Claude Code Opus Plan Mode // 1M Context // Testing ++

Okay, everything looks good. Let's test the chat. Say hello in chat. I did an update to my YouTube server, so you can send messages now. should work pretty good, but it can't monitor. Hello, Aneruda. Nice to see you here. Yeah, there's been a lot of updates like on Claude and stuff. Just want to take a look at it today. I think I'm good. How are you? I think like this was pretty interesting. The opus planning set

execution. Kind of looking forward to trying it out. This is what I was trying to do autonomously. Not autonomously, manually. But of course it's hard in between. Hey Edgar, nice to see you. Have you I guess you are an open coder now. But did you take a look at this opus planning set execution uh workflow or automated workflow? Okay, then you have good time. That's the perfect time to start learning. Uh, I guess Elgar. Hello Jack.

Yeah, if you try during the stream, let me know what you think. So I think the basic the idea is that the new model that's opus handle the planning while set does the work. So basically automate what everyone is doing manually. It's just going to save you like a bunch of opus tokens if you want to use it for planning. I thought I could go through some of the comments. I'm going to try it. Hello Narius.

I have done this manually. Yeah, I have too. It's just a hassle, right? There he is. you got um the crypto and that's kind of not the only good news from Antropic. Hello Sherik. We also got uh 1 million token of context. I haven't really looked too much into it. I hadn't had time. Uh, open code really gets true to me. Back in cursor yesterday, the man um, the convenience is something I really missed. Yeah, check

out Claude. Yeah, you should do that, Edgar, if you have the max. Hello Tom. They already updated Claude code with 1 million. That's what I wanted to find out today. I haven't really looked too much at it. Yeah, that's good. Narius, I think that was the best way to solve this. Tom, hello. How are you? I used open for planning in cursor as well. Okay. Yeah, we can use that. I did try out the Thanks Anuda.

Appreciate it. I did try the open coder plan agent to build agent today. Uh Jack, you have the 5x me too on the max 5. So I'm not sure how I would check how much context I have context token I have. I run this. Have you seen this? This is uh claw monitor I think. So this kind of keeps track of kind of your token usage, time to reset everything. So it's pretty cool. Yeah, just look

for I think it's just called Claude Monitor GitHub or something like that. Like always, if you have any links to share, if you have any links or articles or videos, just put it in the I'm going to pin it. Try opening claw set in open code. It will show. Okay, I can do that. But do I need the entropic API or No, I shouldn't. So let's do models cla 4. Should we just restart this open code?

Hello maple dev mpxcc context. Okay. Yeah, I'm not quite sure. Uh, we need to investigate that a bit. So, where do you usually see the context? Is it somewhere here? I'm not quite sure. I'm not so familiar with open coder yet. There we are. Hello. Yeah, if it's L. Okay. Okay. So, okay. It's like that. So, 11k is 5%. Okay. I see. Yeah, that's 200, right? Do you use perplexity? Have used it, but uh not at

the moment. It's it's it's good though. I agree. But I don't do so much research really. It's like I do some but not too much. So I just wanted to take a look at this claw for set now supports 1 million tokens of context. So claude for set. So it's only set supports up to 1 million tokens of context on the anthropic API. So would that mean hello lick if I open up cloud code I switch

to API usage then I would have 1 million is that the way ah okay Edgar two versions So this is the open router provider. Yeah, I guess if I use the entropic API, I could probably do that in cloud code too. So if I do something like uh if I leave here like login and I can select and API usage for billing then I probably would have the 1 million right but I'm not going to use

that. That's going to be expensive. I guess that's the way they did it because it says uh supports uh on the anthropic API. I hope they can switch it over to the plans though. That would be great. They're probably not going to do that though. Long contact support for Sonet forest now in public beta on the API and in Bedrock and Vertex coming soon. Yeah, we know what long context is. API pricing. So, they have different

pricing, right? So, I don't know why they did this. It's a bit strange. So, prompts below the 200k is three in 50 and out. And if you go over 200k, it's 6 in, 2250 out. That's kind of annoying, right? How do you know? Do you get like a warning or something? This is basically a double price. Hello, Sassu Moava. Nice to see you. Yeah, I just looked at Claude Monitor, but I don't see the tokens. Yeah,

I think it's just the API Jack. I agree. I guess if you use Entropic Direct, you will see the million. Yeah, but it doesn't look like it's it looks like something like 200, then you switch to like another API or something and then you're going to pay more. The plans are also using the API, I think. So, but but I need to know when um but for the plans, this wouldn't matter, right? This will be baked

into the price. So it's probably not on the plans, is it? But you can do if you have like a code base, you can have prompt caching, right? This would uh reduce latency. So if if you have like a big code base, this is pretty nice because you can cache basically your whole project and then you can save some money and latency. Yeah, this is really good for like bolt new lava ball cursor, right? I bet

they are pretty happy. API with customers tier 4 and custom rate limits. Ah, do I have tier four? Okay. So, it's like a special call. You're going to do the betas context 1 million H. Okay. Not exactly what I was hoping for. I was hoping it's just going to get baked straight into cloud code. So, it looks like it's going to be like a Yeah, I I agree, Edgar. I don't think this is going to make

it into the plans. Not yet, at least because it looks like an separate call here. H okay, maybe in a few months, I guess. I don't think I'm going to use it then. It's just going to get so expensive. But I can see like cursor maybe offering this on the plan. I'm not sure. Would you pay more? Would I do like 5x with the 1 million context window for 100 or for 200? I'm not sure. Before

I before I hit 200k, performance degrades. Yeah, I remember we talked about that. So, I'm not quite sure. I'm really happy with um for longer things. I just use this Gem Google AI Studio. They already have the 1 million, right? I think this is great. I use it all the time for long things. Videos. I do like a couple of videos of context, maybe some PDFs. I fill it up here like four 500,000 and it works

pretty good. I did like 500 PDFs. Yeah, it's really nice for video. I do my timestamps. It's just a great It's probably the best free option here, right? By far. But this is kind of interesting. I think this could be pretty cool. I had some fun earlier with the GPT5 and I tried kind of the the plan build agent. So I did a plan with GT5 and I did and I built it with GT5. I created

this uh what did I try? Yeah, I think it was just a stupid game or something. But the game was pretty fun though. It created a game called Snaky. Oops. Python snake. It was pretty fun. So, there's like two modes. There is this mode and then it switches to like a ghost mode. Now, I can go through the walls, but sometimes you can't go through the walls. Now, it's the walls are deadly. Oops. And sometimes I

can go through the walls. When the ring is blue, I can That was pretty cool. It's really responsive. That was pretty fun. Yeah. A Studio is great, right? No, I haven't. I haven't seen the new Arc Hi3 preview. Do you have a link? Could you link it? Is it like a video or something? I can check it out. Arc Hi 3 preview. So that's probably going to be the new um the new benchmarks for like Hello

