← Back to search

Claude Opus 4.8 is here! Will the most powerful automatic development function "ultracode" make engineers shine?

よわよわえふえむ · 2026-05-31 · 36 min
relevance 84 53 words Episode page ↗ Audio ↗
Show full episode description
🎧 今回はGoogle I/O 2026で発表された「Antigravity 2.0」や「Gemini 3.5 Flash」の衝撃、そして最強モデル「Claude Opus 4.8」と新機能「ultracode」がもたらす未来の開発環境について徹底的に語り合います! ✅ 任天堂の新感覚写真アプリ「ピクトニコ!」の魅力 ✅ 爆速かつ賢い!Gemini 3.5 FlashとAI Proプランの仕様変更 ✅ X Premium連携で情報収集が変わる、Hermes Agentの「X Search」検証 ✅ 異次元の並列自律開発!Claudeの「ダイナミックワークフロー」と注意すべき罠 ✅ GPT-5.5 vs Opus 4.8、これからのAI使い分けと「人間が読めるコード」の重要性 AIが爆速でコードを書き換える時代の生存戦略とは?ぜひ最後までお聴きください! ▼ 話題になったリンク ▼ • [Pictonico! | 任天堂]: https://pictonico.nintendo.com/ja/index.html • [I/O 2026: Antigravity 2.0 と新しい Gemini のエージェントツール]: https://blog.google/intl/ja-jp/company-news/technology/google-io-2026-developer-highlights/ • [Gemini 3.5: frontier intelligence with action]: https://blog.google/intl/ja-jp/company-news/technology/gemini-3-5/#gemini-3-5-flash • [X Search | Hermes Agent]: https://hermes-agent.nousresearch.com/docs/user-guide/features/x-search • [Introducing Claude Opus 4.8 \ Anthropic]: https://www.anthropic.com/news/claude-opus-4-8 • [Introducing dynamic workflows | Claude]: https://claude.com/blog/introducing-dynamic-workflows-in-claude-code #ポッドキャスト #GoogleIO2026 #Gemini35 #ClaudeOpus #Antigravity #HermesAgent #AI開発 #ピクトニコ #Techニュース
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Tracking the latest AI model and subscription changes and finding a way to pull closed X information into agent research.
Benefits
  • Unified Claude Pro plan covering Sonnet and Opus
  • Gemini 3.5 flash gives near-instant code answers
  • Cheaper Google AI Ultra tier now available
  • X Search skill works inside Hermes while logged into X
  • Automated news triage instead of doomscrolling trends
Use cases
  • Runs Codex automation 'about three times a day' to pick up only AI news of interest
  • Built an X Search skill usable in both Codex and Claude to retrieve closed X post summaries with source tweet links
  • Switched to Google AI Ultra now starting at $100, matching ChatGPT Pro pricing between the $200 plans
KPIs / results
  • Pictonico mini-games run ~5 seconds each
  • Google AI Ultra now starts at $100 (was a $200-only tier)
  • Claude Opus 4.8 only lost to GBT5.5 in scoring
  • Codex news automation runs ~3 times a day
Tools / build
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
Yes, I'm here. So, I'm not feeling well yet, so I may have a cough or my voice might be a little hoarse, but I apologize for the inconvenience. That's it. I've been getting a lot of AI-related news this time as well, but the first thing I wanted to talk about was, can I talk about something else instead? Yes, it's okay. There's a game I've been looking forward to playing. Yes, do you know about Pictonico? No, I've never heard of it. It's made by Nintendo. Yes, yes. It's an app that, um, uses the photos you've taken and turns it into a game. It's amazing how well it's done, and it's really fun, so, um, there are things that could have been better or worse. Yes. I don't know what it is, uh, it has a Wario-type feel. What was it about, that Wario? Wario? Um, isn't there one that's full of mini-games? Something like Made in Wario? Yes, that's right. Something like that, uh, you're forced to do something like this in a really short amount of time. Yeah. That's exactly what it is. It's like there are some instructions written right away, and you have to do them quickly. Yes, yes. It's like this, and if it disappears, you have to look at it and take immediate action. It's only about 5 seconds at a time. That's especially true for a game with a time limit. Hmm. I think it's really nice because it's done so quickly and everyone can get excited. Hmm, it's just the pre-game. You don't need a Switch or anything. No, you need a camera. Ah, that's right. I see. I said that a game was released. Yes, that was after that. Last time we talked, it was just before Google I.O. Yeah. That's why there was Google I.O., and a lot of things were announced. What we were really curious about was the evolution of Gemmy. Yes, we were wondering what would happen to it. However, there was a big evolution, or rather something unusual. Well, I'm sorry that there is no UI or anything like that on the screen. Yeah. At Google I.O., Anti-Gravity 2.2.0 was released. Yeah. What type of this, uh, is it? You're a subscriber to Open AI services. Well, if you're a subscriber to Chat GPT, the image sharing is an application that looks like Codex Up. Even now, when it's Anti-Gravity 1.0, is it called 1.0? I don't know, but when it's Anti-Gravity, it's