1. Podcasts
  2. The Neuron: AI Explained
  3. BONUS: GPT-6 Sol vs Claude Opus 5.5 LIVE: Which AI Model Is Better?
The Neuron: AI Explained

2 Oct 2026 · 57m · 20 chapters

Episode artwork

BONUS: GPT-6 Sol vs Claude Opus 5.5 LIVE: Which AI Model Is Better?

In short

Live “model release day” comparison of Claude Opus 5.5 vs OpenAI GPT-6 Sol and GPT-6 Luna, focusing on performance, pricing, and practical workflow settings (quality tiers like high/extra high/max). They also run agentic coding tests, including “Doom end-to-end, but with cats,” and discuss caching/cost dynamics and planning strategies.

Guests

No named guests appear in the transcript. Hosts/speakers include Mark, Grant, and Corey (plus chat participants).

Key claims

  • Biggest differentiator is frontier pricing: GPT-6 Sol/Luna input/output costs drop sharply (e.g., Sol output $20M→$10M; Luna input/output also reduced).
  • Claude Opus 5.5 is cheaper than Opus 5 and performs near Claude “Fable 5.1” on most work; Opus 5.5 pricing cited as ~$4 input and ~$20M output.
  • Claude “cash reads”/caching is now clearer and improved (tiered discount vs confusing prior behavior).
  • Recommended tiering: Sol on extra high for accuracy; Luna on high for best speed/cost; Opus 5.5 for planning-heavy workflows.

Notable examples

  • Benchmarks: “Automation Bench” and agent workflow tests; “CatDoom”/“Catacomb” game demo generated by Astra vs Sol vs Opus.
  • Safety note from Claude model card: automated behavioral audit claims less likely to take hard-to-reverse actions and more resistant to prompt injection.

Written by AI. May contain mistakes.

Chapters

0:46

Excitement for Model Release Day

Hosts express their enthusiasm for model release day, discussing recent AI updates.

“They've been getting progressively better from both companies as we go.”
2:55

Pricing Dynamics in AI Models

A discussion on the changing dynamics of pricing among AI models and companies.

“I think we're really looking at, Grant and I were talking about it a little while ago, and I was saying that the dynamics really changed because it's not just them anymore.”
4:06

DeepSeek 4.1 and Usage Patterns

Hosts share their experiences with DeepSeek 4.1 and comparing usage costs.

5:45

Evaluating GPT-6 and Claude Opus 5.5 Features

Analysis of the features and pricing of GPT-6 Soul and Claude Opus 5.5.

“Yeah, this is the real kicker with the OpenAI news, in my opinion, is that GPT-6 Soul and GPT-Luna Soul are at Input tokens are exactly half price over 5.6.”
7:55

Understanding Caching in AI Models

The hosts explain caching in AI models and its impact on usage costs.

“and it performs at the level of Claude Fable 5.1 on, caveat, most work.”
10:01

Real-World Experiences with AI Models

Hosts share personal experiences and opinions on using different AI models.

“And, you know, you can see there it's 20 cents per million tokens, 60 % less than Opus 5.”
12:21

Analyzing Benchmark Performance

Discussion on benchmarks for various AI models and their performance metrics.

“So if you were running Automation Bench, which is a test of business workflows across apps, X-High is the best option there.”
14:00

Evaluating AI Models: GPT-6 Sol vs Claude Opus 5.5

Discussion on the differences in performance and usage of GPT-6 Sol and Claude Opus 5.5.

15:17

Choosing the Right Model for Development

Advice on selecting the best AI model based on the type of application being developed.

“Initial take on 5-5 seems less verbose, a bit faster, maybe because it's not throwing out extra words.”
17:46

Cost and Performance Considerations

Comparing costs and performance metrics for coding tasks between various AI models.

“make sure that you are fact checking yourself there.”
Show all 20 chapters
20:07

Optimizing Workflow with AI Models

Strategies for utilizing different AI models effectively in workflow and project management.

“But it's not doing the hours and hours of work in between so much.”
23:19

AI Safety and Model Reliability

Exploring the concepts of safety and reliability in AI models and their implications.

“I'm going to go long in on my other computer Yeah let's see here what am I bringing up?”
28:00

Navigating AI Models for Game Development

Learn about the nuances of different AI models for game development.

“Open street map, mostly some basic typical nav functions.”
30:58

Testing AI with a Unique Prompt

Explore how to utilize AI models to create a game variant with a creative twist.

“And the way that we did this last time, Corey, correct me if I'm wrong, was make the game Doom end-to-end.”
33:41

Exploring Coding Tools and Platforms

Discover various coding tools and their applications in software development.

“I've heard good things, but I've not used it yet.”
35:39

Advanced Coding Strategies and AI Assistance

Gain insights on managing multiple coding projects and leveraging AI for efficiency.

“it's kind of a no-brainer depending on the amount you need what do you think of sui2 for coding um i will be honest i have not used that um but i know people like it.”
42:02

Exploring AI Instruction Following

Discussion on instruction following in AI and testing various models.

“But I'm thinking out loud, I don't know since they just released Opus 5.5 I don't know if there's a Luna Sure, why not?”
46:50

Live Game Development Insights

Live demonstration and discussion on developing a game using AI models.

“It's every time my computer dies because I have too many agent processes running, then it restarts.”
52:58

Comparing AI Models: GPT-6 vs. Claude Opus

Comparative analysis of GPT-6 Sol and Claude Opus 5.5, discussing features and capabilities.

“I'll come join again and enjoy your toys.”
56:00

Wrap-Up and Next Steps

The hosts summarize the session and discuss future plans for testing AI models.

