My Real AI Tool Stack: What I Actually Pay For in 2026

Introduction

I run multiple businesses across the UK, Saudi Arabia and Dubai. I build my own software tools. I use AI every single day. And I’m going to show you exactly what I pay for, what I use, and what I have dropped.

No affiliate links. No sponsorship deals. Just an honest breakdown of what a real business owner’s AI tool stack looks like in 2026.

Most of the tool stack videos you see online come from people who have never run a business. They are reviewing tools in a vacuum. I’m using these tools to run actual operations, build actual software, and serve actual customers.

By the end of this article, you will know exactly what is worth paying for and what is a complete waste of money.

The Mistake I Made

Before I show you my stack, let me tell you the mistake I made.

Last year, I was paying for six different AI subscriptions and maybe using two of them properly. I had overlap in tools, doing the same thing. I was paying for features that I did not need because a YouTube video told me I should. And I was switching between tools every few weeks, chasing the latest FOMO model release.

Here is the thing they do not tell you. Switching between AI tools costs you 15 to 30% of your initial setup investment every time you swap. Your staff lose their prompt mastery. Your integrations break. And you spend weeks reconfiguring instead of working.

Consistency with one well-configured tool beats chasing the latest model every time.

I eventually sat down, audited everything, and cut my stack down to what actually delivers value. What I am about to show you is the result of that audit. Every tool earns its place or gets dropped.

Coding Tools: My Development Setup

Let me start with the biggest area of my stack, which is coding.

All of my internal software tools, I have built personally for my business. I run two coding setups in parallel.

The first is Cursor with Claude on Anthropic’s Max Plan.

Cursor is an AI code editor. It looks and feels like VS Code, which is another IDE, but it has AI built directly into the coding experience. I run it through Terminal on Claude, and I find it is the quickest and best way to do it. It understands your entire codebase. It can suggest code changes across multiple files and can refactor whole sections of code in context.

The Max Plan I am on is the £90 a month plan. I was on the £200 plan when I had loads of stuff going on, but I have ramped that down to £90. Although recently, the tokens have been getting absolutely vacuumed up for some reason. Apparently, they are trying to fix this. But yeah, the £90 a month plan is more than enough for your business.

I basically have switched between Opus 4.6 and Sonnet 4.6 and sort of balance between the usage limits depending on what I am doing, like if it is task specific.

Also, Cursor Pro has a subscription which is about £18 a month. And with that, you also get Cursor Bug Bot, which I use as my independent auditing tool. So essentially, what happens is you do any changes on Cursor, you then push it into your remote GitHub project, and Bug Bot is built into the Cursor Pro plan. It audits all your code and then spits out comments and replies. I really like that as someone who is not a software developer myself. This gives me another layer of checking. I suppose auditability on the code that I produce.

So that is one side of the setup that I use.

Clothing Cursor to be honest, I will do most of the building there, like building the plan. So from plan to initial MVP. And then once I have got the basic working version of the tool, generally I switch it across to Codex, and I use GPT 5.4 normally on medium or high. I basically use that. That is kind of a bit of a sniper approach. If you know exactly what you want to do, you just direct Codex and GPT 5.4, and it will fix those issues one by one. That is how I like to build.

But I like to start with Claude to build the outline and the structure of the tool. I think Claude is also better on the front end as well, if you develop in like your dashboard or whatever that might be.

Most of the stuff I will build, I will run locally for myself and for the business. But some of the stuff we will put on Railway. Our main ERP is on Railway. When it is on Railway, we are hosting a custom domain, and then basically all our team can get into it. So that is where we have our custom-built ERP for the business, which covers all the marketing, sales, and cash flow.

If you were looking at going with something like Auto or whatever, it would be costing you an absolute fortune. You have 10 users on there, you are talking at least £300 to £500 a month. This basically costs me for the domain and cost me for the hosting. Pretty much that is it, you know, £10, £20 a month, and I have got a full custom ERP system for my business. You could not do this two or three years ago.

When you run an SME and a small business, every single penny counts. So if you can develop your own custom tools for your business, this is absolutely advantageous in this day and age for business owners. And this is what I teach on the channel.

So Codex with GPT 5.4, normally on medium or high compute. Codex is opening eyes. Code and Assistant is actually a really nice tool to use. You know, when you are running code in Terminal, it looks a little bit like you have just transported back to 1993, and it is a little bit scary for non-technical people. But I would say Codex is a much cleaner and nicer app to use. So it is kind of a bit of a personal choice which one you want to work with.

