HUGE OpenAI DevDay LEAK! “o” AI Agent, Sonnet 5.5 BEATS GPT-6, MiniMax M3.1 OUT & 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: HUGE OpenAI DevDay LEAK! “o” AI Agent, Sonnet 5.5 BEATS GPT-6, MiniMax M3.1 OUT & More! AI NEWS · Channel: WorldofAI · 16:43 · Watch on YouTube

Topics: AI agents, AI content generation, AI infrastructure and hardware, AI models

Summary

The video covers leaks and announcements ahead of OpenAI DevDay, including a rumored always-on assistant called O for long-running tasks and signs that OpenAI may expand ultra-fast API access. It discusses reports about Anthropic’s upcoming Sonnet 5.5 and the newly launched MiniMax M3.1 Flash, with the host highlighting coding and image-generation examples. The video also mentions a temporary free offer for LongChat 2.5 Preview and showcases an Opus 5.5 agent building a large Minecraft-style game project over roughly 25 hours. Several claims about unreleased products and model performance are presented as leaks, speculation, or the host’s own tests.

Key points

OpenAI code references reportedly reveal O, an always-on assistant designed for long-running tasks; its details and relationship to the rumored AION/Aeon system remain uncertain.
O may be able to handle emails, use a cloud environment, execute code, research, and potentially coordinate multiple agents.
A speed selector with standard, fast, and ultra-fast options was spotted in the Responses API playground, suggesting broader access to ultra-fast inference could be coming.
The host reports that Sonnet 5.5 is expected in the coming weeks and says partner testing and early evaluations suggest strong performance, while noting these claims are not all confirmed.
MiniMax M3.1 Flash moved from a reported GitHub leak to an official preview release during the recording, with MiniMax positioning it for everyday development.
The video compares M3.1 Flash and GPT-6 image-generation examples, describing Flash’s results as coherent and impressive for a speed-focused model.
LongChat 2.5 Preview is described as free on OpenCode for two weeks, with a million-token context window, multimodality, and zero data retention.
An Opus 5.5 agent reportedly created a Minecraft-style game project with 775 files, 100,000 lines of game code, 26,000 lines of tests, 1,000 automated tests, 783 images, and 184 sounds.

Chapters

Start	Chapter	Summary
00:00	OpenAI, Anthropic, and MiniMax news	The host previews DevDay leaks about OpenAI’s O assistant and API speed options, alongside reports of Sonnet 5.5 and MiniMax M3.1 Flash.
03:07	Firecrawl sponsor segment	A sponsored segment presents Firecrawl’s developer index, its claimed benchmark results, and its integration options.
04:10	The rumored OpenAI O assistant	The video examines reports that O could be an always-on assistant for long-running tasks, while stressing that some details remain speculative.
06:17	Ultra-fast API and DevDay expectations	The host discusses API speed tiers, possible wider ultra-fast access, and hints about upcoming resets and DevDay announcements.
09:27	Sonnet 5.5 reports and examples	The video covers reported Sonnet 5.5 testing and pricing, then compares a Minecraft-clone generation with Astra.
12:03	MiniMax M3.1 Flash and LongChat 2.5	The host notes M3.1 Flash’s official preview release, shows its generation examples, and describes a two-week LongChat 2.5 Preview offer.
14:39	Opus 5.5 builds a game project	The closing demonstration describes an Opus 5.5 agent working for around 25 hours on a large Minecraft-style game built mostly through code.

Quotes

we’re starting to see some huge leaks and glimpse at what OpenAI could be launching. — 00:00

O is rumored to be OpenAI’s answer to agents like Rockbot, Mana, and Hermes Agent specifically designed around long-running tasks. — 05:12

MiniMax officially launched M3.1 Flash preview. — 12:03

We’re talking about 100K lines of game code, 26K lines of tests, a thousand automated tests, and 775 files. — 15:11

Concepts and entities

OpenAI, O, AION, Astra, Anthropic, Sonnet 5.5, GPT-6, MiniMax M3.1 Flash, Firecrawl, Cerebras, LongChat 2.5 Preview, Opus 5.5

Full transcript (original language)

