GPT-6 Astra came out on September 3, and naturally I started using it almost immediately on another project. This time I’m building a website in static HTML rather than WordPress, which is unusual for me because I’ve been building custom WordPress themes since long before anyone called this vibe coding. But at this point I’m making enough websites with AI that I don’t particularly want another recurring hosting bill every time I have an idea. GitHub Pages can host a static website for free, so all I really have to keep paying for is the domain name. Of course I want the .com. Part of the game for me is getting these websites to rank.
AI has also changed the calculation around static websites because maintaining them manually used to be annoying. If I want to change navigation, structured data, a reusable component, or something else across dozens of pages, I no longer necessarily have to open and edit all those files myself. A coding agent can do it. I still like WordPress and will keep using it where a CMS makes sense, especially for sites that need nontechnical editors or more complicated dynamic functionality, but some of the historical inconvenience of static development disappears when an agent can maintain the project for you.
The project gave me a convenient excuse to test GPT-6 Astra, which is absurdly powerful on paper. OpenAI says it reaches 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3, along with state-of-the-art performance across software engineering, computer use, browsing, and other professional work. More relevant to what I actually do, OpenAI specifically calls out Astra’s stronger visual judgment when building websites, applications, games, and other rendered experiences. Plus users are getting Astra through Codex and ChatGPT Work rather than GPT-6 Pro in ordinary chat, so for me it belongs mainly in the agentic-development part of the workflow anyway.
I also happened to receive two banked Codex resets during the rollout, so obviously I was going to spend them. But I’m not planning to use Astra for every task just because it is the new strongest option available to me. Sol is still my normal Codex model and my favorite for the kind of development work I do. I know I could theoretically stretch my usage further by dropping to Luna for simpler implementation work, and I may do that occasionally if the task is extremely basic, but realistically I’m probably going to keep clicking Sol. The point of optimizing my subscription is not to make the actual experience worse in pursuit of perfect inference thrift.
What I’m much more interested in is how much I can get out of two $20/month subscriptions: ChatGPT Plus and Claude Pro. I’ve paid for the $100 versions of both services at different points, and there are absolutely periods where I would upgrade again. But after using these tools heavily for design, development, writing, SEO, websites, a local LLM chat app, and an Android habit-tracking app, I’ve developed a process that lets me get a ridiculous amount out of $40 a month. I have no idea whether anyone else routes their usage exactly this way, and I’m not claiming I have uncovered the objectively optimal AI workflow. I just know that this one works very well for me.
How I Split Work Between ChatGPT, Claude Design, and Codex
AI workflows vary wildly when it comes to what tools are used for what part of the process. The tools are intuitive enough that many people don’t start with any particular tutorial and just figure out what works for them. As someone who uses AI for design, development, writing, SEO, and assorted projects I probably did not need to start, I’ve gradually developed my own preferences based on where each product adds something and where its limits become irritating.
Because I have ADHD, I’ve also hyperfocused on my AI workflow and spent many hours getting dopamine hits from AI responses, something I’ve already written about in my posts on AI-induced hypomania and ritual reuptake. AI and ADHD are a match made in heaven for me. I have too many ideas, and now I execute them and bring them fully to life at a speed that used to be impossible. My scifi series is unfinished, but I already have a coming-soon website and my own custom app for working on it with both API-based and local LLMs. I also vibe coded a personal Android habit-tracking app despite having no previous Android development experience.
Still, website development is where I have the strongest background and therefore the best frame of reference for judging the output. I already knew HTML, CSS, JavaScript, PHP, WordPress, SEO, and enough design to know when something looks wrong. That matters because AI becomes much more useful when you can tell whether it actually solved your problem or simply generated something plausible-looking.
