpenAI says an internal model has resolved more than 100 longstanding math problems; an independent advisory group will help assess and communicate the results.
Google’s Project Suncatcher will test TPUs in low Earth orbit, with early orbital testing planned before two larger satellite experiments in 2027.
Google is investigating launch forces, radiation, heat management, solar power, and laser links as challenges for space-based AI infrastructure.
Anthropic says it will charge for some requests blocked by safeguards in biology, distillation attacks, and frontier model development; it says few accounts encountered the blocks in testing.
The host recommends using low reasoning effort for routine Opus 5.5 tasks and increasing it for harder problems.
The Claude Code low-priority command may allow use after normal rate limits, but requests run more slowly and only when spare capacity is available.

Chapters

Start	Chapter	Summary
00:00	AI news roundup	The host previews stories about Claude Sonnet 5.5, a possible ChatGPT Pro Max plan, OpenAI’s math progress, and Google’s Project Suncatcher.
03:09	Sponsor: Monad	A sponsored demonstration shows Claude using Monad to retrieve and compare live pricing information through a single tool account.
04:10	Claude Sonnet 5.5 previews and reported pricing	The host discusses reported model specifications and pricing, then showcases early examples of generated art, SVG controllers, and a Minecraft-style game.
08:51	ChatGPT Pro Max and OpenAI’s math model	The video examines reports of a $500 ChatGPT tier and OpenAI’s claim that an internal model has resolved more than 100 longstanding math problems.
11:56	Google Project Suncatcher	The host explains Google’s plans to test TPUs in orbit and the engineering challenges of powering, cooling, and connecting satellites.
15:02	Anthropic safeguards and Claude usage tips	The video covers charges for certain safeguard-blocked requests and tips for choosing reasoning effort and using Claude Code after rate limits.
17:06	Closing	The host wraps up with reminders to follow the channel’s other AI news resources and subscribe.

Quotes

OpenAI says this new internal model has begun training on August 28th — 01:35

Google has already tested its chips against extreme vibrations, radiation, and vacuum conditions — 02:08

Sonnet 5.5 could be an incredible competitive model — 05:12

Google is preparing to launch a prototype satellite that is carrying its TPUs into low Earth orbit — 12:26

you can try running low priority — 16:33

Concepts and entities

Claude Sonnet 5.5, GPT-6 Soul, ChatGPT Pro Max, Codex, Cerebras, Bell, Project Sun Catcher, TPU, Monad, Claude Code, Timothy Gowers, Martin Hairer, Ravi Vakil, Edward Witten, Anthropic, OpenAI, Google

Full transcript (original language)

00:00 As always, it’s never a dull day in the AI space because we’ve got some new huge stories to talk about. Starting off with Anthropic, the new Claude Sonnet 5.5, which could be coming very soon. And this is something that is based off of yesterday’s leaks, where we had talked about how Anthropic seemingly is still testing this model. But today, we actually have our first few previews of what it can do. And honestly, it looks like Anthropic could be reviving the Sonnet series in a massive way. If this

00:31 launches with the pricing and efficiency being suggested, which I’ll reveal later, Sonnet 5.5 could be a serious competitor to GPT-6 Soul, especially based on these early previews. And in my opinion so far, it looks a lot better than GPT-6 Soul. This leads us to OpenAI because they’re reportedly preparing a new higher-tier ChatGPT subscription called ChatGPT Pro Max, potentially priced at a crazy 500 per month pricing plan. So far, references to fastest work

01:03 with Codex are some of the biggest differences spotted compared to regular Pro. With speculation that Cerebrus’s infrastructure could power it alongside potentially higher usage limits. And obviously, we will be breaking down what you could actually get with this new plan. And speaking of OpenAI, remember Bell? Well, there’s another major development around OpenAI’s internal AI capabilities. OpenAI says this new internal model has begun training on August 28th, and it has now resolved