00:00 It’s never a dull day in the AI space, and with OpenAI’s DevDay coming up on September 29th, which is about 72 hours away, we’re starting to see some huge leaks and glimpse at what OpenAI could be launching. From new models to potentially ground-breaking AI agents. Starting off with what looks like OpenAI accidentally leaking one of its biggest DevDay launches, called O. This is reportedly a competitor to products like Rockbot and Hermes agent. It could be

00:30 powered by a variant of Astra called Aeon, and it is designed specifically for long-running tasks. References to O briefly appeared on the ChatGPT upgrade screen, while internal configs reportedly mentions of O as a display name and negative O as an email suffix. Now, this is suggesting that it could potentially handle emails and other tasks for you. Aeon may currently be an internal implementation of ChatGPT’s custom agents, but OpenAI could be

01:01 turning this into a consumer-facing assistant. But that’s not all. OpenAI is also reportedly preparing to expand its ultra-fast API after DevDay. A new speed selector with fast, standard, and ultra-fast modes have been spotted within the response API playground. Combine that with the recently spotted 500 ChatGPT Pro Max plan, which is promising the fastest responses, and OpenAI seems to be heavily doubling down on speed and infrastructure. Meanwhile,

01:32 Anthropic might throw another jab at OpenAI right before DevDay because Sonnet 5.5 is getting extremely close to releasing. Partners reportedly accessed a newer checkpoint that is even better than the first one. And like I had mentioned, if Sonnet 5.5 launches with discounted pricing like GPT-6 sold, this thing could be revolutionary cuz its initial testing has been showcasing some crazy results and it is something that is even outperforming GPT-6 Soul across

02:04 a surprisingly number of tasks while potentially being significantly cheaper. Over at MiniMax, M3.1 Flash was spotted in an early commit to MiniMax’s GitHub repository before the references were removed. And from the early outputs, this model looks seriously impressive for something focused on speed and efficiency. Here’s an SVG of Super Mario drinking coffee with Peach generated by the GPT-6 Soul and then take a look at what the M3.1 Flash had generated. I’m

02:36 not saying it’s necessarily better, but this is seriously a competent result for what appears to be a fast and efficient model. Now, there’s a lot more to cover, so let’s just simply dive into it all. But guys, it would truly mean the world to me if you guys can go ahead and give this video a like as well as subscribing as it truly helps me make videos for you guys on a day-to-day basis. So, really appreciate it if you guys can do that. You hit an error in your app, you paste it into your agent and it starts guessing because general web search

03:07 hands it blog spam, aggregators, and a 4-year-old Stack Overflow thread. That’s what the Firecrawl developer index fixes. 70 million primary sources with readmes, docs, issues, and mergeable requests refreshed daily. So, paste the actual error in plain English. You get back the issue, the pull request that fixed it, and the doc with the match passage already extracted. No second screen. On DevDay, their open benchmark

03:39 of 1,100 real developer queries, it beat every other index on recall. And on issues and pull requests where breaking changes actually get discussed, it is more than double native web search. One line wires it into Claude code parser, Codex, or any MCP client. No API key to start. 1,000 free credits every month. So, stop guessing farcrawl.dev/developer index. And thanks to Farcrawl for

04:10 sponsoring today’s video. Let’s start off with probably the biggest OpenAI leak right now for Dev Day because OpenAI O has officially started appearing inside ChatGPT’s code. According to the leak references, the lowercase O is actually the product name of OpenAI’s upcoming always-on assistant. It’s reportedly supported with 63 languages and the references suggest that multiple pro tiers could be able to access it. Now, another code name for this is AION and OpenAI hasn’t

04:41 exactly been subtle about this either cuz OpenAI accounts have been posting what appears to be the O logo while Tibo, who is one of the main product leads at OpenAI, has been teasing on X about O in the last couple of days. We’ve also seen O briefly appear on ChatGPT’s upgrade screen while the internal configs reportedly reference O as a display name and negative O as an email suffix. That could suggest that this assistant will eventually be capable of actually handling emails and completing tasks on

05:12 your behalf. But here’s where things get really interesting. O is rumored to be OpenAI’s answer to agents like Rockbot, Mana, and Hermes Agent specifically designed around long-running tasks. Instead of giving it a task and getting a response a few minutes later, you could potentially give O something to work on and have it continue operating for hours, days, or potentially even weeks. The underlying system has reportedly been associated with the code name AION like I had mentioned, potentially using a variant of Astra