based on VS Code, I'm sure. Yeah, yeah. I wonder if it's VS Code, but I had a slightly unpleasant experience interacting with the AI. Well, I think the good thing about it is that it has a proper editor, so you can edit it with the editor. Well, as AI has evolved in various ways, the opportunities for humans to edit in this way are decreasing, or so it seems. Well, I think the UI has probably changed, but yeah, something called Anti-Gravity 2.0 came out. So what do you think? I'm going to shake it loosely. No, just when this came out, at the same time, Jiminy's plan also changed a little. Yeah, yeah. Well, at that time, in my case, YouTube Premium was originally included. It's quite useful. Yeah, yeah. Yeah. I can't use it all the time, um, well, I use Codex on a regular basis, so it doesn't really cut it down, but anyway, I use the 3.5 flash Yes, but it's fast, and when I want to ask a question about a code or something, I can listen to it very quickly. Yes, the speed is extraordinary, it kind of feels like a Cerebras, in my experience. From what I've experienced so far, flash is faster without being, well, smarter. I think it's been said that the pros are smarter, but right now I don't know which one is smarter. Yeah, yeah, yeah, yeah. I've been using it thinking that 3.5 flash is fine, but yeah, yeah. That's why it's really good. In the sense that I often think like this, 3.1 Pro is definitely better. I get the impression that it's thinking about things. Hmm, the 3.1 Pro is still quite good in terms of daily conversation and such. Yeah. I see. Yes. I see. Why, well, I guess it depends on the situation. Yeah. Also, that's it, isn't it? Two things. Well, this one can be used as well. Moreover, it looks like the limit for each model is set for each model. Well, I don't know how much I can use this one. Well, if that's the case, I'm thinking that I'll be fine with Claude Pro's contract. Now I'm feeling like I want to have Sonnet and Opus. Yes, it can be used for those models as well. It's not that you can actually use it. Um, that's right, that's right, that's right, that you can use it with this plan. Wow, that's right. That's right. So, I can't use a lot of Claude Pros with Opus. In that case, I would like to combine them into one, but Claude's co-work function is useful in its own way. After thinking about it for a while, I said yes. Well, I guess I should unify them. Well, Claude's bill ran out this morning or yesterday morning. Oh, I forgot to start the medicine. Hmm. Big failure. Big mistake. Well, it's the cheaper plan. Well, I guess it's Claude Design, too, so, uh, yeah. In comparison. As for the reason for doing that. Also, that's it. Gemini CLI will disappear. That's right. In September, it will become Anti-Gravity CLI. Well, Anti-Gravity CLI has also been created. Well, I haven't completely migrated yet, but it seems that Gemini CLI's functions are basically like making an investment. Well, that's also kind of interesting. Well, this one was also huge, wasn't it? Google AI Ultra subscription now starts at $100. It used to be a crazy price. Do you mean you can now get a $100 plan between the $200 plans? Yes. That's right. So it's now exactly the same as Chat GPT PRO. It's new, isn't it? This idea is new. I think it's new. I don't think people would think about creating another pro under a pro. Yeah. I think this is new. Yeah. So it's Ultra. It was Open Air when they called it Pro. That's right. It's exactly the same thing. That's why all three companies started doing the same thing. Basically, Anthropic was the first to do it. Yes, and then Open Air. That's right. It happened when they changed from message-based to token-based. Shortly after that. Yes, yes. That was announced. So this time, Google made an announcement like this, and yes, it received quite a bit of criticism. Now that this new subscription has started. Well, even for pros. It's token-based. It's like a usage fee. It seems like something like that has come out. So, if you look at the Gemini page, no matter how much you use it this week. Well, you'll see that familiar screen. It seems like he's saying something about improving things. Yes, that's all. We've talked about Gemini. Eh, Gemini's screen isn't shown. Now, let's talk about Gemini 3.5. I thought I'd show it a little more there. I see. Did I start talking about it? Ah, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no. Well, the Gemini 3.5 flash. Well, of course the specs have gone up dramatically. Yeah. It has a lot of better points than the 3.1 pro. Yeah. If it comes out like that, it's a model. Yes. If it doesn't have the sound? Well, it does. Just a little. There's a