For most projects involving a UI, my workflow now looks roughly like this:
ChatGPT chat: planning + SEO + writing + research + image generation + rough mockups → Claude Design: visual refinement + layout + spacing + UI structure + SVGs → Codex with Sol: functional implementation + debugging + testing + deployment
Those borders are obviously porous. Codex can plan. ChatGPT can make polished-looking mockups. Claude can write code. Astra can increasingly design and implement something inside the same workflow. I’m not sorting tools according to some rigid belief about what they are capable of doing; I’m sorting work according to where I personally get the most value from the limited parts of each subscription.
Unlimited ChatGPT Text Chat Is a Major Part of the Value
One of the biggest advantages of ChatGPT Plus for my particular workflow is surprisingly boring: I can keep talking to it. OpenAI’s current pricing page explicitly describes “everyday text chats” as unlimited across Free, Go, Plus, and Pro, subject to its usage policies. It even gives Free users unlimited text chats with GPT-5.6 Luna, while eligible paid plans get Sol for Instant and higher reasoning settings. OpenAI does not publish a neat number telling me exactly how much Sol inference I receive before something changes behind the scenes, so I’m not going to claim that Sol itself is literally unlimited. But I am a very heavy chatter and have not personally reached a point where normal ChatGPT usage became something I had to manage or where I routinely found myself forced onto Luna.
That is very different from how I experience Claude Pro. Anthropic’s $20 Pro plan has a session-based usage limit that resets every five hours, along with a weekly limit across models, and how quickly you burn through it depends on the model, feature, conversation length, attached files, and other factors. The limitation is not inherently unreasonable, but it changes my behavior because a long exploratory Claude conversation is consuming the same subscription capacity I may want to use for something more differentiated later. Codex on ChatGPT Plus has its own five-hour usage window as well, along with a weekly allowance. That means both of the expensive parts of my workflow can temporarily disappear on me, while ordinary ChatGPT text chat remains available.
With ChatGPT, I can spend a long time deciding what I actually want before I ask an agent to touch anything. I can discuss architecture, go down an SEO tangent, research a technical question, write some page content, generate a visual concept, change my mind about the feature, and eventually emerge with a much clearer specification. There is no reason I need filesystem access while I’m still doing that, and there is definitely no reason I need to burn through my Claude allowance doing it when regular ChatGPT gives me so much conversational room.
I Usually Plan Development Before I Open a Coding Agent
You can absolutely open Codex or Claude Code with a vague idea and tell it to start building. Sometimes I do. But planning is important to me, and I generally find it more efficient to separate figuring out what I want from asking an agent to implement it, especially when I’m working on something unfamiliar or a project where architecture, content, and design are intertwined.
When I decided to build my Android app, for example, I first used regular ChatGPT to understand how personal Android development worked, how I could sideload an app without publishing it through the Play Store, whether future updates could preserve its local data, and how I wanted the UI and logic to behave. By the time Codex actually started building, I had a much better understanding of what I was asking for. I wasn’t expecting the coding agent to simultaneously teach me Android development, invent the product, design the interface, determine the data model, and then implement everything correctly.
The same applies to websites. I may talk through what pages a site needs, what a particular interface should accomplish, whether WordPress or static HTML makes more sense, how a feature could be structured technically, what the responsive behavior should be, or what the agent needs to be careful not to break. I’ll also use ChatGPT for content strategy and SEO before implementation, because deciding what a page should rank for or what information belongs on it is not improved by giving the model command-line access.
This is not because Codex is incapable of planning. It’s because agentic usage becomes more valuable to me once there is something concrete for the agent to act on. If I have already made most of the subjective decisions elsewhere, the coding agent can spend its time coding rather than dragging me through decisions I could have made in a normal conversation.
It also means that hitting a Codex limit does not necessarily stop the project. If I manage to burn through Codex with hours left before my usage window resets, I can continue planning the next feature in regular ChatGPT, work on page content, research something, generate image assets, or prepare much better instructions for what I want Codex to do next. By the time the agent becomes available again, I often have a cleaner batch of work ready for it.