“1 p.m eastern and we will be back and trying it out so get your prompts ready get your get your hot takes.”
A man walking through an airport, asking his phone a question

Ask about this episode

Get the gist, explore an idea or ask about something said.

Claude

On The Neuron: AI Explained, what are the key points of this episode?

Searching your listening

Transcript

Automatic transcript. May contain errors.

0:00To the Neuron Live. We'll give everybody a minute to trickle in here while we're getting started, getting set up. Funny enough, we were doing an interview 30 seconds ago and had slam, done, scoot, open. So it's very...

0:15Corey Noles: Model release day, you know. it's it's model release day and that that gets gets crazy and uh it's also the funnest day in ai if we're going to be honest everybody loves new model i guess not everybody loves new model day but i sure love new model day the ai safety people do not love new model day they do not love new model day but you know sometimes they're just no fun okay to open the stream that might have been

0:51oh goodness yeah uh but yeah really excited about today i see an update button in my codex that i'm gonna let do the thing that it does and uh hope that in a moment my work codex has the new ones but yeah yeah crazy new model day we've got uh claude dropped opus 5.5 and and blasted through

1:15Corey Noles: some serious benchmarks it looks like i think the new benchmark to compare these model releases is who has the cooler landing page because they're really they're really going at it with these They're really going at it. Yeah. They've been getting progressively better from both companies as we go. Yeah. I guess it's like, well, what's the excuse? You got to use your own supply. Nobody's going to like your model. Nobody's going to like you if you don't build the best new model landing page ever. Mark, I totally agree with you. Mark is already calling us out on the AI safety people. He said, hey, we can be fun to respect.

1:55Corey Noles: and honestly I think safety is really important same I do too and I'm just having fun it's a nice new day I mean no offense hey if you're just joining us looks like we got a flood of people shout out in the chat where you're tuning in from Corey's in St. Louis, I'm in Los Angeles let us know where you're from yeah it's always exciting to see the wild places that people pop in from uh well um yeah for everybody just flooding in welcome we're glad you're here we're excited to talk about uh some of the new models that dropped today we've had uh a few cool drops we've got opus 5.5 we're checking out here and it looks like open ai came out with uh gpt6 soul and gpt6 Luna as well, which is also really cool.

2:51A lot of, and honestly, the biggest takeaway from all three of these to me is pricing at the frontier. I think we're really looking at, Grant and I were talking about it a little while ago, and I was saying that the dynamics really changed because it's not just them anymore. Like we've got Grok coming out yesterday at$6 per million output. it we've gotten meta who's nipping at their heels and and at this rate probably has a new model coming in the next month and uh it getting increasingly competitive at four dollars or free if you're on open code which is crazy and uh i think that that kind of pricing makes it worth looking at some of those other companies so you've got to stay competitive like the truth is the open ai models have been good to me the cloud models have been good but the more i'm spending the the better that value proposition needs to be for me to keep my wallet open

3:55Corey Noles: yeah well especially because you know you it's sort of a reinforcing feedback loop where the more you use the more you want to use them and then but then the flip side of that is now there's a constraint which is the cost and the more that you want to use the more the cost really matters and I don't know if you mentioned that as forgive me if you did I was looking at the chat which by the way shout out Montana Boston LA Brevard North Carolina I'd like North Carolina I've never been though Atlanta Portland DC Germany and Ireland we got a lot of people in the chat Toronto wow but but my point with that is um one thing you didn't mention or at least I didn't hear you mentioned is deep seek v 4.1 and that's another one that is front another level and very cheap so yeah yeah i spent i spent my weekend because i was at my codex limits as happens from time to time i've been burning like my token use has gone this year from or from late last year of a few thousand tokens a month i mean a week to i'm burning closer to three billion a week now which is a lot now i'd like to add i'm doing that on a subscription plan for the most part and that's very reasonable like the truth is what i'm getting for my dollars even with the rate limits is a lot and uh but doesn't mean it's enough so i spent the weekend figuring out like open code and running MuseSpark for free and using DeepSeek v4.1 Pro?

5:32I can't remember. 4.1 something. Yeah, and for close enough to free that it's really cool that we have these options, but let's take a look at what we got. Yeah, this is the real kicker with the OpenAI news, in my opinion, is that GPT-6 Soul and GPT-Luna Soul are at Input tokens are exactly half price over 5.6. So you can use GPT-6 Luna for 50 % of the input cost of 5.6 Luna today. So there's no reason to go back. Output tokens for Sol fell from$20 per million output to$10 per million output, which is a huge difference. and Luna went from 20 million, 20 cents per million input, a dollar 20 per million output to 10 cents input, 50 cents output, which is bonkers.

6:32Corey Noles: And one other thing we didn't mention is this whole Jev phenomenon, which if you're a developer, you know all about Jev, but if you're a regular person, you're like, what is Jev? Well, Jev is named after Jevon's paradox, which is the cheaper something is, the more people use it, which is kind of what we were getting at earlier. and uh that's jev competitive pricing right there i don't even i don't even know if that's you know comparable to dbc uh deep seek 4.1 but yeah go ahead and in stuff that is broadly usable like the thing with jev is like you have to have a use case uh like again probably everyone has a use case or two but uh like i use luna so much most of my scheduled automations most of my agents run on Luna because it's enough.

7:19The truth is it's enough. I don't have problems with it. If I was getting poor results, I tick up. Uh, but I'm able to do enough just ticking the reasoning up and down that, uh, you know, there's, there's a lot of stuff I do with it. That's pretty heavy and runs every day just to make sure I get everything I can out of my, that I have bled my single$100 bill as dry as possible.