little bit at the bottom. Well, it's a benchmark that I don't really care about. When it first came out, it got quite a bit of bad press. Ah. Ah, that's right. Yeah. It didn't seem like it had the kind of expectations it had at all. It was like I couldn't be entrusted with the coding. Hmm. Well, I don't know if it got a lot of surprises or surprises. Yeah. It kind of felt like that. It had a bad reputation, though. However, what Mika said earlier, I don't think it's bad as far as I'm concerned. In terms of benchmarks, well, Terminal Bench isn't bad, so it kind of feels like that, though. Yeah. Ah, this is exactly that. Oh, this isn't it. Isn't this the other one? Spark? Isn't this Antigramity? Isn't this something that will come out next time? Gemini Spark. Ah. I think it's something that's only available in America right now. Yes. Well. Spark. Yes. Yes. Like co-work. No, Spark. Spark. It's a new thing. Yes. This is amazing. After all, it's like my own assistant. There are all kinds of things like this. After all, it was said that there would be something like a chat that can be done while using MCP and skills. Co-work. I've heard of it. Co-work. For some reason, I couldn't use it for Geminis for a long time. I couldn't do that kind of thing. Yeah. Well, maybe I feel like this spark will become the basis in the end. Yeah. Well, it's just that sometimes it's annoying to be able to do anything. Especially when it's running on a local application. For example, if something like a virus is brought in, there's a possibility that the local area can be contaminated as well. Yeah. Well, that's a hassle, isn't it? Yeah. Yes. I'm wondering if it would be safer to ask Gemini from the web version. Yeah. That's right. Hey. Okay. And then this is changed. Well, it's not about Gemini all at once. Yeah. Well, that's what happened this time. It was X Premium. Using that account. Last time. We talked about it last time. That Hermes agent. I don't know which one. It was announced in a big way that you can use the function there. Yes. What else was it about that Groke? It seems like you can use something like the Groke code or something like that. Yeah. Yeah. I felt like there was something like that. Was it Groke? I wonder if it was about another contract. I'm sorry for a moment, but I'll leave it at that. Yes. There's something I'd like to talk about a little bit from everyone. Well, I just saw this news. I haven't used the Hermes Agent itself, but I do things like search the web, and I use Codex to do things like that. Well, I do it as part of my research. Inevitably, the information about X was kept in a closed area. There was no way to retrieve that information so clearly. Well, I thought it would be possible to do it using this, so I tried installing it. There was no need to install the Hermes agent separately. It was inside the Hermes agent. Well, this... um... Well, I was able to use just that part. So... Well, I was able to use it while logged in to X. Well, when I used it, I was able to search for it, retrieve the information that I needed, and use the results. So, I actually tried that Codex skill. Well, I guess you could call it Codex. I created it myself as a skill that can be used in both Codex and Claude. Hmm... What is that? This X Search is... Um... It's not something that searches for X posts. Um... Well, that's not it. It's not. You can't search for those posts. Ah, it's a function to search summaries of posts. How troublesome. What? Um... It's not a replacement for that API in X. Well, there's a contract for each API. That's right. Yes, yes. So, I don't think this is okay. You can't retrieve a specific post, or someone's timeline. That's why you can't retrieve your own timeline separately. Hmm. That's a bit inconvenient. Um. Well, I want to store my own information, like AI information. There's also something that's currently being talked about. I wonder if there's something else. Something like detecting an obstacle. When searching for something like that, I'd like to do some research. I've been using this skill recently. It's become a bit useful. I wonder what it is. Erm. Station. I wonder what it is. Grounding. I wonder what it is. Um. After doing a search. Isn't there something like that source?What does that mean?Ah.The source is.The tweet appears.The link to the tweet appears.Ah.The tweet appears properly.Oh yeah, that's right.That also appears.I see, I see.Well then.It doesn't seem so bad.Yes, that's right.I was wondering what it meant when a summary appeared.Search for the summary.Post it. A post like this... Well, it seems like it was the seed of this summary. Is that how it's presented? It's different