How I Use ChatGPT for Writing, SEO, Research, and Image Generation
Writing is another place where my use of AI varies enormously depending on what I am actually creating. I’m not using AI to write my scifi series, yet I’ve read thousands of scenes from my story that it produced. This helps me bring my worldbuilding and character development to life for testing new concepts and often sparks new ideas. ChatGPT is my go-to for this, hands down. I started with 4o, moved onto 5.1, spent a period exploring Claude when I was less happy with several newer GPT models, and now Sol has become my favorite again.
For writing that is less creative in nature, or sometimes not creative at all, I’m much more willing to simply let AI write. Writing for SEO is highly formulaic and something LLMs do very well, and my interest in doing all of it manually is admittedly low these days. I don’t think you can make a huge distinction between human slop and AI slop when it comes to formulaic types of writing created primarily for SEO purposes. Humans were producing enormous amounts of boring, repetitive search-engine content long before ChatGPT existed.
For my affiliate and informational blogs, the conversation could be limited to me having a topic idea, asking ChatGPT to help with search intent and the outline, and then letting it draft the post. If it is supposed to write in a relatively neutral tone, the result is often passable on the first try. Neural Ecstasy is different because the articles are based much more heavily on my opinions, experiences, and in-depth research. Even when a post is heavily AI-assisted, it is typically the result of long conversations I’ve had with ChatGPT, research I guided, an outline we refined, and substantial human editing afterward.
Providing references of my own voice helps, but ChatGPT still has some tendencies I routinely have to remove:
- Writing like a LinkedIn post with one short sentence, sometimes not even a full sentence, per paragraph
- Overstating the thesis repeatedly as if it becomes infinitely more profound every time
- Overly quirky wording related to raccoons, goblins, and anything Victorian
- “That matters because…”
- “It’s not X. It’s Y.”
Sometimes I’ll write most of an article myself and ask ChatGPT only for line edits. If I don’t feel like getting feedback that may annoy me, I explicitly tell it to stay focused on glaring errors, factual problems, and awkward lines rather than trying to redesign my prose. That works surprisingly well.
ChatGPT also handles a large portion of my research. For Neural Ecstasy, that can mean sharing an X post and asking it to find similar sentiments on Reddit and elsewhere, finding actual research papers that do or don’t support some random idea I’ve had about psychology, or collecting online reactions to a new model release. For my other websites, it may mean much more straightforward keyword research, product research, or technical questions.
Image generation belongs in this same stage of the workflow because ChatGPT is where I create most of the visual assets I later use elsewhere. I generate hero images, backgrounds, illustrations, icon concepts, article graphics, and rough page mockups. A ChatGPT-generated mockup can be genuinely useful and sometimes gets surprisingly close to what I want, but it is still an image of a website rather than a functioning interface. The spacing can look convincing without actually being consistent, text can become distorted, and a layout that looks great at one fixed aspect ratio can fall apart once someone has to translate it into responsive HTML. That is usually where Claude Design enters the process.
Why Claude Design Gets Most of My Claude Usage
I have increasingly little interest in using Claude as my main general-purpose chat tool. Part of that is personal preference because I find the tone of some of the newer Claude models more grating than I found Opus 4.6 and 4.7, especially when I’m talking through controversial topics or dark fiction. Part of it is simply resource allocation. If every long Claude conversation consumes the same limited subscription capacity I want to use elsewhere, I would rather have most of those conversations in ChatGPT.
Claude Design is different. Anthropic launched Claude Design as a dedicated environment for creating designs, prototypes, slides, one-pagers, and other polished visual work. It can create a first version from a prompt and then lets you refine it through conversation, inline comments, direct edits, and custom controls Claude can generate for the project. It can also work from a formal design system, although I usually do not bother with that feature myself.
For UI work, what I like is that Claude Design is actually constructing the interface rather than merely giving me a static image of one. That gives it a much better foundation for real spacing, typography, hierarchy, alignment, and the relationships between components. ChatGPT can give me a visually exciting first concept, while Claude Design is much better at taking that concept and making the whitespace make sense.