7:44Corey Noles: Yeah, so that's GPT's pricing. I think the surprise was this one here, which is that Claude is also a lot cheaper now. So we've got Opus 5.5, and it performs at the level of Claude Fable 5.1 on, caveat, most work. It's not all work. And costs 40 % less to run than Opus 5. And I was actually doing quite a lot of Opus in a$200 a month subscription. So this is very interesting to me. It is. You know, and I won't even give them a hard time on most work. Like, the fact is, every task is so different now. Every run you do, it's really hard for me to say, oh, well, I built this website, and then I built that website.

8:32Like, the fact is, there's so much difference between them now. I have yet to get hands-on with this, but it's in a price I'm willing to pay, which is the first time I can say for a Frontier Anthropic model in a minute.

8:47Corey Noles: Look at that. It is$4 per input and$20 per million output. So basically it was competing with GBD 5.6 on price, and then GBD 6 was like, slice it in half. so they're definitely competing on price like this is a price work they're at least staying in the range of difference they've tried to stay is what i would say you know what i mean like there's always been this you know you can't be more than double more like because double more it makes zero sense i don't care how good it is it's uh but fixing their cash reads is a big thing too claude's cash reading has been just a nightmarishly confusing mess of of of scale for since it was really explain that just for people who don't know what cash reads are just like a very brief like one line thing yeah yeah sure so when when you send an agentic prompt to a model it's got all of this history that is data that you have already sent the model those tokens that it has already seen get charged at like one percent of the rest of your load and that's why you can run three billion tokens on a hundred dollars once in a while and uh it happens it is most of that is caching because like it's grabbing the full context in your prompt of that agent every time and instead of charging you full token rate they're not claude did this but it was like a scaled tiered discount thing where everyone else's version was like well these are 25 cents per million cashed uh and it was very basic it was very trackable and understandable where claude's was more complex.

10:43And now it's way better. Yeah.

10:46Corey Noles: And, you know, you can see there it's 20 cents per million tokens, 60 % less than Opus 5. So, yeah. Should be good. But Will Herons in the chat says, we can say it doesn't read cache. So that's interesting. Please tell us more about that. Have you tested it? Also, if you have tested these models, we want to hear from you. We also want to hear if you have something specific you want us to test. We have our six prompts that we tested GPT-6 on. So Corey, I imagine we'll run through at least a couple of those. Yeah, I think so. I think so. Most of them take a while, so it might be tricky to get through too many of them.

11:26Corey Noles: Do like two or three or something like that. But there's no reason we couldn't do two or three. Grant, which ones do you think? Definitely Cat Doom. We should definitely do Cat Doom. Definitely Cat Doom. Do that in Opus 5. Opus 5.5. I want to see Opus' version of CatDum. I will. I'll pull that up in a minute. Let's see here. Well, while you're researching that, I want to read Cabadrin's comment. So Cabadrin says, I've been using Opus 5.5 today. It spins significantly faster in code reviews and responses. So that's interesting. Nice. Nice. That's good to hear. Yeah. I am real quick taking a look here.

12:07Corey Noles: I'm going to look just very briefly over some of the highlights here. Yeah, that'd be great. Automation bench. So we've got Astra here, which is the highest on that. And then Sol. Looks like Sol on extra high is probably the best performance you're going to get there per cost, which is good to know. Then we've got don't put it on max. It doesn't perform as well. So that's helpful to know. On what? No, on Macs. So if you were running Automation Bench, which is a test of business workflows across apps, X-High is the best option there. It's funny how that happens with some models. Was it Fable or Fable 5.1 that had this one point where it was like, oh, well, if you cross that threshold, it wasn't a lot.

12:57I think it was GPT-6.

12:59Corey Noles: My recommendation is I've just been using everything on High because it seems to be the best middle ground between quality and not costing too much. But yeah, as you can see here, obviously Astra on max is going to be the best value you're going to get, or I don't know about value, but best performance. And then Luna on max outperforms. So you don't want to do Luna on extra high when you're trying to automate things. You do Luna at max, which is very interesting. Interesting. and this kind of breaks it down here by the score that it got in a little bit grant yes what do i am uh fishing for

13:46Corey Noles: our our project prompts here yeah so this is the model and effort and the best score here and the cost and then they benchmark it against cloud opus 5 max and fable 5.1 which is comparable but this is not the new opus 5.5 that just came out so we'd have to look at that agents last exam which evaluates agents on complex professional workflows gpt sole max is the best 54 6.4 you know i never run them on max i just i have a i used to only do it on max now i rarely do it on max but this these models i think you're going to want to run on max it seems you get significantly better increases there assuming the benchmarks hold true yeah this all comes with a caveat and if you've watched any of our previous episodes you know we just don't care about this like you got to just test it and your workflow like this is marketing and they're trying to tell you like hey you should give this one a shot look at how good it is yeah but it's you know you got to just try it on what you're trying to do ultimately i generally with whichever model i'm using i live at medium or high depending on what i'm working on and then kind of bump up or down just depending on what I see.

15:17Ryan's take. Initial take on 5-5 seems less verbose, a bit faster, maybe because it's not throwing out extra words.

15:26Corey Noles: Yeah, oh my gosh. It's not saying load-bearing every five seconds. You're absolutely correct that I'm saying load-bearing. The word load bearing is a load bearing Word in my sentence structure

15:42Grant for the life of me I am I am not finding

15:49Corey Noles: What the doc? I can send it to you The doc with the With the actual prompts If you want to If you want to take over sharing your screen And talking through some of this I can go find that Info for you real quick Yeah, I can do that. Hang on. Let me get one up. Yeah, we're looking at GPT-6 right now. If I'm working on an app, what should I have it on? Can you elaborate a bit? What kind of app are you building? If you don't mind saying. Because it matters, I would say. Broad strokes, yeah. Yeah, you don't have to give it all away. How complex are you looking? Is this a thing with accounts and logins?

