Gemini 3.5 Pro X-High, MiniMax M3, DeepSwe, New Claude Models, MiMO-v2.5 Upgrade, & More! AI NEWS

Added: 2026-05-27

[00:00:00]As June approaches, we're starting to see some absolutely wild AI news with major model drops and backend leaks beginning to surface ahead of the upcoming releases. First off, Google is back again with some new updates reportedly being prepared for the Gemini 3.5 series. During the recent Google IO dev conference, Google mentioned that Gemini 3.5 Pro would be launching within a month. But now, new back-end flags are appearing that references of an extra high or X high thinking variant of the model will be dropped soon as well. That could potentially hint at Google finally addressing some of Gemini's previous reasoning effort limitations that held earlier versions back in long horizon task. Meanwhile, the Miniax team is also slowly approaching release of their new M3 model, which was also teased today on X. The company hinted at a brand new sparse attention architecture, and there are reports suggesting that the model could also be open source, which would be a gamecher, and I personally believe it will be releasing in June. On top of that, there are four new leaked Claude Lab products and feature flags showing up in backend logs under the code name coons, squares, bitboard, and claude spaces. Now, the specs tied to these features are genuinely interesting and could point towards Enthropic expanding Claude into a much larger productivity and agent ecosystem. We're also seeing new benchmark rankings being updated across the board. Pricing war is heating up and surprisingly Mimo 2.5 Pro now reportedly cost roughly the same as Deepseek version 4 Pro which is a massive shift in value proposition adding in fresh anthropic cloud code updates more genic tooling improvements and several additional leaks floating around right now. And yeah, there's a lot to cover, so let's just simply dive into it all. 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. Starting things off with Google, which is reportedly preparing to introduce a new X high thinking level, and this is for their new Gemini 3.5 Pro model. This is similar to the higher reasoning effort modes that we've seen from Open AI as well as enthropic models. and this could potentially launch alongside this new model that they're planning to drop in June. Finally addressing some of the reasoning depth inconsistencies previously known to Gemini models. And we have all seen Gemini models struggle with long horizon or more complex task.

[00:02:30]On top of that, Google also appears to be preparing for a brand new Gemini Live model with potential voice cloning capabilities, which is honestly pretty wild. The leaked model identifier showing up in the backend system is Gemini 3.1 Flash Live VR EAP. Now, the live naming heavily suggests real-time multimodal interaction, while the VR and EAP tags could point towards early access or experimental features tied to advanced voice or immersive interactive systems. If this ends up being true, Google may be getting ready to push Gemini Live much further into real-time AI assistant territory. Next up is Miniax officially teasing M3 today on X.

[00:03:12]One of their biggest details that were mentioned today is the brand new sparse attention architecture which could end up being a massive deal for long context AI models. And we all know Miniaax is one of those models that doesn't actually have a long context. This is something that will be resolved, I believe, with this next release. And based off of my internal connections at Miniax, I do believe that they're going to be releasing a model in June. But essentially, this new miniax sparse attention approach changes everything completely because instead of deeply processing the entire context all at once, the model first performs a lightweight scan across everything that identifies the most relevant sections and then it focuses on heavy reasoning only on those important areas. It's kind of like how humans would use the index or table of contents in a massive textbook before deciding which page is actually worth reading carefully. Now, the results is potentially huge cuz it could be up to 10 times faster for context processing, around 15 times faster for decoding speeds, and dramatically lower compute requirements.

[00:04:18]And honestly, this is becoming one of the most important architectural tricks for enabling ultra-ong context AI systems without requiring absurd amounts of GPU power and infrastructure. Now, here is a pretty interesting tweet, and this is another interesting technical detail listed over here when comparing Miniaax's new sparse attention approach against something like Deep Seek's version 3.2 ESA architecture and version 4 CSA system. Essentially, the main changes being discussed in this finding is that Miniaax's implementation is reportedly based on GQA instead of MLA.

[00:04:54]It also uses block level selection similar to CSA of DeepSeek version 4.

[00:04:59]But the major difference is that the tension is performed directly on the real KV cache rather than operating in compressed dimensions. That's actually a pretty important distinction between these different models because it allows Minia Max to retain stronger contextual fidelity as well as reasoning quality while still achieving the massive efficiency gains that a sparse attention architecture is aiming to achieve. Next up, it looks like Enthropic is preparing to launch four brand new Claude Lab products or experimental features that recently appeared in backend logs and internal feature flags. The leaked code names are tunes, squares, bitboard, plot spaces. There isn't official confirmation on what each of these feature flags does, but based on the naming and early specs floating around it, it genuinely looks like Enthropic is expanding Claude far beyond just a chatbot or coding assistant. Some of these appear to be tied to collaborative workspaces, persistent agent environments, organization systems, and potentially even customizable AI workflows or shared project spaces.