I usually create a Claude Design project for each UI and build multiple pages inside it. Sometimes I start by giving Claude a rough mockup I created in ChatGPT. Other times I give it screenshots of pages I already built and ask it to create another page that fits the same visual language. I may generate the hero imagery or other assets in ChatGPT and then ask Claude to build the actual layout around them. Because Claude Design works with something structurally closer to a website than an image generator does, it can turn a concept into a much more coherent UI.
This process involves a lot of refinement. I may adjust a section repeatedly because I think the whitespace is off, ask for a different layout, decide the revision feels too symmetrical, change the component structure, revisit the typography, adjust mobile behavior, add another visual, change the image crop, or ask Claude to create an SVG that fits the page precisely. Claude is especially useful for SVGs because those assets remain crisp and editable in a way that image-generated icons often are not.
There is also a recognizable “Claude website,” which is something I actively try to avoid. If you prompt Claude to design a website without much direction, it often produces something polished and high-end, but it also falls into recurring visual habits. One of the easiest examples is the pill-shaped eyebrow label above an H1 in the hero, though there are broader similarities in spacing, card treatments, rounded components, and the overall startup-polish aesthetic. I have adopted some of those choices myself because they look good, but I do not want every website I create to look like it emerged from the same Claude template.
That is why I spend so much time refining. I bring in my own imagery, push against layouts that feel generic, change shapes and spacing, ask for less symmetry, incorporate ideas generated somewhere else, and keep modifying the result until it feels specific to that project. The end result is usually much more original than what I would get from prompting Claude once and accepting its first attempt.
This can also absolutely annihilate my Claude usage. I can blow most of a five-hour allowance refining one complicated page because I will keep noticing things. I’ll move something, decide a section needs a different rhythm, hate the revision, revert it, adjust mobile, create a new SVG, change the content hierarchy, and then decide the entire thing needs ten percent more breathing room. This is apparently recreation for me.
Why I Barely Use Claude Code Even Though I Pay for It
Claude Code is included with Claude Pro, and I don’t avoid it because I think it is bad. I’ve used it plenty, and I have heard many developers argue that Claude Code is better than Codex. I’m not convinced there is a universal answer to that question because the tools and underlying models change constantly, different developers care about different things, and both products are extremely capable.
The reason Claude Code is not central to my personal workflow is mostly about where I spend my Claude allowance. Anthropic explicitly says Claude and Claude Code share the same usage limits, and IDE usage counts against that same pool. If I have already spent a large portion of the current window refining a UI in Claude Design, there is no separate fresh bucket of Claude Code capacity waiting for me afterward.
At that point, switching companies becomes useful for reasons that have almost nothing to do with a Claude-versus-OpenAI coding argument. I can take the completed Claude Design output, screenshots, exported files, or other visual references and give them to Codex along with the actual project. Codex can then reproduce the design while connecting it to whatever is really underneath it: WordPress templates, PHP logic, JavaScript interactions, static HTML, an Android project, or something else entirely.
That workflow has become one of my favorite uses of AI:
ChatGPT planning + images + rough mockup → Claude Design refinement → Codex implementation
By the time Codex starts, much of the subjective work has already happened. I know what the page should say, I have a strong idea of what it should look like, I already have the image assets, and I usually have either a visual reference or actual HTML from Claude Design. Codex can focus much more narrowly on turning that into a functional implementation. Could I keep the entire process inside Claude and move directly from Claude Design into Claude Code? Absolutely. The problem is that I would run out of Claude.
Sol Is Still My Default for Codex
Astra is interesting because it is brand new, performs extremely well on benchmarks, and OpenAI is specifically highlighting stronger visual judgment for the type of UI work I do. I’m going to keep testing it, especially on harder projects where I want more autonomy or where the visual and technical parts are intertwined. But Sol remains my normal choice inside Codex.