from the topic, isn't it? Hmm. Something like that. You can check it out on the web. Yes. It's not like it's some kind of thing. I don't think so. Hmm. It's difficult. I don't really understand it either. No, but somehow. X tends to gather information like this. It's especially easy to gather information about Japanese people. Yes, I think it would be interesting if the power balance could be a little more imbalanced. Uh. Well, I guess the current situation is still very low. U. U. I feel like things are starting to get better. U. U. Even recently, some people have started to feel like they can communicate normally in multiple languages. That's right. That's why. I started following people like that. I thought I could read them. U. U. I feel like it's a bit of a hassle for those who post. I don't post, though. Yes. U. I guess the chance of people getting involved has increased even more. U. That's certainly the case. U. Only in those cases. Maybe it's better not to translate it. If it's said in English, it might be a little less damaging. I don't know if it's raging or not. Uh, uh, uh. U.U.U.U.U.U.U.U.U.U.U.U.U.U.U.U.U.U. U.U.U.U.U.U.U.U.U.U.U.U.U.U.U.U.U.U. That's right. That's right. If possible, I've stopped the habit of going to see what's trending when I have free time. I look at my own timeline, but I've stopped doing that because I end up looking at trends. Now, I'm just running this from Codex's automation about three times a day, and only picking up news that I'm interested in. SNS is great, really. Yes, it's not good. I understand. I've come this far, but... I think everyone thinks that's not the best news, and that's true.However, at the time we were recording today, there was an announcement today, and today was yesterday.The date changes every time, so let's not worry about it.When Claude Opus 4.8 was announced, it was said that it was a very honest model, or that it had such improvements.It was the only one that lost to GBT5.5, but it's a little difficult to score around here. I think you can see that the score isn't everything by comparing 4.7, 4.7, and GPT5.5.The score isn't particularly great. I'm not buying all of them, but I think there are people who like Obus 4.7, but personally, GBT 5.5 has a better feel, or rather, it feels like it's working hard, so there's talk that the model has gotten smarter this time around, or that it's become more obedient, or something like that, and the dynamic workflow that has been a hot topic. I wonder if some data has been officially released.I don't really understand, but it seems like there were some people talking about something like that.Yes, well, this workflow came out.Well, I think that one is more, um, rather more prominent.Yes, and this is what is called dynamic workflow.Yes, something like that. From what I can see, it looks like Lazy Docker or something like that, but ah, that's true. Well, that's what it looks like. It's a TUI interface. Hmm, how do you call it? It's like, um, it's like hitting a huge job all at once. To put it intuitively, there are two types of them, or rather, two types of workflows, which run this agent, this agent, and multiple subagents in parallel. Well, there's a workflow that allows you to do a lot of things in parallel, and then you make it even bigger, and you do it all at once to the goal, and it's called Ultra Code, and it's written in the book below, and it's called Ultra Code, and it has a function, and it shines, and it shines. Yes, it exists, so I was wondering how to use it, and I researched it, and I researched it, and when I had it done a few times today, I was asking questions like ``Should I do this task using workflow or ultra code?'' I was thinking, ``This task isn't that big, so I don't need a workflow,'' or this one is small, so I don't even need a workflow. They say things like, well, basically, if it's an ordinary task, it's a workflow, it's really big, it's something like language porting, it's like there's a ridiculous number of files, and what else? Or when things change when a version is updated, it's like using Ultra Code to do something that would take several hours or days. So, like the goal in Codex, Claude also has a goal, but what Ultra Code does is to do it in parallel using a platform called workflow, and then autonomously continue to work hard. It feels like it's going to work, it's going to work, it's going to work, and I've done one thing. I see. I see, it's like making all the changes all at once, and testing and everything, and somehow passing it all on your own. It seems like they'll do something about it even if they don't have a clear design in mind. Just tell them what