[00:06:04]Cloud spaces especially sounds interesting because it could hint towards enthropic building a more persistent operating environment for cloud agents rather than an isolated chat session. And considering how aggressively the industry is moving towards longunning agent memory systems, collaborative AI tooling, these leaks honestly line up perfectly where the market is heading right now. Xiaomi is back with a new upgrade to Mimo and this is where they just announced a massive pricing overall for the Mimo version 2.5 series with API cost reportedly reduced by up to 99% alongside 5 to eight times more usable tokens on existing plans which is just unbelievable. This is with the unified context pricing and simpler billing overall system. Now, one of the biggest takeaways here is that the Mimo 2.5 Pro now is reportedly costing roughly the same as Deepseek version 4 Pro, which is a huge shift in value proposition, as well as AI companies beginning to compete more aggressively on efficiency. They also confirmed that the MIMO version 2.5 TTS will remain free for a limited time with the company stating these improvements came from major inference optimization and serving efficiency upgrades across the MIMO stack. And if you haven't seen my video testing out this model, it's actually pretty underrated in what it can actually do with such a great pricing structure. Also, a brand new Agentic coding benchmark called Deep Sway just released today and it's aiming to test models on far more realistic software engineering tasks. Compared to older benchmarks like Swaybench, instead of scraping existing GitHub issues and PRs, Deep Sway builds tasks entirely from scratch to avoid memorization problems and contamination, which is a big problem when you have new prompts being tested out with different models. It is focusing more on longer horizon engineering work that better reflects actual developer workflows. But interestingly, you can see that the Open AI GPT 5.5 reportedly already scored around 70% on this benchmark, which is honestly pretty insane considering how much harder these tasks are supposed to be. It's another sign that Frontier coding models are rapidly improving at handling more realistic endto-end software engineering problems. And this is why I believe the Open AI GPT 5.5 is the best model in certain cases for long horizon workflows. Another benchmark update is where the Quen 3.7 Max has debuted at number four on code arena which is insane. This is where it is excelling at frontend task. It is quite exceptional at back-end coding logic and becoming the highest ranked Chinese lab on the leaderboard. The model is reportedly surpassing GLM 5.1, DeepSeek version 4, and in certain cases it even out competes the Opus 4.6 six in aentic web development task, which is honestly pretty impressive. Claude Code shipped a new security guidance plugin today for Cloud Code that can help identify and fix vulnerabilities while you're actively writing code. The plug-in is now available for all Claude Code users directly through the plug-in marketplace using /plugins. And essentially, it is definitely going to be helping you with real-time debugging, auditing, and security analysis throughout all your workflows. This next tool or skill you can say isn't actually affiliated with Enthropic. I just thought it would be helpful for most of us developers. And this is a new open-source agent skill that's called React Doctor. And essentially, it is something that's designed to analyze and fix bad React code patterns automatically. It focuses on things like unnecessary rerenders or state management and messy architecture, showing how you can essentially work on debugging a lot of these different components while you're working on other workflows. I thought this could be helpful for a lot of you guys. So, I wanted to mention it in today's video.

[00:09:57]And as I always end these videos off with something interesting happening in tech or AI today, Figure AI essentially announced a major commercial agreement with Catalyst Brands to deploy humanoid robots at scale across all of their operations. Catalyst owns brands like J.

[00:10:15]C. Penney, Aerapostel, and Brooks Brothers with the first deployment beginning in Renault, Nevada.

[00:10:21]Essentially, this is where they're going to be deploying all of their figure one robots within these different operations. And it honestly feels like one of those moments where humanoid robotics is slowly shifting from cool demos on Twitter into actual commercial deployment, which is kind of crazy to think. But regardless, the bigger takeaway here is that companies are now seriously testing whether humanoid robots can economically replace repetitive logistics and warehouse labor at scale. And if these deployments actually work reliably, this could end up becoming one of the biggest new industries created by the AI boom over the next decade. And it is pretty scary if you think about it. 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. But that's about it, guys, for today's video. I hope you found this video to be interesting. I'll leave all the links that I used in today's video in the description below.

[00:11:26]But make sure you take a look at the second channel where we post AI news.

[00:11:29]Join the newsletter, join the Discord, follow me on Twitter, 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 the latest AI news. But with that thought, guys, have an amazing day, spread positivity, and I'll see you guys fairly shortly. He's suffers.