16:33Is it a game? Is it a little work tool?

16:37Corey Noles: Is it an iOS app, a web app? So it looks like GPT-6 Sol, if you run it on extra high, is probably the most accurate per the cost that you're going to get of the GPT models. see that correct what on do you say max or high this is gvt6 sole on extra high okay so you've got answers with any factual errors it's the lowest at 4.5 percent at the cost of 13 cents per task so over here you can see astra has higher factual um higher factuality i don't know if that's word. It has a lower error rate, let's put it that way, but it's much more expensive. So this is the sweet spot right here for that. So lawyers, any lawyers in the chat, check your AI generated citations, check your fake case law, use GBT6 Sol on extra high and make sure that you are fact checking yourself there.

17:50Corey Noles: oh my gosh coding okay this is what the people want to know coding okay so they've got clod in here okay gpt6 soul versus astra so okay soul on max costs two dollars per 14 cents for coding tasks which is comparable to Astra on medium, but a little cheaper. So if you're going to run Astra on anything lower than high, you might as well do Sol Max. It's going to be a little cheaper for you there. Nice. Yeah, the flip side, someone just pointed out here, is that another factor is time. Extra high and max are slow. That's true. You know, if what you're working on requires that extra juice, cool. and it just depends on the task for me we got some more information here definitely deals with accounts and logins going to start as a web app eventually i'd like to get it on ios and android later on honestly i would probably use soul uh i what i do is i will let astra plan i'll be like I've got this thing I want to do.

19:08Here's how I want to do it. Build me out a phased build plan, plan the architecture, tell me what I want. And then I take that over to soul. I work there and sometimes I'll tell soul even, you know, use Luna sub agents wherever it's helpful and good enough and let it decide. But you know,

19:31Corey Noles: usually if you have it do a technical implementation plan, so it explains exactly how to implement it, not just say like, okay, hey, go build this feature in the planner mode, that's going to be better when you take it to the subagents. I always tell it, prioritize quality, but don't be wasteful. Yeah. You know, like if we don't need the steam engine to drive the nail into the wood, a hammer's fine. Oh my God, look at this. GPT-6 Luna on max is comparable to GPT-6 Sol on high. wow actually maybe better if i'm reading that correctly at five percent of the cost yeah dude that well and look how look how it compares look how that line compares even over to astra yeah like it's not that big a gap no you think about the money no it's not it's really not yeah wow even wait what is uh uh-huh i couldn't see go back up real quick and go over to the luna line what's the third tick down this is luna on high for eight cents that's where i'd live really absolutely where i would live i would live at luna on high i mean like if you look there you're looking at the difference in like 52 and 58 or something wow uh you know gonna notice that and you're gonna get a good amount of speed at medium uh or at high that's probably where i would live if it's me uh with the knowledge that i can tick it up anytime i want and and you know yeah put the pedal to the metal you gotta ask yourself when it comes to speed um this is my personal take is okay is it more um can you do it uh quicker if you either a use a less intelligent model that's really fast but it has a lot of mistakes and you have to do a lot of rework or b use a slightly more um you know more expensive but more intelligent model that makes less mistakes and you can just roll with it um that's kind of the trade-off and i think a lot of times people say like hey you know intelligence is is worth it because it's worth not having dumb mistakes in the code and you know they'll take that trade-off see like my approach to it is i have astra plan and do my reviews yeah so like it does the planning the worker bees go do their work and then astra goes and does a code review and a security review and fixes some little things here and there.

22:18But it's not doing the hours and hours of work in between so much. And, you know, you can... It's still quick. Yeah.

22:28Corey Noles: You can also speed up by paralyzing, right? Like, yeah, sure, you can use Luna because it's faster. You can also do three different parallel chats of Sol, you know what I mean, and have to be working on three features at the same time. So there's different ways to approach it. You kind of have to figure out what works the best for you. What works for you. Same with whether you're using this or whether you're using Opus. I mean, like, if the truth is, if you feel like Claude performs better for you, if you feel like GPC figures performs better for you, go for it. The truth is the margins aren't that big between them in terms of quality.

23:07but in the end the one that you feel like responds the best to the way you talk to it and the way you ask it questions uh if it's the one that makes you better it's it's the right one it's kind of the same idea with you know whether it's baseball bats or tennis shoes or whatever the case is if it's if it's what you feel like this i just feel like for me i do better with this uh cool it's the right one maybe uh give it a shot but don't be afraid to try something different either I spent my weekend trying different things And I was quite impressed

23:41Corey Noles: Yeah everything's pretty good at this point Hey do you want to take over sharing your screen? Sure Are you ready? I'm going to go long in on my other computer Yeah let's see here what am I bringing up? Whichever Whether you want the page or I'll share the links to both of these In the chat That is OpenAI's Page this is claude's page i'll also go ahead and share the model card for claude i've got claude's up now yeah we spent a good good bit of time on gp6 there's been a good bit of time there yeah that is a sexy page even with the dull type the dull type makes it even cooler like that uh it feels authentic i like it so what happens why isn't it playing it was doing neat things for you grant i'm very disappointed maybe it's just when you load it yeah there we go whoosh that is fun let me make sure i'm over here so when grant hits us with his other computer i can click him in okay performs the level of Claude 5.1 very cool see what they say about safety it's been a big topic lately achieves the best scores of any model to date on our automated behavioral audit cool alignment suite to test Claude across thousands of simulated scenarios much less likely than recent models to take hard to reverse actions that's good or act outside the boundaries.