05:45 optimized for these long horizon tasks. That leads me to think maybe it might be minor as in Astra minor as the model that is powering it. But the idea is that O could get its own cloud environment where it can set things up, execute code, research, and continuously work until the objective is completed. There are even rumors that it could involve multiple agents communicating and working together on the same task. Now, some of the details are still leaks and speculation particularly how Aon and

06:17 Astra fit into the final product. But Open AI already has much of the infrastructure needed for hosted environments, long-running tasks, and agent workflows. And with Dev Day coming up on September 29th, we might not have to wait very long to find out what O actually is. Another potentially huge Dev Day announcement is Open AI expanding its ultra-fast API to significantly more developers cuz a new speed selector has been spotted for the response API playground, which gives developers three options: standard,

06:49 fast, and ultra-fast. And Open AI has now added ultra-fast directly in its API documentation as an access control service tier. Now, ultra-fast itself isn’t completely new cuz Open AI officially previewed it last month for select customers where GPT 5.6 so could generate up to 750 tokens per second and run up to 14 times faster than standard. It’s powered by Cerebras and access has remained extremely limited so far. But these new references suggest that Open

07:21 AI could be preparing to expand ultra-fast access around Dev Day, potentially making this insane inference speed available to a much larger group of developers. And when you combine this with the recently spotted 500 ChatGPT Pro Max plan promising the fastest responses, there seems to be a much bigger story here. Maybe those people get access to this new ultra-fast mode. I am just speculating, but that could be the thing. And OpenAI has seriously doubled down on infrastructure and

07:52 speed, so we could be entering a point where developers can essentially choose how much they value latency. Standard for normal workloads, fast with latency matters, and ultra-fast when you need frontier models responding almost instantly. And with Dev Day right around the corner, I wouldn’t be surprised if the infrastructure ends up being one of the biggest themes of the event. 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

08:23 free. Now, there’s a couple of other things I wanted to mention, and this is where Tibo had essentially said that can’t wait for Dev Day, and he had emphasized that this is going to be the best Dev Day that we had ever had. So, that is obviously setting a really high standard for this Dev conference. Another thing is is that a lot of people have been asking about the resets, and the timing is actually insane cuz Tibo came out and said that more resets are coming next week, which means that we

08:55 may be getting a new model, maybe the Astra Miner is going to be released to help with that claim, but I’m just speculating over here. But, the thing is is that we have a lot of resets coming, and that is just insane, and I really love what OpenAI has been doing. Next up, we have even more information about Claude Sonnet 5.5, and I genuinely think Anthropic could try to drop this right around Dev Day as another jab at OpenAI. Anthropic has officially confirmed that Sonnet 5.5 is coming in the next few

09:27 weeks, but leaks claim that partners have already gotten into access to another checkpoint that is supposedly even better than the first one. There are also reports that Sonnet 5.5 has begun pre-testing inside Claude code with testers describing it insanely fast, efficient, and having more of that natural human-like communication style people loved from the previous Sonnet models. Those test claims haven’t been officially confirmed yet, so take them with a grain of salt, but I have tested

09:58 out this model myself and I’ve really loved the capability of it. I even showcased a couple of examples in my last video. But, the biggest thing here is potentially the price to performance ratio. The rumored pricing is $2 per million input tokens and $10 per million output tokens and 20 cents per million cache reads. Although importantly, that’s also exactly what the current Sonnet 5 costs, so we don’t know yet if those numbers actually belong to 2 or the 5.5. I remember what just happened

10:28 with Opus and Anthropic says Opus 5.5 became over 30% faster while costing 40% less on typical workloads than Opus 5. So, if Sonnet gets anything close to the performance jump we had saw from Opus 5 to 5.5 while maintaining Sonnet level pricing and speed, Sonnet 5.5 could be an absolute monster. And I will say this model is better than GPT-6 all all rounded. So, that is also another wild claim coming from me myself and this is

11:01 based off some of the tests that I have ran. I know you guys are fed up with the Minecraft clones, but this is where the Sonnet 5.5 high, which is what we see on the top, had generated this incredible Minecraft clone with texture as well as a lot of other elements that have been added to the original or to a clone that mimics the original sandbox clone. Then on the bottom we have Astra, which was able to create this Minecraft clone on high thinking mode and you can see that

11:31 Sonnet 5.5 looks a lot better as it was able to add in obviously the shader pack as well as even a map on the top right. So, it added in a couple of other features that we wouldn’t even see with what the Astra model was able to output. So, that just shows alone how capable this model is in terms of its creativity as well as its output scale. So, initially when I was recording this video, MiniMax M3.1 Flash had only been spotted in early commit of a pull request on MiniMax’s GitHub repo.