Reuben. Taskmaster AI. I remember we had a look at that. Okay, so there's the link. Didn't we have a look at Taskmaster? Uh, we had a video. I don't think I had a lot enough time to watch it yet, but it was this like Reuben. this one. Uh, I had been a bit busy yesterday, so I didn't have time to watch the video, but I'm going to do it. Is it this one? Looks pretty cool. Yeah,

I know, Tom. It's uh this one, right? I'm going to have a look at it. Yeah, I'm definitely going to check that out maybe tomorrow. Anyone running safety hooks in cloud code to prevent dangerous commands or claw reading envir uh sensitive but do you run like um do you run uh give me a So, what event do you run? Do you run a pre-ulus event, Maple? Like the pre-ulus event to check if it's a dangerous command.

I guess it has to be the pre-ulus event. Running code in YOLO mode in dev container. That's pretty fun. I think I read a post about that. I kind of want to try it. I'll show you. I think it was hacker news. Okay, I guess I have to just go in here. Let me try to find it. Can I search? Maybe next page. Yeah, this one. Claw code is all you need. He ran uh VPS. I

think previous tool checks for RM. Uh yeah, that's pretty clever. Yeah, that rm-r command can be scary, right? But have anyone had a look at this? I thought it was pretty cool. It's like a big big article, but he's doing Hello uh JSlime. Nice to see you back. You can get a preconfigured dev container from Antropic. Hello, Nick. that uh gives access only to their servers and your local commands. Ah, I think I've heard of that,

Reuben, when you when you mention it. So, this guy did this um giving a root VPS to a robot and saying go fetch. The first thing I thought I was doing was exhending that loop maybe indefinitely. So he fired up a cheap VPS on Hetner, installed uh an authenticated cloud code, messed around to write its own prompt, building autonomous startup and looping logic to keep running forever without my input. I thought it would looked pretty funny.

So, it's just going to go indefinitely with the dash dash dangerously skip permission command, right? And it's going to have this waiting three hours before restart. I hit a small snag snag andropic decides that running Claude as root with dash dash is not allowed. You can get past this by dumb nanny state like running. Okay. Is sandbox equal one? Not financial advice. Yeah, this was pretty cool. Just got here. Uh title real men. Uh we checked

out that J slime. Uh, and we kind of found out that um it looks like the context window is um a separate call. So it looks like it's the 1 million is like a beta call for tier 4 API users. So it's probably not going to be in cloud code yet. I think there might be some way to set it up, but um I think I'm just going to wait a bit to test out the 1

million context window to see if it comes into the plants. But it's pretty hype though. I tried that some time ago, but I wasn't able to authenticate because I needed a browser, but I need something remote. Yeah, I see. Uh, I can link the article Edgar if you want to check it out. Antropic devs uh said to decontainerize their projects, yolo and walk away from the computer. Tell uh until done with all tasks. Yeah. And you

can do the I have done some hooks sometimes that is called I like this disable all hooks by the way. I think it's stop the stop hook and I use like twio to send me a text message when uh claude code stopped. So, I got like a text message. Claw code has stopped running and then I can rush to the the Mac and start it again. Was pretty fun. Okay, so this is Arc Hi3 preview. Didn't

they do a video on this? Let me see. It's 18 minutes. That's a bit too long. YOLO. Yeah. Yeah. There there is a video here, but it's 87 minutes long. That's a bit too long. But basically, how does this work now? So okay. What am I supposed to do here? Ah, so you can maybe you can have like some um should we try it with uh what is the sneak peek? There is no instructions intentionally. You

must play the game to discover controls, rules, and goal. Okay, so there's no instructions. Let's do Can we do this on manus or like shot GPD? So let's do GT5 thinking I guess. Can we do agent mode? Solve this. Can we see the container here? The first I did on accident. Yeah, I'm I'm not quite sure if I'm going to spend um Okay, so let's see if um chat GPT can solve this. It's just going to

try to find the rules. Do people think it's going to solve it? It's a Norwegian. Sorry. Now it's in English. So, now it's just going to try to figure out how to do the controls. Hello Tim. Let's do a poll here. Uh, chat GPT going to solve this. Okay, Edgar, thanks for dropping by. Oh, it's cheating. Why did it do that? I'm going to want No. Yeah, J slime. It's pretty cool. I watched a YouTube video

the other day. Uh, repo that let you remote into your CC to control it from your phone. Yeah, I've seen some um messages from this too. What is it doing here? What is it? I'm just going to follow up. Don't leave this URL. Maybe. Yeah, it's not going to solve it. Let's give it like a couple of minutes. Hello Jamie, you use tail scale and terminals on iPhone. I I kind of want to try that. So

how do you Is it like a lot of work to set it up, Jamie, or is it pretty straightforward? If it leaves the the game again, I'm just going to stop it. Net bird or tail scale. Okay. Now it's just clicking. Okay, so it found out it could click and the the color changed to red. So isn't it like the solution is in the middle, right? So we're going to move the red cube to the middle

and you're going to have white cubes and maybe blue that something like that. Pretty easy. install tail scale uh on the your VPS and your phone. Uh then just copy the tail scale IP for the server and SSH into it from the Terminus app. Cool. Maybe I'll try it this weekend. So, how is chat GPT doing here? It's trying at least. Has anyone seen this uh Pokemon streams? Here's a question for uh some uh Claude Code

veterans. Can I create a custom slash command slash test command that will just run a bash command and output the results? Yeah, you can do that something like something I can show you. So it says Okay, so now it has a plan. It was in the region. It's going to try to fill some routes with the colored squares. Seems like everything I do runs to the CDM sends to the API, and summarize the results. Okay, I