25:19More resistant than open to Opus 5 to prompt injection. I don't know that any or that's one of those things that's kind of like safety in general. I don't know if some of these are problems we ever solve. I think they're problems we face. And I think AI safety is going to be just another thing like cybersecurity where there's always a frontier of that. and you're sort of always working with new capabilities, new problems, new scenarios, new surfaces, new means of attack. I know I just heard Grant. I could be crazy, though.

25:56Corey Noles: No, I'm here. I'm just pulling up some stuff. I'm going to comment. Recommend models to use for a navigation app. Are you thinking about doing that through, like, you know, like, I bet Google's got a Maps API of sorts or that there's some type of, there's something on GitHub, I'm certain. You could use either. The truth is, any of these, you're going to do good.

26:27If I was thinking navigation, I would probably lean towards one of the biggers, like either Opus 5.5 or Astra, probably. especially like Grant and I were just saying, we have different approaches in terms of how we do it. He'll work within the bigger one longer and I tend to just kind of plan and then review with the big model. But either approach you take, I think you definitely want the bigger model involved in the project. But I'd say Sol could do it. Opus Here could do it. You know, depending on your specific features, you know, just take patience. A lot of people say you don't have to have your AI plan, but I really, really disagree with that.

27:17I have gotten such better results when I go into a project and have an idea. Like, I want to do this. Let's put together a phased implementation plan. So we've got, like, I know the features we're going to have. I know the layout we want. I even go and a lot of times will use, I use chat GPT image to generate a mockup. Like, here's kind of how I want it to look. I want it to feel like this. I want this kind of buttons. Maybe I want these colors. And then I'll take that as part of my project plan that I feed in and say, hey, here's what I want. I want it to do these things. I want it to look like this.

27:55Go build a project plan and an engineering plan and bring it back and we'll get to work. Open street map, mostly some basic typical nav functions. But yeah, one complex routing logic. Honestly, that would probably be fine. That would probably be fine. Let's see here.

28:23Corey Noles: Corey, I think I've got the prompts for you. I'm going to email them to you real quick. Interesting that you all use Astra for plans. uh the truth is i don't really use anthropic anymore um it's been quite a few months i still i still i still check out the models i still test them out i'm still impressed with the work they do i just got to where i decided i would rather have a big subscription than two small ones became more practical for me uh so yeah i'll have astra plan and uh because astra astra is I think significantly stronger than Fable 5.1 probably not than Fable 2 when it comes but uh I feel like the Anthropic models do better with more holistic plans but GPT knows the code base better that's fair and and the truth is um you know do what works for you uh none of these are bad uh in my opinion uh you really can't go wrong Grant you're back uh did you hear me talking or know never okay were you talking i once said i think i hear grant's voice uh it's all good so i sent you the prompts um of which we will choose like two or three to test i've also updated both of my apps claude and uh codex so we can test them whenever we're ready excellent excellent all right let's see if mine is i've got them in my my work codex too you've used your entire limit.

30:02Okay. Let's see here. Yeah, I think that's fine. I think if you want to dive in, let's see what you got.

30:12Corey Noles: Okay, let's do it. That sounded so aggressive. No, that's fine. Alright, so what I'm going to do is I'm going to pull these up in a window here. Mark, this is so true for me too. What is it? Read it out loud? Planning is super useful. I found myself down some bad paths when you just start without having an outline or structure uh kelly roberts says i've also allowed it to talk me out of a build during the plan phase it made good points i thought about and so i pivoted i've absolutely changed my mind when it's like boy this thing you want to do this is this is a lot okay so for the first test why don't we do cat doom and And the way that we did this last time, Corey, correct me if I'm wrong, was make the game Doom end-to-end.

31:07But with cats. But with cats. That was the whole prompt.

31:10Corey Noles: All right, let's run it. Make the game Doom end-to-end, but with cats. Which do you have it on? So this is on Astra, so let me make sure that I switch it up. We're going to do Sol. Go to Sol. what did we say was the best x high for soul let's do that yeah that's probably plenty and then over here we got opus 5.5 let's also do x high extra is what they call it and we'll do make the game doom end to end but with cats

31:53Corey Noles: go all right we'll see what happens we'll see you in a half an hour yeah it might take a while uh if it does i'll steer both of them with the same prompt and i'll say hey make a demo version i can play asap yeah um oh i'm still that's funny did it just tell you you hit your limit no i'm approaching my weekly limit in fact actually i think there's a reset i might need to do that yeah i read earlier that they were doing a reset

32:26yeah good good call outs here from everyone what are people saying people usually use top model for planning and for execution a low-cost model for example people using astra for supervising

32:37Corey Noles: deep seek for one flash for the work yeah that's a great pattern uh you just have to you know you need a third party uh what's called a harness to do that uh cory's been using open code i've yeah I think quite highly of open code. I also really like Devin or Cognition. The one that they create is really, really good. And that's more for professional software. I haven't played with Devin yet. I keep seeing people talk about it. Another one's Pi. Yeah, Pi is like a very local, easy to configure one. Basically, if you wanted to do that, you should just go to the GitHub for Pi and ask Codex or Cloud Code to help you.

33:16Corey Noles: be like, hey, I want to make, I want to, you know, fork and customize this Pi harness. Can you help me set that up? And you would have it help you do it. I just dropped, I did this open code guide yesterday for basically how to use open code for free with the desktop version if you want. So I just threw the guide in there. This is what I did all weekend is I, I built two obnoxiously stupid, useless websites. The world could live without but i couldn't resist uh and uh i did them both with uh with muse spark 1.3 through open code for like zero dollars and and truth is i mean from zero to it's it's uploaded to my repo it's posted to a github page and it's it's live it's it's connected my google analytics account to it so I can watch traffic and see how they do.