they really want to do, and then they'll do their best to do it really well. Well, the first instructions are important. So if you have a clear design like this, you've already solidified it really well, then it seems like it's okay to use a workflow or something, but if the goal is clear and you don't care about what's inside, you might be able to do it in a nice way with Ultra Code. It's really easy to understand, isn't it? Well, if you wish that the goal is to make the code that is moving with a jig work 100% at the end, it will come true like this. What, what, what, what? What was moving with a jig will now work completely 100% at the end. When the result is exactly like that, even when it comes out, the contents will be clear. Well, it's possible for things to become incredibly messed up, right? Of course, it's possible that the structure isn't what you expected, or that there's some messed up code, but I wonder how things like that will turn out in the future. I wonder how far people will go to see them. Well, I mean, if this bar is fixed, then there will be people who maintain it. I don't like it, right? Well, if that happens, it won't really be a product that you touch. When you think about it, in that sense, this bar itself is amazing, but you might have the impression that its appeal as a product has diminished a bit. Yes, that's right. Well, I believe that you've looked at everything properly inside, but when I said that I would move on to the last part. It seems that the maintainer of the last part said something like, ``Don't do that at the end.'' That's right. Yes, I'm sure. I think it was the maintainer of the last part. Yes, the person developing the last part told him that it would be better not to do that, so he decided to make it the last part. Yeah, I see. Maybe it was necessary to prove this. Ah, that's true. It's possible that he didn't actually want to make it the last part. Yeah, well, this is kind of difficult, isn't it? I'm trying to figure out how far and what is best for which language. Yes, I feel like there are a lot of people who choose the ending too carelessly. Yeah, I think it would be nice if people looked at it calmly and looked at it calmly. Well, I think the general public is looking at it calmly and thinking about it. I think that's true, but yeah. Well, there will be a lot of people in the world who don't like that, so yeah, the last thing is hot right now. The main premise is to make the code readable even without it. Honestly, yeah, because you can't read it yourself, and you ask AI for help every time something happens. It's too painful, isn't it? I was thinking of asking, but the token has already run out. I'm sorry, I can't work right now. If that happens, it'll be very difficult. Yes, it would be very difficult if I were to go into a field where AI cannot be used. Well, I don't feel like I can do it either. I don't know how to do it, how to develop it, how did I do it? I totally understand how you feel, but on the other hand, I was forced into that field, and... I think it would probably be fun if you were told to do your best from scratch.That's why I think it would be fun in its own way.To be honest, there was a lot of news this time.That's how much time I spent, but it's like this.Since this time, umm, isn't open AI included?Yes, yes, there are other things other than open AI, aren't there?There are various announcements from major companies. I just said that there have been a lot of moves. It's all good. It's probably going to be complicated from now on, but it's really in that kind of territory. Now that we've come this far, yeah, that's right. I'm starting to feel like I want to develop it. That's right. I can sign up for everything. I don't have to apologize. Other than that, I really want to talk to GPT5.5 professionals, so I'll continue to do this. I'm going to press Chat GPT, but there's something that's comparable to that, right? Well, Google has one, right? It's Ultra Sync, isn't it? It's not Ultra Sync, it's not Sync, it's Deep Sync, uh, yes, yes, yes, yes. I'm not interested in deep sync, but I'm really interested in it. Well, on the other hand, you're right. No one is even talking about 5.5 Pro. Ah, I do think it's best to let him design it. Yes, it's really a Matsusaru product. Yes, nothing has been decided. When you start from the real thing, it's really good, so yeah. Sorry, I went off topic a little, but yeah. It's like this. We're sharing our thoughts on things like AI news and other things, so if you'd like, please subscribe to the channel and click the good button. Thank you very much. Yes. So, thank you again this time. Thank you, thank you.