may have to read that again. Custom slash command. Uh, that will just run a bash command and output the results. We can try. No, this doesn't look too good. I'm just clicking the same. It just keeps clicking the same um square every time. Okay, this was a better plan because now I can do action one to five plus click. Just being able to host Net Bird yourself is just a bonus when it comes to privacy and

security. Cool. Haven't tried that. This isn't looking too good for chat agent. Nice to see you. How are you? I think he just gave up. But now it's going to start looking for um I think we're going to call this So, it's just going to start searching for solutions again. Just keeps cheating. So, we're going to stop it. So, it found a Let's see what happens. It did cheat, though, but it says it found a solution

on X. I don't think so. It has some kind of reasoning log here. I don't know what it is doing. But that's cheating. Yeah, I guess this is quite hard. So, let's stop this. Um J Slime L5 sound waves or sound waves in CC to use plan mode first before using a sub agent or do you prefer sub agents assisting with the planning? I think I just use plan mode first. You could mention in your plan

that you have different agents. You can tag them. I think that's better. Yeah. Uh JLine, let's see. Did you mean like So if you do because this is open code, I meant opus plan. on if it uh uh let's say can do try two things I wanted to See, Pedroine, nice to see you. Uh, we're not building anything. Wow. RX, are you serious? That is crazy. What the hell? I never experienced that. Are you here, RX?

Really? Wow. Wow. I don't really know what to say. Thanks so much. I hope you are here. That is crazy. Thanks so much. Didn't expect that. Wow. So, RX just gave 50 memberships. Wow, that's just insane. Okay so that that is a lot. So, what are you up to, RX? That was pretty crazy. I didn't expect that. So, basically, that would mean like everyone got a membership. Uh, wow. Uh, I really appreciate it. Thanks so much.

Is there anything you want to check out, RX or Thanks, man. Really appreciate it. That's pretty crazy. And that's perfect timing because uh I have some plans to ramp up the membership, too. I'm going to post more stuff on the on the the GitHub. So, I guess I have some people to add on the membership GitHub then. That's just crazy. Yeah, John. Sorry. It's It's random. I I can gift you a membership. I don't know how

to do it, but uh How do I do that? I'm going to gift you a membership, John. I'm just gonna I think it's here. Um, okay. So, let's see now. I don't know what's going to happen. Okay, John got one. Okay. Uh, I can do more, but I can't do I can only do five now. But thanks, Tim. You will probably the next one. But RX, thank you so much. That was really I didn't expect that.

Kind of threw me off what I was doing here. Hello, Nathan. Nice to see you. Yeah, that was really nice by RX. I didn't didn't see that coming. So now suddenly we got a lot of members here. Almost enough to just do like a membersonly chat. That was pretty crazy. Uh I never seen that before. So yeah. So now it's going to be a lot of green people in the chat. That's pretty cool. And that's perfect

because I have some membership plans coming up. That's going to be nice. Yeah, Pedron, you will get one. I'm going to I can only gift five at a time. And maybe I can do Okay, I I can do five more. Let's see what happens. So, it's just random. I can't control it. So, it's just going to be random. Okay. So, yeah, Tim and the Pedro, you didn't pick get picked this time, but next time. But, uh,

now we are a lot of people with um, one day you'll get it. So yeah, that was perfect timing because I'm going to keep I'm going to create like a new repo for members where I try to post some stuff. Most Are you serious, RX? Uh, is insane. I don't know what to say. Uh, thanks so much. I hope you I hope you got a lot of money. I guess you do if you can spend this

amount of money on memberships. Uh, but yeah, thanks a lot. Uh, no, I really appreciate it. just kind of threw me off um a bit. So yeah. Yeah. I don't know exactly how it works. I've never I haven't seen this happen before when I was streaming. So, but uh a lot of people got it though. Yeah. Uh, Benjamin, I'm good. We've been just having like a membership spree here, so pretty crazy. So, RX has been giving

out crazy memberships. I was a bit thrown off. Yeah. Is it Mr. Beast? I don't know. Maybe. But I really appreciate it. Pretty fun. Uh yeah, that's kind of the first I've seen like videos of this, but I never had it happen to me before. So, sorry, I've kind of forgot what I was doing here. Anyway, I was a bit thrown off. But I think we were trying to look at uh here and do some commands.

Is it command? I think so. Let's check the membership is uh a while back I used to have uh I used to have like um you will get there. I used to have like updated GitHub repo. Yeah, thanks Maple Dev. That's what what I kind of wanted to do. That's why I started streaming again because um if you just do videos, it's really hard to get to know people and I don't really have time to sit

in the comments all day. And I tried GitHub uh Discord servers, but uh I just feel I don't spend enough time there. So, I was wanted to try streaming again and I think it's just a better way and then maybe we can do something else. So yeah, pretty fun. So I really appreciate everyone tuning in. Thanks, Nathan. Really appreciate it. Uh we're just doing some just checking out some news and stuff. So um we just uh

commands right. So yeah, thanks RX uh Mojave. He's been going on spree. Pretty crazy. So, uh, hello Liqua 1 million context window. That's only for the API page or does it work on the Yeah, that was what we was looking at. So, we kind of had a check on, uh, the documentation here and it looks like where did I I don't know if I put it up. Yeah, we can see here it looks like the 1

million context window is uh like a separate API call. So, you have to do a a 1 million context here and you need a tier 4 uh usage tier, right? But, uh hopefully I can get into the plans soon. So, J Slime uh what command uh did you want to try out? I kind of forgot now. What he was this one was community forum. Maybe we're going to do something. I'm going to think of something. So,

JLine, what was your command you wanted to try again? Okay. Yeah. Um, do you mean this? I'm a question. Command. Is it command or is it commands? It's commands. Is it this? I guess you try this. I think uh Nathan, you have a tier four. Yeah, I I might have it too, but uh I don't know what to try it on, right? So you do you mean this uh you get the API call here? Is that

what you mean? Or like the response? This is kind of your issue. You only kind of wanted this. Hello Harsh. Yeah, I know what you mean. You just want the command line to output the results. I don't know. I haven't tried it. Let's uh we can kind of try the ask code. I guess we only want But what do you mean output the results? enable gifts to get one. Okay, I I don't know really know that

system. Hello, Decorate. Can we buy our way to tier 4? I guess you just have to spend a lot of credits. So, I guess uh J slime you meant something like like this, right? But uh it is working though. If we do if you do sleep 10 right post tool uh post tool call hook if tool cost matches x stop. Yeah, maybe something like this. Yeah, but what if you do? So, it is I' I've used

uh So, this is going to sleep now, right? For 10 seconds. Is that what you mean? Like this? So, this doesn't return anything CC user setup. So, I wouldn't mind a custom command like CC like us search that would be run. Ah, okay. I see what you mean. Hello, MS call Python script on user prompt submit. and check for uh run my test. You can then block it from going through the API. Okay, that's pretty interesting.