01:35 more than 100 long-standing open math problems, alongside its claimed Navier-Stokes breakthrough. Its progress apparently surprised mathematicians inside OpenAI so much that the company is now working with an independent mathematics advisory group to help review and communicate these AI-generated discoveries. Moving over to Google, they have just announced something insane. Project Suncatcher. It’s moonshot to explore putting AI compute in space. This is where Google

02:08 is preparing to send TPUs into low Earth orbit to see whether a space could actually host massive AI infrastructure. Satellites could take advantage of near constant sunlight while clusters of TPUs communicate through high-bandwidth laser links. Google has already tested its chips against extreme vibrations, radiation, and vacuum conditions with its first orbital TPU test coming before two larger satellite experiments in

02:39 2027. And lastly, there’s a new way to keep using Claude Opus 5.5 after hitting your rate limits. They’ll be slower and only works when extra capacity is available, but it could make Opus 5.5 way more usable. I’ll show you how to actually get this started as well as a couple of tips and tricks, which is going to help you use Opus 5.5 better. Now, there’s obviously a lot more to cover, so let’s just simply dive into it all. Your AI agent can write an entire

03:09 app, but ask it to pull live data, and suddenly you’re juggling API keys, accounts, and subscriptions. That’s where today’s sponsor, Monad, comes in. Think open router, but for agent tools. You get 2,000 plus tools through one key. Web scraping, enrichment, SEO, and social data. Let’s give Claude a real task. Use Monad to fetch Cursor’s current pricing page and compare its individual plans. Claude

03:40 selects the tool and makes the call through Monad. The tool returns the page data directly into the conversation. Claude then turns that real data into clear comparison with a link back to the pricing page. You pay per tool call from one balance without separate provider accounts or subscriptions. Every sign-up automatically gets $1 in free credit with no card needed. So, check out Mona through the links in the description below or the pinned comment and spend

04:10 less time wiring up APIs. Thanks to them for also sponsoring today’s video. Let’s start off with what we had talked about yesterday because Anthropic is currently still testing Claude Sonnet 5.5 spotted under the model select Claude-Sonnet-5-5. And so far we have learned quite a bit about this model. Apparently, select partners were given access for just 24 hours and forced to use has now been retired. Sonnet 5.5 is also reportedly

04:41 coming with a massive 1 million token context window with up to 128k tokens of output. Thinking is adaptive by default while disabled thinking is apparently only available for low, medium, and high effort, not XI or max. But, what really caught my attention, and this is something that we had talked about yesterday, is the rumored pricing, which is $2 for 1 million input tokens, $10 for 1 million output tokens, and just 20 cents for 1 million tokens per cache

05:12 rates. If those numbers definitely hold up, Sonnet 5.5 could be an incredible competitive model that is going to be on par with GPT-6 Soul and maybe even better based off of some of these new leaks. Now, obviously, we already know that Sonnet 5.5 is coming. We already mentioned it yesterday in our video, but today we finally got our first few previews of the Sonnet 5.5. And so far, these previews are insane in terms of its capabilities. Like, in different

05:43 boxel art generations, you can see that the Sonnet 5.5 is putting up a serious fight up against the GPT-3 so which just launched and in my opinion the creative touch as well as the elements that the Sonic 5.5 is able to add is just insane and with that pricing I truly believe it is going to be a huge upgrade over GPT-3 so that’s just my two cents and my prediction. I’m not saying that’s going to happen. But based off of what I have seen so far I am leaning towards using

06:14 the Sonic 5.5 over GPT-3 so in 3GS it was requested to create a castle and this is one of the outputs that it had generated. Not something that is crazy or anything but still a pretty decent generation that adds in a decent amount of texture as well as elements. Now these two controllers were both generated by the Claude Sonic 5.5 and this was done on X high reasoning and you can see the quality over here. And these are two controllers that were

06:45 coded out in SVG and the level of detail is kind of insane cuz you can see the textures of actually the buttons being fully coded out thoroughly in SVG code with proper lighting and shaders being attributed to this SVG generation which is kind of unique on its own. You can see that it was able to thoroughly generate the Xbox controller in an accurate fashion but when it came to the PS5 controller I’m not saying it’s not accurate but it is a little more to not look exactly like a PS5 controller.