There is a theoretical optimization argument for switching to Luna on simple tasks because I could stretch my allowance further. I know that. Realistically, I’m not very motivated to do it because I like Sol, it reliably handles the kind of development work I throw at it, and optimizing my subscription does not mean deliberately making myself use a weaker model just because the task might technically survive it. The bigger savings come from where I perform the work rather than constantly downgrading the model.
If I do the planning, research, writing, SEO strategy, image generation, and visual ideation outside Codex first, then by the time I open the agent I am using its limited capacity for the thing I actually need it to do: interact with the project. That is also why Astra does not fundamentally change the workflow for me. It is another option inside the Codex stage, not a reason to collapse every part of the process into one giant agent prompt. Maybe its visual capabilities eventually reduce some of my reliance on Claude Design, but it has been out for three days. I am not reorganizing my entire process around benchmark charts quite yet.
What Two $20/Month Plans Let Me Do
The useful thing about combining ChatGPT Plus and Claude Pro is not simply that I have two companies available in case one gives me an answer I dislike. The products have different constraints, and I have found different places where each one adds the most value. ChatGPT gives me an enormous amount of conversational room. I can plan, research, write, generate images, create rough mockups, and generally figure out what I want without feeling like every tangent is stealing from Codex. Claude Pro gives me Claude Design, which remains the visual environment I most enjoy using for serious UI refinement. Codex then gives me a separate implementation agent without consuming whatever Claude capacity I just spent making the interface look good.
Even when I manage to exhaust both constrained parts of that workflow, I can still keep working in regular ChatGPT. This happens more often than I would like to admit. I have absolutely burned through both Claude and Codex with three or more hours remaining before their five-hour windows reset, and the fact that regular ChatGPT remains available means the project does not have to stop entirely. I can plan the next feature, prepare content, conduct research, generate graphics, create a rough mockup, or simply talk through what I want to do once the agents are available again. By the time the reset hits, I often have better instructions anyway.
When the $100/Month OpenAI and Anthropic Plans Are Worth It
I have paid $100 a month for both Claude and ChatGPT before, and there are plenty of completely practical reasons to do that. If you’re using AI heavily for software development, client work, research, content production, or other professional work where additional capacity directly saves billable time or produces revenue, the larger allowance can easily justify the price. A five-hour usage window becomes a much bigger problem when stopping means delaying actual work rather than temporarily abandoning whatever unnecessary website you decided to build on Saturday.
Of course, at an actual company you are more likely to be using one of the business, team, or enterprise offerings rather than buying everyone individual consumer subscriptions, and those have their own pricing and usage structures. I’m focusing on the $20 and $100+ individual plans here because that is the calculation I’ve personally made for my own projects. The broader point is simply that heavier professional workloads can justify paying for more capacity without ADHD entering the equation at all.
There is also a case for the stricter limits being useful rather than merely annoying. I once saw someone on Reddit describe upgrading to a higher AI plan as feeling like “a raccoon with ADHD on crack,” which unfortunately resonated with me immediately. In the same discussion, other users said they actually preferred having a hard five-hour limit because it forced them to stop, think about what they were doing, and come back with a better plan instead of continuing at light speed. I understand the argument even though I am personally very fond of continuing at light speed.
The $100+ plans can feel much closer to unlimited when you’re coding heavily, even though they still have limits. The smaller plans impose substantially more friction, but as I described earlier, hitting my Claude Design and Codex limits does not mean I have to stop working altogether. Regular ChatGPT is always available for planning, research, writing content, generating image assets, making rough mockups, or figuring out what I want to do once the agentic limits reset. That is one of the main reasons the $40 workflow works so well for me: the limits slow down one mode of work without necessarily stopping the project.