Arian Claude Code Sonic got a 1 million context as well. We had a look at that and it looks like it's just on the API off the plans now for now. I can't confirm 100% but that's kind of what it looks like, right? When the user submits a prompt, check for run my test. You can then block it from going through the API. Okay, I'm going to try that. Maybe not now, but I'm going to check

it out. Yeah. Uh JSL, if you do some more testing, let us know what you find out. Can be interesting. I have done like if you were here on the live stream I did like so you can do something like let's try the new plan by the way let's do opus plan mode yeah I think that's a good idea from Maple So, we want to plan where we're going to do uh we're going to do a

fetch. Then, we're going to do a bash sleep 10. Wait for 10 seconds. Do a new fetch. any website bash sleep for 10 seconds. So we're going to do three cycles of that. Can you set up this plan? So this is what I've been using sometimes to do like a long run. It works pretty good. So you do fetch sleep 10 fetch like a cycle. Sonic, I think you enable it on YouTube or something. I'm not

quite sure, but I'm going to donate more memberships. So, you could do something like this, like fetch, sleep, fetch, fetch, sleep. And you can set like a long run here. So it's going to fetch, sleep, fetch. So it's going to sleep. When it's done, you can fetch sleep. If you want like, let's say you wanted to fetch like an update on a site every five minutes, you can do it like this. I tried it. It works

pretty good. Fetching and sleeping sounds like my life. Yeah, pretty much. It's just a work around. Hello. Hello. Hello, Unreal. Link. Okay. So, that's fine. So, that works. Should we do the Let's try Let's try the same thing I did with um Let's delete this. Doesn't really matter too much. So, let's do SLM models. We had a look at this. We can do a quick recap. So, they did a new release. So, they finally automated the

Opus planning plus the set execution combo. Uh, this is something I've been doing manually, but uh, now we have the Opus plan mode. So, this is going to use 4.1 in plan mode. Otherwise, it's going to use Sonet 4. This is really nice because new mode lets Opus handle planning while Sonnet executes the work. basically automates what everyone has been doing manually. So instead of keep switching between models like I've been doing, you can now just

go opus plan mode, you can go to plan mode and you can do create a snake game in Python code with a surprising twist. Save it to snaky2.py. Pi create a plan for this. Yeah, it's a great feature. I think it's pretty cool. It's just going to save you a lot of Opus tokens, at least on the 100 plan, right? I think it's going to be pretty good. I'm looking forward to testing it on like a

bigger project. Okay, so here's our plan. We have a portal game polish. So this plan now was generated by Opus, right? And when we auto accept now, I don't have to stop it and switch model because now we're going to do set autonomous. And if you're going to do any planning or thinking I think in between here, it's also maybe going to be switched to opus. I think so. That is the idea, right? But I haven't

tested it enough yet to see how it works in practice. But the comments here are pretty positive. Antropic is cooking. I agree. Nathan says he's on the 200 plan and I think uh almost unlimited Opus tokens. People under 100 seem to get anxious, but 200 is great. Yeah, that is kind of my uh experience. I can run a bit of opus, but it doesn't take long before I get the approaching limit on opus, but hopefully this

could help me out a bit. Uh, I also had like a video. I did like an MCP server with GD5 that did the planning and I did the execution with uh, Sonnet. But I think this uh could be really nice. Oh, we have some errors. It's just a typo here. So, here's our game. We have snake with a portal twist. Okay, it didn't work. Oh, it's just a timeout. Let's try it. Python snakey2.py. Uh okay. Wow,

that was hard. So, we can go through these portals here. And if we do, if we have a portal, we can Was it just one portal? So if you suddenly got here and you have like a subscription and a membership, uh that is going to be RX. He gifted 100 memberships. So yeah, pretty crazy. I never experienced anything like that. So that's pretty crazy. So KML uh you can say thanks to RX. Yeah, was really generous.

Pretty pretty crazy. Um, I felt spoiled. I almost only use Opus. So, Nathan, do you see like any big differences if you are working on something advanced? I don't know what you're working on, but thanks K. Appreciate it. On the $100 month plan, I typically use Opus for your planning and reviews and Sonet for the context. Yeah, me too. So, Jlime, I guess um this is perfect for you too then the switch like the auto switch

here. Unreal. So, RX did 100 membership giveaways. That was pretty crazy. So, he gifted 100 memberships. Okay. Yeah, J Slime. I think it's pretty cool, but I haven't tested it um too much yet. I was switching manually. So, this is nice feature. Yeah. Yeah. The padron I think you need to enable being able to receive memberships. I'm not quite sure. Yeah. I think the switching is pretty cool. So it would be nice if we get switching

and they will add the 1 million context in like uh plants uh a mobile I have an old account was set as a brown brand account. Okay, I see. I'm going to do some more gifts if people are interested. Anyway, we'll do it like the the next days or uh next week. So, I have to change profile to be added to the membership maybe. I'm not quite sure. Yeah, this was this like so let's try to

build a 3D snake emit opus planning and a sonet execution and I'm going to kind of monitor uh here. So when we run this now, my model distribution so far has been 97% Sona, 2.5% opus. So I kind of want to see when I'm going to run out here, right? But if you're doing like a big project, it's just going to eat a lot of tokens fast since I'm not, it's probably not going to be too

bad. Look at this snake game. Whoa, that's hard to play. What happened? There's a wall there. I don't understand anything. How do I go down? I I don't know what this is. It's a gravity. Well, how do you connect uh your claude code monitor? I can show you. So if you go to claude code I think it's just called cloud code monitor or something monitor. Uh I'm going to link it. So I think if you go

here uh just go to the there's like a pinned post You can find this web page here claude code usage monitor and you just do when you install it just do claude monitor and you can get this nice overview here when you're going to reach your limit how far you come. You can create a private Reddit invite only. Yeah, maybe we have to do something Maple. I I'm agree would be cool to do something like community