07:18 Regardless it did add add in all of the elements of a PS5 controller and still it did do a pretty decent job overall. If you want the best AI tools, workflows and drops before everyone else join my free newsletter with the link in the description below which is completely free. Now here the Sonic 5.5 had generated this Minecraft clone and I got to say the textures and everything that has been generated to represent similar attributes to the original Minecraft sandbox game like the flowers. You even

07:49 have the ability to attack the different mobs. It did do its best job to create these mobs with decent texture and it looks pretty good overall. When you actually kill the sheep, you get wool as well, which is kind of nice. Now, it didn’t add in anything extraordinary like the animations when you are breaking blocks, but regardless, when you break different sorts of leaves, you can see that I got a stick, which is pretty unique. You have a crafting system and I believe there are cave

08:19 systems, which I would need to find. And I just found that right over here. You have ores and when you fall, you also get the damage animation as well. But, let’s actually go up cuz these caves don’t look that nice. Actually, do this. Let’s see how deep we can go. Overall, it does look pretty good. I think that is gravel. Yep, that is. And it did do a decent job with the generations. Even the physics where you can place the gravel block and it falls down

08:51 instantly like the original game. So, that is actually pretty cool. You even have a hunger bar and then you have the crafting system. Moving forward, OpenAI is appearing to actually launch a brand new higher tier ChatGPT plan that is called ChatGPT Pro Max. And this thing could cost a ridiculous $500 per month USD. Now, we have actually talked about this before on the channel where OpenAI would be exploring a much more expensive subscription, reportedly going as high as $2,000 per month. So, I guess

09:23 somebody finally gave them some sense and brought it down to $500, which is still absolutely insane. But, the plan has now been spotted in development with the fastest work and Codex tag and it is currently being in some of the different catalogs for some users, and the only difference shown compared to the regular Pro Plan is that this is something that is powered by Cerebrus infrastructure, which is actually not 100% confirmed

09:54 yet, but that is one of the things people had saw, potentially giving users much faster performance alongside possibly higher usage limits. letting ChatGPT handle much longer running tasks. So, Pro Max could potentially tie into that direction as well. And with OpenAI DevDay coming up, we may not have to actually wait very long to find out what this plan is going to be getting you. Moving forward, remember how we talked about OpenAI’s mysterious

10:24 internal model Bell? Well, we’ve got some pretty crazy new information around OpenAI’s internal AI progress in mathematics. OpenAI says an internal model has begun testing and training on August 28th, which is probably one of the strains of the Bell checkpoint, and it is now reportedly resolving more than 100 plus long-standing OpenAI mathematical problems across multiple areas in mathematics itself. In addition to its claimed solution to

10:54 the Navier-Stokes Millennium Prize Problem. And apparently, the rate of progress has surprised even mathematicians within OpenAI. Because of this, OpenAI is now working with a new independent advisory group of mathematicians featuring some extremely respected names including Timothy Gowers, you have Martin Hairer, as well as Ravi Vakil, and Edward Witten. Their job is to essentially help assess how significant these AI-generated

11:24 mathematical results actually are, and how they should be communicated to the public and research community, and how academic standards should be maintained as AI starts contributing more directly to mathematical research. And just to be clear, OpenAI hasn’t publicly said that this internal model is Bell, but considering what we have previously heard about Bell and the timing here, this definitely looks like it is the Bell checkpoint. If you like this video and would love to support the channel, you can consider donating to my channel

11:56 through the Super Thanks option below. Or you can consider joining our private Discord, where you can access multiple subscriptions to different AI tools for free on a monthly basis, plus daily AI news and exclusive content, plus a lot more. Now, moving into the main story of today’s video, because Google is literally sending its AI chips into space. Yes, seriously, cuz Google has revealed a major update to Project Suncatcher, its moonshot to explore whether we could eventually build