34:11So, yeah, I need to check Pi out. I've heard good things, but I've not used it yet. I've also been seeing a lot of this Orca.

34:20Corey Noles: Someone else mentioned Orca. I've been seeing this a lot lately as well. Haven't tried it, though. Good call. How are we? What are we doing here? Well, it's both very slow, I think, because we've got extra high. You know, we joked that live demos aren't what they used to be. Yeah, yeah, Muse 1.3 is awesome. Like, it's really significantly better than what I was expecting when it came to building. I also appreciate the old school emoticon there. Yes. So thank you for that. Yes, thank you for that. That is, needed that today. Copilot for GitHub is definitely worth it, especially if you're getting, if you're new to coding.

35:03Corey Noles: I would say it's probably pretty easy to use. And we just did a GitHub for beginners live stream that walks through how to do that. So you're welcome to use GitHub Copilot to do that. Cassidy, who did the GitHub live stream with us, recommended it, obviously, because she works there. But she walks through how you would use that in a workflow like this. Thanks for providing the link. And it's a great one. Yeah, and the truth is, you know at its core you're kind of paying for co-pilot which means your subscription in there is getting you uh open ai models it's getting you anthropic models you know um it's kind of a no-brainer depending on the amount you need

35:52Corey Noles: what do you think of sui2 for coding um i will be honest i have not used that um but i know people like it. Sweet 2 is definitely up there. The one from Cognition. Yeah. Yeah. I just haven't tried it yet. I've been meaning to do like a full deep dive, like let me use Cognition for a week and see if I can replace both Cloud Code and Codex with it. Ryan Carson, who we had on the show, he's previously taught a million people how to code back before Coding Agents. He's pretty much all in on cognition at this point and he likes it and he recommends doing cloud environments so i did both of these tests uh in the cloud actually i might have done this one local which might impact it so let me oh one's cloud one's local let me just double check that

36:45yes orca is nice only if you're multi-work tree tasking otherwise you don't get much more than other desktop apps got it so if you're like like vincent from open claw and you've got 42 codex windows open on your desktop uh uh that makes sense was it he was talking about last night herder for multitasking yeah i was i was looking at that i've the most i've ever had open and running at once is like four and and it felt a lot like those are rookie numbers you gotta get those numbers up dude it felt like juggling flaming bowling pins uh i was having to make sure i'm not screwing things up over here now to be fair they weren't all in like one cohesive software project either that might be more manageable instead it was like I'm building a game I'm building an app I'm building a this I'm building a that and I found it pretty easy for me to get confusing and trip over myself yeah for real

37:57Corey Noles: that's fair yeah it not to you know just repeat everything Ryan Carson says but he also was saying that like in his experience because he's full-time coding for his own startup right so at the limit if you are really excited about agent coding tools and you want to maybe start your own business or create your own product with it, you might be doing something similar to him. So this is why he's a good test case for this. And he's since, I think, hired a couple of people. But he started off when we interviewed him back in March, he was doing everything solo. And he was saying that basically every 20 minutes he has to stop or he can only focus and make critical decisions approximately every 20 minutes.

38:40Corey Noles: So you want to try and set up your loops so that they don't come back to you with things all the time. And you can kind of just offload some of the decision-making and say, look, I've approved for this path. This is the plan. You're going to just go and execute this so that you're not constantly having to juggle those. Yeah. See, and what I would like is like I use, when I've got multi-up, I'll have voice mode commandeer them and be like, you know, go tell this one to do that, tell this one to do that. what is missing from that is i need voice mode to recognize when one is done and come to me and be like hey cory this one's done hit a snag what do you think look at this uh and i'll go yeah looks good i approve tell him to keep working and he'll be like hey go to work and then i move on and that's how i would like it to work sure yeah that makes sense preferably with my feet still up in the recliner yelling at a laptop three feet away.

39:42Corey Noles: Amazing. Speaking gently, too. I do a lot of stuff with a laptop that's in the floor beside me. That's where voice is so handy, is to just be able to go, you know, all right, hey, how's that one going? So you leave it on all the time, and you just shout commands at it when you have something for it to do? Yeah. Yeah, quite often. Whoa! Works really good. You can get, if you've got the pro plan, you get 15 hours of voice a day, Grant. That is extremely generous. It is extremely generous for how good voice mode is. And on a rolling, excuse me, on a on a rolling 24 hour basis. So like anything you use before 10 a.m.

40:21disappears at 10 a.m. tomorrow. And so like but I mean, you could absolutely leave it on for I could leave it on and do some days for my entire 10 hour work day.

40:33Corey Noles: and that's extremely nice because the amount of times that i'm opening tabs to start a new conversation to to do a new task is is probably ridiculous at this point yeah oh ridiculous yeah the number of let's just count one two i don't even know how many windows i have open three four this is a this is a great question i have eight eight chat gpt windows and codecs open right now uh yes i always joke about what i have running in codex but the truth is what's more hilarious is the number of cloud windows i have open but i'll go turn a thing on that i know i needed an hour and i just leave it away and move on all right so somebody in the chat christo flo ro excuse my pronunciation there said how long is this going to take lol good question we did use extra high so extra high is um you know takes an extra long time should we do a second prompts while these are running and do something like let's maybe do the cheapest best ones we can we can find would you pull up that like what's the best combination of cheapest best you know he's he's in the zone he's he's commenting to somebody okay one more time okay let's do a response you're good hey shut up on top fact i mean look we we both uh need that uh instruction following uh it was actually absolutely meant as an apology but it did not that its tone did not deliver that way at all i i i forgive you okay so let's test the cheapest best versions that we can find.