thing. I'm going to think of it, but that's nice. We also have something else that is called Hello Platal. Yeah, you're a member now. So, a big thanks to RX for donating 100 memberships. Can't believe it. So, we also have um CC usage. What's it called? Uh, and so you can see the last 30 days, right? Um, so this is the last 30 days. I have spent for API cost $532 in API cost on cloud code,

but I only played $100, right? So that's CC usage. You can kind of check your monthly monthly setup. So I spent $500 for 100 CC users monthly. Okay, maybe that's the thing. But how what is the command? Is it npx? I haven't really seen. Hello Martin. Okay, here we got it. That was better. So, we have done some haiku. Yeah, that was much better. Uh Jamie, thanks. So, in August so far, we spent $200. And maybe

it's is maybe it'sn't monthly. It's maybe like one and a half month. I'm not quite sure, but I don't think I have spent that much. Some people are like $2,000. Yeah, Jamie, I am on the $100 plan. Hello. Uh, long one. Uh, Ano, something weird is happening. I can't access cold code at all on cursor. Okay, that's pretty strange. Did you try to do claw code version? I had some issues with that one time. or like

clawed version. I was kind of on like an old version here. I couldn't open it. So, Jamie, you're on the pro plan, but thinking of upgrading. Yeah, I find a $100 pretty good. I never run into any issues, but I'm not like Nathan and using Opus all the time, though. Yeah, Martin is also on the 100. I kind of like it. It's kind of in the sweet spot for me. For 200, I kind of feel I

have to use it all the time, but for 100, it's perfect for my use, building like smaller projects and not kind of managing like a huge project. I don't do so much managing. Yeah, I think 100 is good for me, but some people need the 200. Uh, I know a lot of people here have the 200. So, this year I'm also uh I play something called fantasy football. I also I'm doing like my fantasy football team

on Clo. So, if you're interested in fantasy football, I have a new YouTube channel where I post videos on Claude Code and FBL. I'm only using Sonnet at the moment. Yeah. So, Jamie, you can uh definitely try the uh Hello. Definitely try the new Opus plan mode slashmodels. I guess I have to be in cloth first. So that's pretty good. So let's try out our um get all um fetch chat messages. Who has been the hero

of the the stream so far? Yeah, I know Martin. So, hello Mug. 1 million contacts also were forcloed code with anotropic account. You could do like a separate call, right? But I don't think it's in the plans yet. So, we ask Claude code who has been a hero of the stream yet. So, it says RX is clearly the hero of the stream. Here's why. RX has been an absolute legend of the stream. He gifted 100 memberships.

Multiple people are thanking him. Comments like RX is a legend. Thanks RX. RX works at Antropic. Many new members expressing gratitude towards him. So yeah, we have a Claude code kind of hooked up to the chat here. So uh it's pretty cool. Pick your uh emoji. Hello Phil. So now we can do a send a message with cloud code. Hello Rick. I have been I don't know how people use many people use it but I really

I enjoyed open router with uh open coder. It's pretty cool though. Yeah jam it's pretty cool. Smooth. I agree. Do you use T-max? I haven't, but I have had a look at it. So, I don't know how many of you have tried open code, but uh I think it's pretty cool. I really enjoy that we can try different models. So, we can do Kim and K2. We have the new local models from OpenAI, GLM 4.5. We

still have Horizon Beta. We don't know who what um provider this is yet. I built an MCP for T-Max Gemini as the orchestrator and it's been sub code sessions as sub agents and can check in with their progress. That's cool. You're a mod for Adam on the open code team. Okay, cool. Jamie Quentry code is really good. Yeah, I enjoy that too. J Slime has a cloud code router now to set up um in case Antropic

is having issues. Nice. So, how does the other models perform in Cloud Code? Uh Jime, have you tried uh Kim K2? Crush is pretty sexy. Open code seem not to rank as high as gush coders benchmarks. Okay. Yeah, there's a lot of testing on this. But what is crush? I haven't se Is that like a different version of open code? Has anyone tried uh GPT? I guess you mean oss via MCP server uh on CC? I

have not. I just tried GDF5 Maple. Hello, push the model. How's the testing going so far? Uh we haven't tested it too much uh push but um it looks like it's going to be a very good feature at least the planning mode from Opus because it's just going to save you it's going to save you a lot of tokens and you don't kind of have to skip the manual part. Crush crush is the OG open code

different code base now. Yeah, I know there been some drama there, wasn't it? Or something. I haven't really follow along too much, but we can check it out. Chararma, they're using charm. Okay, looks pretty good. I might try this, too. But is there any like performance differences or is that basically the same? Is it mostly UI? Uh Rick, how do you decide what's worth coding? Most of the times I just think of something, code it, doesn't

work. Does it work? And I never use it again. Lol. Yeah, I see what you mean. That is what cool about like using cloth code. But uh if you have like a good idea uh just go for it, right? I built my um video course platform uh using cloud code and that has uh worked pretty good. Love the TUI though. Yeah, it looks good. Have you heard about the GitHub news? No. Tell me more, Pedroin. I

wish open code had auto context compression though. Yeah, I don't know. Maybe you kind of do you lose kind of track of but I guess you can see it in the top corner right? Create uh animated mine. Craft TNT explosion in HTML. Oh, okay. So, this is over. I forgot about that. Don't you kind of see the tokens uh Jamie in the in the up in here GitHub news. Okay. So Oh, you mean like compressing? Yeah.

Okay, I see what you mean. So if you want to do it, you need to do it manually. Save it to a file, I guess, and just read that file. Could be a bit annoying. GitHub is no longer independent at Microsoft. Ah, so they are Microsoft now. I guess works well for cloud models, not so well for local or others I tested so far. What about Kimmy K2? Okay, I see. I can check it out. So,

it's slash compact. Yeah, they have that. Okay. Awesome. just save it. So, how is your impressions of uh GPT5? Has anyone like gone full GP5 on like cursor or open code or anywhere? I've been pretty happy with GT5 for the things I've done, but I haven't done anything like complex yet. Yeah, Chis, we checked it out, but it's like a special call on the API. It's uh it's like a where is it? I can find it.