When I’ve temporarily upgraded for my own projects, the reason has been less financially respectable. Sometimes AI is basically candy for my ADHD tendencies, which is why that Reddit comment about feeling like a raccoon with ADHD on crack hit a little too close to home. I can get completely hyperfocused on a project when the page is improving, a feature has finally started working, every useful response reveals another obvious improvement, and the entire process keeps feeding itself. I have absolutely burned through both Claude and Codex with multiple hours left before the reset, and in that state the usage window feels less like a reasonable product constraint and more like an obstacle between me and the thing I have suddenly decided must exist tonight.
When I’m building an entire site, migrating content, creating templates, fixing responsive problems, implementing complicated functionality, or starting an app from scratch, I may genuinely want to stay in that flow rather than take the forced break. I have upgraded during exactly those periods because paying more was preferable to having the project abruptly stop while I was still completely locked in. The stricter $20 limits may encourage more reflection and restraint, but sometimes I knowingly choose the raccoon-on-crack option.
The fact that I can justify the $100+ plans during those periods of intense hyperfocus does not mean I need to pay for them every month. For normal ongoing work, the two $20 subscriptions force me to ration the expensive parts without preventing me from continuing to think or create. When I know I am about to spend an entire weekend tearing a project apart and rebuilding it, maybe I upgrade. Otherwise, $40 gets me surprisingly far.
AI Is Changing Which Development Workflows Make Sense
My decision to build the newest site in static HTML is a small example of a broader shift I keep seeing. I still like WordPress and still build WordPress websites. For sites that need a CMS, nontechnical editors, complicated dynamic functionality, or the huge ecosystem of plugins WordPress provides, I am not replacing it with a folder full of HTML because an AI model looked at me funny.
But some of the friction that historically made static websites inconvenient is disappearing. If an agent can update components across the entire site, fix markup, change navigation, adjust schema, and maintain the codebase for me, the absence of a CMS becomes less annoying on projects where I am the only person managing the content anyway. Suddenly the fact that GitHub Pages is free becomes more interesting.
The same pattern shows up elsewhere. Before AI, I would not have decided to learn enough Android development to build one tiny personal habit tracker because the activation cost would not have been worth it. Building a separate local app for chatting with LLMs would also have competed directly with actually writing the scifi series the app is supposedly helping me write. Before Claude Design and modern coding agents, producing a highly polished custom UI meant either spending much more of my own time on design and implementation or bringing in someone else.
AI has not made my existing expertise irrelevant. It has made that expertise easier to apply across more projects. I still need to know when the code is behaving strangely, when the SEO strategy makes no sense, when the content is bullshit, and when Claude has produced another attractive interface that looks suspiciously like every other attractive Claude interface.
If you don’t know what good looks like, AI can still help you produce garbage faster. Very scalable.
Experiment to Find Your Ideal Workflow
I’m not proposing that everyone needs ChatGPT Plus and Claude Pro or that my particular split is objectively optimal. Someone who cares much less about visual design may get almost no value from Claude Design. A developer who strongly prefers Claude Code may happily keep most of the workflow inside Anthropic and pay for more capacity. Someone else may decide Astra is so good that they want to keep nearly everything inside Codex.
Your workflow will also depend heavily on what you already know. Mine reflects the fact that I care about SEO, writing, design, and development and enjoy being involved in all of them. I don’t particularly want an agent to disappear with a vague idea and return with an entire product I never thought through. The planning and refinement are part of what I enjoy, even when they are also the reason I end up spending three hours adjusting one page.
What has worked for me is being deliberate about where the constrained usage happens. I use regular ChatGPT aggressively for thinking, research, writing, images, and rough visual concepts. I let Claude Design consume most of my Claude allowance because design is where Claude currently differentiates itself most clearly for me. Then I use Codex, usually with Sol, when I want something implemented in the real project.
Astra is now another option inside that system, and I’m interested in seeing where it proves meaningfully better. For now, two $20 subscriptions let me get an absurd amount done, with the $100 plans waiting for those periods when I have decided sleep and restraint are temporarily less interesting than finishing whatever project has captured my attention.