42:22Corey Noles: If we go back to those model cards. But I'm thinking out loud, I don't know since they just released Opus 5.5 I don't know if there's a Luna Sure, why not? Yeah, tell it, yes, live, Grant, live! By the way, I have a approve for me on which is like auto mode, but it's not accept all permissions so i do get prompted for some things um that's fair someone said hey new to gpt will astra return yes astra didn't go anywhere yeah it's not gone it's there it's still in there see look i can choose it here if i wanted to choose astra yep this is more like okay you can't afford to run astra all day right so you got it yeah for real i mean shout out in the chat who is better i i think at this point you have to say messy but i don't know maybe that's controversial maybe maybe spicy okay so let's check in with this the game now has a clear structure three connected maze levels a key and exit objective several enemy types and a final boss all right looking good looking good over here we've got let me set up the foundation first shared pixel helpers dev server asset preview tooling then parallel lines to art and audio while i build the engine okay okay hey grant in codex over there would you hit the little preview window in the top right corner that opens the sidebar so we can see what it's done if anything nothing to report yet um what's that tab at the top on new tab i zoomed in and made my screen so much worse yeah this is just like once it has something to report it will pull it cool sometimes it'll just blast stuff up there out of the blue and uh maybe turn on dark mode to prevent eye cancer

44:31hey

44:31Corey Noles: hey I will honor the request I apologize I feel that in my soul I have a couple things I use light mode on but it's rare where are those I just have not bothered to ever do that in here interesting Sirius says on my codex screen i only have light medium high extra high what's what's your plan

45:05Corey Noles: it might it might be a plan difference i just noticed someone said they on theirs only have soul 6 light medium high extra high without max ultra soul 6 yeah oh yeah you probably have to be on the pro plan to use those okay excellent excellent this this is possibly one of my favorite comments today and it it so sums up how ridiculous i feel when i'm annoyed that i'm waiting for ai you know i will say that these are kind of annoying like i don't know about you guys but the codex app for me on mac like half the time the buttons don't work like it doesn't minimize for me like now i can't get back to like the actual thing it's is it a crappy mac um no it's fine mac cool cool i have a crappy mac is why i was asking it's just causing me grief up here I have the how much Mac do I need to put an app on the iOS store and here's what I have to spend.

46:24Give me the best eBay solution to my problem.

46:29Corey Noles: I'm just stuck in an appearance trap here and I can't get out of it. Weird. I might have to restart. Nothing's going to take even longer. No! Alright, let me troubleshoot that. I'll bring my screen back up in a minute. Okay. Here in just a few, I have to slide out for another interview, funny enough. You need to restart it once in a while.

47:02Corey Noles: Hey, bro, bro, I do, okay? It's every time my computer dies because I have too many agent processes running, then it restarts. By agent processes, he means 42 Chrome windows and 176 tabs a month. Yeah, and it's all the agent running Chrome windows that it never closed. Chrome is my problem. Yeah. Yeah, it's real. Gosh. This is a bug in Trap2BT, I'm telling y 'all. I haven't seen that before. You should tell them. I always say that when I find a bug. Last night, I was using, this was pretty great, I was using Muse. metamuse and uh and i can't remember what it was now but it was like would you like me to tell the team you'd like to have that i can send a direct message to the team as feedback and i was like yeah yeah i would like that that's amazing all right we got it back up and up and rolling here

48:07Corey Noles: Go back to codex Maybe you'll get lucky And it'll be able to continue Yep Oh, looks like we have something to report Actually, we have a game I think this was the first one Yeah Okay

48:27Okay Okay.

48:30Corey Noles: All right. Let's see it. Catacomb. This is what the last one called it. Interesting. The purge. All right. Let's check it out. Oh, this is kind of sick. Oh, we're on and popping. Okay. Controls are as expected. Okay. We got some yarn over here. 14 yarn rounds. Oh. grant mine has secrets in the game that i've not found the one i did you gotta keep playing yeah i do well it's only three levels i still can't believe it made the boss chairman meow that's pretty good i was like ouch all right uh reach the portal press e got it Did it give you multiple weapons? So far, I mean, the projectiles are pretty good.

49:26Corey Noles: This is also the most sophisticated cat enemy that I've seen. We just keep getting yarn rounds. For 10 minutes and 30 seconds, though, Grant. Was that the actual count? Yeah, it's still like top three cat games we've had to date. Yeah, yeah. And we'll have to run the numbers on this when we publish the article, like how much that actually costs per task, right? Yes. We'll do that. Yeah, I'm anxious to see, do you have weapons? You do have health. You do have to collect keys. That's big. Keys are big. I got fish. Fish for vitality. That was pretty good. Seeing if it says anything around your HUD that looks interesting or out of the norm.

50:13Floor cleared.

50:14Corey Noles: It's simpler. I think sometimes they overcomplicate it and this is a little bit better I'll also say the walls look good yeah they do the overall there's like a red glow on it for some reason well you're in hell Grant defeat Mialuk that's pretty good Mialuk is pretty good Mialuk for you all don't know it's a reference to malik which is just like everything getting terrible eventually oversimplification but basically what's that is acquired find the get your fish so that you have vitality y 'all okay this is great yeah it's it's it's pretty good you know it really is said three levels and that's it so let's now compare that to opus over here i mean i mean here's what i'll say compared to hellcat which i dropped in the chat a little while ago it's 50 per million tokens for astra this is 10 per million tokens for soul and i would say this is 90 of the way there or more so this one is uh 20 minutes so roughly double the time and so far no game so far nothing to report no okay but who knows it might blow us out of the water it might i'm anxious to see but i have to slide away and uh i have actually i have a call in riverside now that i think about it that's problematic i'm gonna try something and hope it doesn't break your live stream well i'm gonna bring down my screen for a minute here and I'm going to set up those other two prompts before you jump out.