It's like a one context one m in the beta header in your request. So, you kind of have to I don't think it's autonomous. Usage limit in cursor for free tier is not much. Yeah, I've tried it in cursor on the the free, right? Oh, useless duck. Very sexy. Very expensive. Yeah, right. think it's going to be expensive. They kind of changed up the pricing, too. Um, I'll be right back. Yeah, I would hope we'll get

to the plan soon. The context window. That'll be pretty nice. So yesterday I had a lot of fun by doing um I really wanted for a long time to do like um do like um poly market app. So yesterday uh I just created one. It was super fun. And we can do what I point them. Oh, do I have to do some uh I need to update the events, I guess. Yeah, I had a lot of

fun with this project yesterday. It was super fun. So, if you're going to use it, you need to update kind of you need to go to Poly Market. You need to find kind of the new uh So, it's going to be today uh at PM I guess here. So, I need kind of a slug, right? And I need It's a bit complicated to set it up though. Then I need to go to my repo. I need

to grab this. Right. This here. That's this. I want to grab this. Grab this. Look. I guess it's just change the I need the slug and I need the event. So, what I'm going to do next on my Poly Market app is I'm going to I'm going to add some um Was that wrong? I need to add some um okay some polygon I think it's called on the poly market app and then I can um I'm

going to make two bots that can kind of trade on the page. All right. So uh that is my idea. I'm going to try that maybe over the weekend. I'll see. So, that is going to be something I'm going to try over the weekend, I think. Poly app. Let's do cloud. Yes. And do plan mode. Uh let's do models first. Models opus plan mode. Um read uh we can do list path or get familiar with project

update events to the following. create uh a plan for this. Okay, let's try that. Try to update my my project here. So, how are we doing on the if we go to poly market, we go to tech, we go to AI, which company will have the best model at the end of August. So we are still at 70% Google. So there's a lot of rumors about Gemini 3 coming out before September. I saw there was some

Wasn't there like some Let's see now. Looks pretty good. Let's do that. There were some tweets I think Gemini three uh death star tweet I don't know start. Yeah, something like this. I don't know. I saw something. So, I think Google may have something coming up soon. So, that's going to be pretty interesting. I see you used uh slashmodels to choose plan mode. Yeah, but I just set uh I'll show you. Do I have another one

up here? Uh, okay. So what I did the sound waves models there's something new here now you can do there's something called opus plan mode right so now we can do for only planning we can run opus 4.1 and set is going to do all the executions and writing the code lines so we only use opus for planning and this will save us some opus tokens right instead of doing it manually way. So that is kind

of what I did here. All right. So I would definitely try it out if you are at least if you are on the $100 plan. Could save you save you some tokens, I think. Yeah, I think you definitely go check it out if you're on the pro plan. Uh, I don't think it's applied because you because you um you don't have opus anyway, I guess. So, look at this. Just read me. Okay, nice. So, it looks

like it's working. Good. Is chat GPT uh Chat GP5 the best coder for a price? I actually do think so. Has to be, right? I do think so. because uh the price is kind of unbeatable. I guess you could argue Gemini 2.5, but uh I think it's probably going to be it's probably going to be GP5 on the API, right? for the price, but it kind of if you have clothes code and you kind of have

the the plan, you also get a lot of value, right? So, it's kind of hard to say Now you see why I asked about using plan mode over sub agents first vice versa. Yeah, setting my sub agents on plan mode due to them yielding better results could be something I want to try out though. So if you didn't see my video on this, so basically this is uh data from poly market. We are grabbing this via

the API and we're feeding all of this polyarket data into a set model. So, in this case, it's GPD5 mini high. And with all of this information, it's going to try to predict a price, like an implied price based on this data. And it's going to plot it here on kind of the map or the graph. And we also fetch the real Bitcoin price. This is kind of the green dots. And then we can kind of

see what model is coming closest to the real price uh from this data here. So I had a lot of fun with this because uh I also have a setup with open router. So I can try different models just to see. So here now I can switch to I can switch to nano mini nano high for example. I can just go to here and I can do now we are on GT5 mini. I can do nano

high or medium reasoning. Uh this is pretty close though. It's it's $400 above the real price. So, if I restart this now, it's going to be it's going to be with GP5 nano. Has someone tried Figma make? I have not. Is that uh some UI? Is that some UI AI thing? So now we are starting it with nano high right. So now we can run it with uh if we open up like a new can close

this if we do cd poly app and we run this. So here we are on the same code but we only have like a different AI. So I think we are on uh Kim K2 from Moonshot AI running through open router. So this yesterday I kind of compared like I had a look at nano high and compared it to like Kim K2 and I just saw what kind of model were closest to the real PL price.

So it's almost like a benchmark right? So, Kim K2 is thinking the price is going to be like 120,700 based on this inputs. So, that's $500 off. Uh, Nano High is going to guess uh 120,800. So, that's $600 off. So, Kimmy K2 is like $100 closer. And I can just let this run because it doesn't really cost anything. The Poly Market API is free. The Gecko API is free. So I can just gather all this and

just let it run and kind of see where the the data kind of looks after a while. So I'm just going to let it run in the background now to see what happens. So what I'm going to do next is I'm going to add some polygon here. And I'm going to use the API to give the models some chance to make the make a bet. So I'm going to compare like Kimik K2 versus GPD5 Nano and

they can if they think it's going to be a good place to place a bet, they can do it. And then I'm going to give like maybe each model like $50 and let them run for an hour and we're going to see what model makes the most money right? Uh yeah. Okay. So that's kind of Figma. I haven't really ever used Figma. So that's been something I never tried. So yeah, my idea is to create like

a benchmarks where these models can try to see what kind of models that can make month make the most or lose the less least money on poly market. So that is going to be some projects I'm probably going to start this weekend and maybe next week we can try to run the we can try to run them on the live stream maybe and see what models is going to make the most money. And on poly market

we can on open router we can try a bunch of different models. So I tried Kim and K2 yesterday and that was a pretty good model at doing an implicit price here on the bitcoin. So let's try to switch switch uh open router model. Uh let's try there was one model I tried that was really bad. I can't remember what model it was. Open router. We can try the new uh open open AI model. So let's

try the 120B OSS, right? So let's do open AI GPT OSS 120B. So I can switch here. And when I start this now, we're going to be on um the open source model from OpenAI. And we can kind of compare it to Nano. I'm a big fan of the nano model. It's ch so cheap and it has uh reasoning. What is the price of nano? It's just crazy cheap, right? Isn't it something like GPT5 nano zero?