12:03 Interestingly, those references were actually removed before the PR was merged. So, at that time, it looked like we had caught an early leak, which we did, but literally while I was putting this video together, MiniMax officially launched M3.1 Flash preview. And now it is live through MiniMax code, and MiniMax is positioning it as its latest text model built specifically for everyday development, focused on being fast, reliable, and capable of everything from quick bug fixes to building full features. This went from

12:35 being a leaked model to an official release while I was recording this video, which is kind of funny. And it also literally sums up how fast the AI space is moving right now. If you like this video and would love to support the channel, you can consider donating to my channel 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. And so far, I am pretty

13:05 impressed by the results. This is a Pacoda that was generated in Voxel art, and you can see the quality here. This is where it obviously added in the atmosphere and a lot of the elements that we see stated our models drop, but the fact that this Flash variant is able to actually output something like this is really impressive. You can see the cherry blossom, the lake or like the pond, and a couple of other elements that were generated with this specific output. So, I really love the output quality with this specific model. And

13:37 here’s another demo and comparison. The prompt was to create an SVG illustration of Super Mario drinking a coffee with Peach, and this is what the GPT-6 Pro had generated. Obviously, it’s a lot better, and the level of detail is definitely insane, but take a look at what MiniMax M3 1 Flash had generated. This is a model that keeps up surprisingly well. Obviously, not the same level, but it generates a pretty detailed and coherent scene, especially for a model that’s heavily focused on speed and efficiency. I’m not saying it

14:08 outright beats GPT-6 Soul from this one test, but the fact that a Flash model is able to produce something like this is really competitive and really impressive. Next is where LongChat 2.5 preview is going to be free on open code for the next 2 weeks. This is where you get access to a million token context window. You get multi-modality with this model and zero data retention. So, you’re getting a steal with this model cuz it’s quite competitive in terms of its performance, and it’s obviously not

14:39 the best in coding or any other domain like setting state-of-the-art records, but it is something that is quite decent in most of the categories and domains. So, you can probably use it as your daily driver for the next 2 weeks. Lastly, I wanted to end it off with a bang, and this is where I wanted to showcase an Opus 5.5 generation. This is something that continues to actually blow my mind. This model is just so unique and incredible. This is where it built out an entire Minecraft world from

15:11 scratch almost entirely through code, and the scale of this project is ridiculous. We’re talking about 100K lines of game code, 26K lines of tests, a thousand automated tests, and 775 files. With the agent working on it for around 25 hours so far. But, what’s even crazier is that it generated 783 images and 184 sounds through code. It even built a complete in-game guide. So, if

15:42 someone wants to jump in, they can learn the mechanics and actually play the game. But, this is something that is actually quite fairly accurate in terms of what it is like with the original game. You can see that with the command line as well as obviously a lot of these other textures that were generated. This is exactly where things like long-running coding agents start getting really interesting. We’re moving beyond just little demos into models automatically just creating a game. This

16:13 is something where it’s able to obviously massively work for a couple of hours, but is able to output a multi-file project where it’s something that would take a developer multiple weeks being done in a couple of days, which is insane, guys. Now, that’s about it, guys, for today’s video. I’ll definitely keep you guys posted on whatever happens within the next following week. I will also be making my full dedicated video on the new MiniMax model, so definitely stay tuned on that. But, with that thought, guys, I hope you enjoyed today’s video

16:43 and got some sort of value. If you guys haven’t, take a look at the world of AI inside of Club. Take a look at the world of AI 5 coding platform and benchmark, as well as its second channel, the newsletter, the Discord. Follow me on Twitter to stay up-to-date with whatever’s happening in the AI space. And lastly, make sure you guys subscribe, turn on notification bell, like this video, and please take a look at our previous videos so that you can stay up-to-date with whatever’s happening in the world of AI. But, with that thought, guys, have an amazing day. Spread positivity, and I’ll see you guys further shortly. Peace out, fellas.