52:07Corey Noles: Are you jumping out right now? Yeah, I'm jumping out right now. So I hope that, you know, you all have fun. Continue building. I think the Opus will be better for sure. Let's see. You never know. Fables wasn't. Fables didn't beat Astras. It was good. They were both good in their unique ways. but uh is opus making cat theft auto we could we could we could test that do we want to i'm convinced we could vibe code gta 6 before it gets here i mean people have basically already done it i mean it's not going to be nearly as good but yeah they also save a hundred dollars by by using that one uh man yeah all right well i will be back thank you everyone so much for joining.

52:57If you're still alive when I'm done, I'll come join again and

53:04Corey Noles: enjoy your toys. So, you honestly think Astra is still better? Yeah, I think here's the thing. Are you logged in? Yes. I just realized I better do that before I hit leave. I wanted to make sure. Yeah, yeah, yeah. You're not going to kick me out. Yeah, Astra better. Yeah, I'd say Astra is definitely better and it should be. Yeah, I'll share some thoughts on how I use Astra with these tools over here. Let me go ahead and hide this. So we've got, here's what we've got over here. So we've got Claude Opus 5.5's plan. And as you can see here, it's got a plan. It's going to create the engine core. This is the world, the definitions, the game logic.

53:55Corey Noles: Then it's got the renderer plus the hud plus font and frame buffer then it's going to do screens inputs main loop and assets design five levels plus validator and then integrate art and audio play test build single file publish so what i'm thinking is happening here and this is some theories that i've heard and i think actually is pretty accurate at this point which is at a certain point fable is or sorry claude models are a little bit more general purpose and steerable than the gpt models are where the gpt models look really good out of the gate and they kind of optimize for demos and there's theories as to what's going on there i've read some people's takes on that um But I think that the GPT models are pretty good, but maybe perhaps you can go farther with the Fable models.

54:57This is a take that I've read that I think actually makes a lot of sense.

55:04Corey Noles: So Short and Simple says, I'm testing Opus 5.5 to the extreme and see on what its capabilities. Shout out. Tell us what you're building there. That'd be great. while this is going, let's get something else going in GBT over here. What do we want to see? What do we want to see? Are you going to kick me out? Yeah, or we're going to have guests show up in a minute. Well, if they want to be on the live. I just don't want to terrify someone with it out of the blue. Here's what we're going to do, y 'all. What we're going to do is we're going to come back on Thursday for the full two hours and we are going to share our tests so everything so our main six prompts we're going to share the results on both Opus 5.5 and GPT-6 Sol and then we're going to test your prompts live so tune back in this will be Thursday at 10 a.m pacific that's 1 p.m eastern and we will be back and trying it out so get your prompts ready get your get your hot takes.

56:12Corey Noles: Make sure you test the models yourself between now and then if you have access. We're going to do this all again on Thursday. We'll see you then. We'll get a bunch of cool stuff built ahead of time. That way we don't have to sit here and watch paint dry. But you will see the results of Opus 5.5's CatDew. I might run them through Meta too. We also have to do the AAA version where we do a AAA version of CatDew. We do need a AAA version. I'm going to run these through Meta too and see what it does. Just because. But on that note, everyone, thanks so much for joining us. We've got to run kind of urgently, and we'll see you back here Thursday.

56:49Have a great week. See you later.

From the publisher

GPT-6 Sol, GPT-6 Luna, AND Claude Opus 5.5 all just dropped. So which model should you actually use?


Join Grant Harvey and Corey Noles from The Neuron LIVE today as we put the newest models from OpenAI and Anthropic head-to-head.


OpenAI says GPT-6 Sol makes roughly half as many mistakes as GPT-5.6 Sol, bringing Astra-level reliability to a much cheaper model, while GPT-6 Luna is designed for fast, high-volume work like summarization, extraction, and quick questions. OpenAI is also dramatically cutting prices, with Sol reportedly coming in at $2 / $10 per million input / output tokens and Luna at just $0.10 / $0.50.


Meanwhile, Anthropic launched Claude Opus 5.5 just 90 minutes earlier. Anthropic says it performs around Claude Fable 5.1 level on most work while costing roughly 40% less to run than Opus 5, generating output more than 30% faster, and communicating more clearly. Its API price is $4 / $20 per million input / output tokens.


And the benchmark fight is already spicy: Anthropic reports Opus 5.5 scoring 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode, and 1,846 Elo on GDPval-AA v2.1, with particularly strong results in coding, computer use, and professional knowledge work.


On the livestream, we’ll break down:

⚔️ GPT-6 Sol vs. GPT-6 Luna

🤖 Claude Opus 5.5 vs. OpenAI’s newest models

💻 Coding, reasoning, writing, agents, and everyday work

💰 Pricing and what it means for AI users

🎯 Which model fits which kind of task

🧪 First reactions and live testing

💬 Your questions from the chat


If you’ve been wondering whether to use GPT-6 Sol, GPT-6 Luna, Claude Opus 5.5, or GPT-6 Astra, come hang out and test them with us.


🔴 LIVE: 3 PM CT / 1 PM PT


Subscribe to The Neuron newsletter: https://theneuron.ai


Models mentioned:


GPT-6 Sol + Luna: https://openai.com/index/introducing-gpt-6-sol-and-luna/Claude Opus 5.5: https://www.anthropic.com/claude-opus-5-5

More from The Neuron: AI Explained

All episodes