So is that five cents in and four? No, 40 cent. This is this is 40 cents, right? 5 cents in, 40 cents out. So basically free. Did you try GLM 4.5? I think I did. That was pretty bad. We can try it now. So, this is this is going to be my benchmark going forward. So, I'm going to try to fine-tune it so I can do like benchmark tests on new model how good it's going to

be at uh picking the implicit price on Bitcoin. And we can add more data points. We can add like uh Fed rates changes. We can add Trump news and a lot of data points here because it's kind of real time. So you can see uh GPT5 GPD OSS 12B 121,000. It's a bit higher. It's like 700 off. This is more like 400 off the real price. Okay, that was really close. So, let's see here now. Can

your bot make my TI and enjo bag pump? We'll see. We'll see. Shik, but I just wanted to come up with my own benchmark. That could be pretty fun. And I think I'm going to go for this. So, I'm just going to find a way to give it money and uh give it the option to trade because we have the Poly Market API dev docs. Uh there's something called I think it's called just pretty simple Is

it where's the documentation uh for developers? Okay, there's something called just place order. Is it uh place single order? detailed instructions for creating, placing, and managing orders on Poly Market using this API. So, this is what I'm going to set up. So, I'm going to give it the option to to buy here. It's going to be pretty fun, I think. So, Nano is pretty stable in its prediction. doesn't really change that much while the open source

model is a bit all over the place but there hasn't been a lot of movements here now. Uh so yeah but this is almost yeah because the price is 120 now. Okay. So yeah that was what I was doing uh the day I didn't stream. had a lot of fun with this. So, I think I'm going to do 15 uh more minutes. Something put someone put up like here. What is this? Is this your project? CC

proxy dolphin. Who is running this project? Tell me more. It's a flashbased proxy that unlocks AI freedom. Use any cloud compatible application with XIRK me. Okay, Nathan. Thanks. modify cloth system bronze named in honor of Eric Hartford and Sensor Dolphin model series. Yeah, those models were cool. I haven't used them in a while, but I remember when they came out. I think it was was it like a llama version of Dolphin or something and it was

really nice. It was like an 8B model, like 7B model. Uh and censored that was really cool. Do you remember what the core model was, Nathan, on that dolphin? Was it the bloke? I remember him, Lucas. He had a lot of these GIFs, right? It actually intercepts and rewrites the system prompt. Cool. We might check it out. Let me see here. Okay. So, we have Yeah. Play straight forward. fun. I just came to think of something.

I just remember I don't know. I guess most of you um yeah I don't even think those models are here anymore but uh uh I guess the models have gone now unfortunately but when I started uh using large language models like so you can override the default cloth called prompts and make it more aggressive security-wise. Huh? Do you How do you do that? Let me see. Are you using Grock? An entropicate custom PR so you can

do your own original cloth. Here's the Grock pass through. Okay, Grock with custom Gro is not fast and good in practice though. Maybe we can try it someday with Grock. I have a Grock AP here. But I just wanted to do something funny. So I started using when I started using large language model that was probably 2020 GPT3 beta so before the instruct models. Uh let's see uh Excel uh what's your email for promotion on an

AI tool uh dedicated videos. Uh I can send you the email just send it doesn't really matter. You can send to to this one if you want to. This is my agentic email. Hello, Elgar. What did I miss? We can ask uh Claude, but basically you kind of missed RX giving away 100 memberships, Edgar. That was pretty crazy. So, he's a big hero for the community. So yeah. Yeah, just send me an email. Yeah, he's a

legend. So, we can ask Claude what what did Edgar miss and we are testing the polygon uh bots. Yeah, I guess I guess you didn't, but uh it was just crazy. So, he's like a community legend already. Uh, and we've been kind of running the Poly Market. Um, my benchmark on Poly Market. Is RX in the chat? Uh, I hope is. He's been for a while. I think he doesn't chat much, but uh, he was here.

So, what did Edgar miss? So, he missed that RX gifted 50 membership. I guess he did 100. Oh, it's here. Uh so clo opus plan with testing of the 1 million. We didn't really do that. We discussed uh okay so uh yeah it was only 50 messages. You didn't miss anything big but uh maybe the 100 membership was pretty cool. Other than that we've been just playing around. I was just going to tell a story when

I started with GP with uh large language models. the way you had to prompt it before the instruct models. So you had to do something like uh so the way I prompted it like in the GPD3 beta it was like what is the main capital of uh France and I did so I often did like Q and I did a Paris And then I did Q, what is the main capital of uh Norway? And I did

um a Oslo. So every single prompt had to be like uh multi-shot or like few shot. And then you could do something like V is the main capital of uh uh yeah I don't know Spain. It was really slow. So every single prompt had to be when I started using large language model had to be in this format. If you didn't you just got if I just said what is the main capital of France the response

I got was like what is the main capital of Spain? I didn't get the answer. So this was before the models were trained to do instruction following. So they were only they only did like pattern recognition. So if I didn't do it like this, it wouldn't answer this question. So it was really hard to work with. But they changed that with the instruct model from OpenAI. But that was back in like 2021 or something. So it

was just interesting back then. there was no information or anything about large language models online. So we come far since then. And maybe some of the open source models were kind of like this too in the beginning. Explaining this brings me back so many emotions. Were you early sound waves on the large wrangles models? You remember this, right? It was crazy. So, I did some work on some uh yeah, some online businesses. It was so hard

to work with. You had to like set up this crazy few shot prompts and the context window was like 2K. You were on GPT2.5. I guess you remember this pre-instruct prompting then. Unreal. It wasn't that practical to work with. It was messy. Still, OpenAI didn't fix it. I guess they fix it now. That's cool. Unreal. Were you Were you guys on the like the OpenAI forum? That was when the models were just autocomplete. Yeah, JK. They

were kind of autocomplete and like pattern recognition. They didn't really they didn't really they didn't follow instructions. They just did whatever they they did, right? Yeah. I remember that was pretty fun. But we come a lot of ways since then. autocomplete. Yeah, kind of autocomplete. But you're going to have to do it like this. So if I did if I did what is the cap of Spain and I just said like this. So the answer would

be something like what is the capital of France? That was my answer. So I just got this. Yeah, it was pretty fun. Okay. So, let's see our plotting now. Gift 5 Nano is uh $300 off. So, the price has gone a bit up here, like a couple hundred. It's kind of trending in the same way, though. That's pretty cool. Or is it maybe this two? Do you guys agree? Do you Is this Are they both trending

up? So, I guess it's using the data somehow correctly. You can see this one is going a lot up. This greater than 120. So, it's probably using that to trend up. But you could see the potential. Yeah. And I remember the OpenAI codeex model when this came out. I don't know. Is this Oh no. There's two things called codeex. It was called codeex at list sent you an email. Okay, I'll check it out after a stream.