12:26 massive AI computing infrastructure in space. Similar to what Elon had thought about, but Google is preparing to launch a prototype satellite that is carrying its TPUs into low Earth orbit, and this is to see how they actually perform in space. And that is a huge reason why Google is interested in this satellite race, where it is something that can launch these satellites in low Earth orbit and can get near constant sunlight, potentially generating up to

12:56 eight times more solar power than comparable systems on Earth. The long-term vision with this is something that is pretty crazy, cuz Google believes that they could eventually have an entire constellations of satellites filled with TPUs, all connected together and processing massive AI workloads in orbit. But obviously, putting an AI data center in space is insanely difficult. First, the hardware has to survive the launch itself. Google says that an individual component can experience

13:27 forces of around 50 to 100 Gs. So, they’ve literally been shaking these satellites to simulate a rocket launch. Then, there is radiation. Google blasted its Trillium TPUs with proton radiation while they were actively running AI workloads. And the initial results showed that this is something that could withstand more radiation than they would be expected to receive during a 5-year mission to space. But, one of the biggest problems is actually cooling.

13:57 There’s no airflow in a vacuum of space. So, you can’t really cool these chips like a normal data center. So, Google is experimenting with heat pipes as well as radiators to move that heat away from the TPUs and prevent them from overheating. And then probably comes the craziest part, which is connecting all these satellites together. Future Project Sun Catcher satellites could carry each dozens of TPUs and communicate with neighboring satellites using extremely high-bandwidth laser

14:29 links. And this is where Google describes the precision required as basically trying to hit a coin-size target from miles away while both objects are moving. And this isn’t some concept for the distant future. Google is preparing to put its first TPUs into orbit. And then in 2027, it plans to launch two satellites to begin testing this high-bandwidth inter-connection technology. So, we’re genuinely entering a world where it is entering some sci-fi

15:02 world where we have these satellites that are going to have AI workloads functioning in space that is going to be assisting us on Earth. Next up is a pretty weird announcement from Anthropic, and personally, I really don’t like this one cuz Anthropic says that it will once again charge users for certain requests that it safeguards block before Claude even responds. This only applies to three categories: biology, distillation attacks, and frontier large language model

15:32 development. Anthropic says that this is partially because in response to coordinated attacks against its systems in recent weeks. Now, to be fair, Anthropic does say that 99.7% of accounts in its recent testing never encountered one of these new blocks, and the classifiers are supposedly tuned to have less than 0.1% false positive rates. Now, speaking of Opus 5.5, I found a really useful tip that I wanted to let you guys know that

16:03 will definitely save you a ton of tokens and money. Some testing found that the Opus 5.5 on low reasoning effort performed basically on par with Max for task completion, while costing around 12 times less. So, for most everyday tasks, you probably don’t need to actually crank Opus 5.5 all the way to the max reasoning levels. You can start with low reasoning by default and only increase it when you’re dealing with genuinely difficult problems, and obviously this is self-explanatory. You set it to low

16:33 for basic tasks and then set it to max for higher tasks, but sometimes you end up just forgetting, so to be reminded is a great way for you to actually take action on some of these tips. And another really cool helpful thing is that if you were to hit your normal Claude code rate limit, you can try running low priority, which is a command, and it is something that can let you continue using Claude even after hitting your limits. Except your requests will run slower and only when Anthropic has spare capacity available.

17:06 So, if you’re constantly running into Claude limits, this is definitely something that you should keep in mind, and I’ll leave a link to this post in the description below. But, that’s about it, guys, for today’s video. There’s a lot more happening in the world of AI, so I highly recommend taking a look at our school community to stay up-to-date. Make sure you go ahead and take a look at the world of AI, 5 coin platform and benchmark. Take a look at the universe of AI, which is our second channel. Join the newsletter, join the Discord, follow me on Twitter. And lastly, make sure you guys subscribe, turn on the notification bell, like this video, and please take a

17:37 look at our previous videos so that you can stay up-to-date with the latest AI news. But with that thought, guys, have an amazing day. Spread positivity, and I’ll see you guys very shortly. Peace out, fellas.