Excel. Yeah, I agree. Ludovich, it's not hard to predict BTC as rising. Yeah. And the tools here are like really they kind of feed the information. You can see all the positive signs here on the reach, right? Um but it's a fun um it's a fun benchmark. I think I'm going to have some fun with this. That's my plan. And I can add other points. So I can do we could add like something I don't know

what kind of point is funny. What is like a correlated what is correlating with kind of crypto prices? This is maybe too slow. Fed decisions. I guess maybe geopolitics like Trump could be related. I'm not quite sure. I will have to look into it. And I could have like we could have also had like uh x.com information feeding in here from some handles on uh X or Twitter as data points. We can just use the the

Grock MCP server or something. So Reddit sentiment, I don't know. I'm going to try something. So that is the idea. So this is pretty dead. This source is dead. Okay. Ah, thanks L witch. I appreciate it. Yeah, I think it's going to be pretty fun. at least when we when we add the when we add the opportunity to to make orders so we can see what uh model is going to make the most money. So I'm

going to give access to like an account and I'm going to give them like a limit. So, let's say they can spend like $20 both each and let's see if they can make money or if they can lose money. But I need to look at over the weekend how I going to do that. I found this documentation here a bit confusing, but uh I think I will figure it out. Yeah, you're just going to do So,

you're going to do price size buy. So, we need a token. What about sell? I hope you can sell too. What would happen if claude code saw the output instead of What would happen if cla code saw the output instead of cloud code in sort of a display only mode? Cloud code will become its own feedback in real time and could enhance the display visuals based on the root prompt. Try it out. Soundwaves could you made

a notification that sends SMS for this project? Yeah, you can do that. You can just use Twilio, right? Maybe you need some kind of hook. I'm not sure. But if you use Twilio, you can make like a local MCP server that sends text messages to your phone. So you can place an order. Okay. So you can I guess cancel order is the same. Is that a same as sell? I have to look at look at this

over the weekend. I haven't studied it too much. Yeah. Yeah, if you I think if you go to like Twilio API docs, I've used it before and you go to developers or you can just search for SMS. I can't remember it is but it should be it's not that hard to set up. You just need an account. Uh I have made it before on cloud code hooks to send me text messages when the cloud code stopped.

But one thing I want to do has anyone I know it's possible right I think some of you have done that you do like a let's say we did a local host project like a local website and you have like an MCP include code uh it could be like uh I like something is it called browser use browser use, right? This is a really good uh really good um computer agent play right for example. On Mac

you can send notifications via o say osa script. Cool. So I want to do like a a browser use MCP and when the web page is done uh I want to send a request with MCP go to localhost 3000 do a full check of all the buttons on the web page and bring back our report. I know a lot of you have probably done that but I haven't really done it yet. So maybe over this weekend

I might do a video on this. I just want to see how it works in practice. So we're just going to do a local host uh web page and trigger the MCP server and check it. Have you tried that, Edgar? or someone should work, right? Like test that button, scroll up and down, fill out the form, see if everything works. Okay, Edgar, that's pretty cool. Are you doing that now? Logo. So, how is it going so

far? Does it work? Yeah, I'll let you know. Edgar, what is this? Strap the 1 million context rocket to claw cold max plants. Yeah, do it. And this was a good one. This wasn't this what we talked about during the stream. So we want this part added into the max plans. Do it entropic. I mean the report is solid but it got get lost in the pages. Okay. Logo. Did you use browser use or did you

use playright? I really recommend this one. This is really good. This is what um Manus uses browser use. It's really good. But yeah, let's do this. Strap the context rocket to the max code max plants. I'll have both this I double dare you to say you're absolutely right one more time. Every single time if you correct claw code, it's just going to answer you're absolutely right. Anyone using Kim K2? Uh we are running um we did

try to run Kimik 2 here but I haven't tried it like on um on like a real project yet. Yeah, it always says you're absolutely right. Oh, this is pretty good, too. Yeah. Yeah, there are some good memes here, I guess. Even when you're wrong, I know. So, it's a bit sick of fancy, right? This was a good one. But we need to we need to get it into the max plants, I think. Yeah, G5 got

rid of that. You're absolutely right behavior. Okay. What is this? Sonet 4. I think you set model to Sonet for 1 million. Try. Okay. I guess we figure out more um in a few days how we're going to use the 1 million. Yeah. Shonic. Uh I have to go too. I got to be somewhere in 20 minutes. So I think I'm going to call it. So, for the stream, RX, if you're still here, thanks a lot.

That was really nice uh for the community uh of giving all those memberships. That's going to give a lot of people like uh I'm going to do more on the membership side on YouTube. Anyway, uh that was really really nice of you. So, I really appreciate it if you're still here. Uh if not, so thank you everyone for tuning in again. We didn't do the most so much this stream, but uh I think we went over

like a few things that was new with cloth code and some interesting conversations. So yeah, Lucas, Jamie, John, say thanks to RX. That was really nice. And yeah, I think we are kind of getting like a good community here. People are working on some cool stuff. So I think we need to do one stream where we can kind of look at some projects. So, Edgaria, I have you on LinkedIn, so let's this weekend maybe I'll send

you a link to the project. Pedro, Ludovish, Jake, thanks for tuning in. Tom, thanks RX. Yeah, pretty cool. So, I wish everyone a good day and um yeah, I'll be probably back tomorrow. I don't have any big plans, but I might be back tomorrow. Uh, I might do like a mountain hike. If I'm doing that, I don't have time. But if I'm not going up the mountain, I'll be back tomorrow. Nathan, thanks for the engagement. And

yeah, thanks Soundwaves. Nice to meet you and talk about the old days. So, have a nice day and I'll see you uh all again hopefully soon.