I say Codex is actually easier to look at and work with in the app. But I think Cursor and Claude is actually better for beginners because I think the models are a bit more beginner-friendly. It kind of explains stuff as well. A lot more when you are building things, it will tell you exactly what it is doing and what it has done wrong.

What I always do when I am building anything is actually read what it is doing. It sounds mental. And this is where I hate the Terminal and VS Code is because I think vibe code is literally typing, “build me an ERP SaaS for my water company,” and then someone just plugs it online. That is like vibe code. But what I try to do is I do like layered code of small functions and features. So you do one tiny little function. It might be, I do not know, add a delete button to that article card. You put it on there, you test it, you run it through Cursor Bug Bot, you pull it, you test it online, and then you fix it. And then you go back and test it in production. And then you run agents for security or refactoring, whatever that might be along the way. So you are building something that is pretty solid.

That is why I actually hate the terminology vibe code.

Why do I run two setups? Because they are good at different things. As I mentioned, Cursor and Claude is really good for interactive development and for big picture stuff. But Codex for me is kind of like once you have got something built, if you like layer new things on there, or if you are doing debugging or fault finding, I think Codex with GPT 5.4 is just ninja.

The important thing is that I am building real business software with these tools, not just demos, not side projects. Actually, the systems that I am building, I am using them in my business daily. And I think this is the best way to do it.

They are internal tools. They are not public-facing. I mostly see a lot of them, and my team share some of them. Also, what we did when we went to production, we hired a senior developer to review the codebase and make sure it was pretty safe and there were no glaringly obvious things like APIs in the code and all that sort of joyful stuff.

We do that, and that is what we recommend on this channel. We always recommend develop your own internal tools, get them to MVP standard, and then hire somebody who is a professional senior systems engineer who can review the security and the structure to make sure there are no leaks.

So that summarises the coding tools.

General AI Tools: ChatGPT, Claude, and Claude Code

For general day-to-day work, I use ChatGPT. Not as much now. I think I am using it less and less. Although I did check a lasagna recipe last night, and it was absolutely excellent. So I guess it is not really business-related, but good lasagna recipe on there.

And also predominantly now, I am using Claude and Claude Code. To be honest, I was a bit skeptical about Claude Code to start with, but actually it is a bit of a game changer.

I think the way you need to look at these tools is I think there is quite a clear distinction between how you should use these models. So I think first of all, it is all on sensitivity of data. So if there is anything that is sensitive for my business, that is all offline. It is all offline models we use, open-source models connected to our ERP system. So all the data stays on our servers basically.

But then the frontier models, which are ninja like Claude Code and all the rest of it, actually what we do with them is we use them for marketing, we use them for coding, building the structure of things. But all the stuff that goes into their software and tools is all kept offline for obviously GDPR and security and all the rest of it.

Claude Code is amazing. It is really, really amazing. Things like to-do lists, things like connecting it to your website to do blog posts and content and marketing, and all that sort of stuff is super cool. It can generate flyers and presentations.

I am going to do a bit of a video on Claude Code, but I have only really started using it the last week or so. So I am just kind of working it out myself. But I have built some skills. I have built some MCPs. All these cool tech words. Maybe I sound like a cool software engineer. I built some MCPs and plugged them into my software. So there are certain functions, like in the marketing and content generation and even some of the sales stuff, project tracking, and stuff like that. It is all plugged into Claude Code. And I am basically managing probably about 30 to 40% of my business in Claude Code in the chat box. Which I believe is pretty cool.

So going back to general day-to-day work, ChatGPT and Claude, probably 80-20. I would say with Claude, you can do things like email, drafting, content, planning, research, analysis, document reviews, obviously nonsense of stuff as well. But I use it in a lot of things I do, really.

And generally, the rule of thumb for me is anything sensitive, do not use it, or use open-source offline models which we have in the business. Which I will do another episode about later.

And if you are going to generate anything that is going to be publicly facing from your business, you need to properly check it before you post it. There is so much AI slop online at the minute. It is absolute trash. Because obviously someone just learned how to use ChatGPT, and they have just typed, “create me a LinkedIn post,” banged it up there, “create me an image for a LinkedIn post,” banged it up there. They think it is amazing because it is the first time they have actually done it. But actually, most people are just like, what is this absolute shite.

So yeah, check it. And you need to make it custom. All right? If you just start using ChatGPT, go online, have a look on YouTube, and look how to properly use it. Because typing in a prompt with no context, no background, nothing specific to your business or enterprise or you as a person and posting it online is absolute laziness. You should be ashamed.

Practical advice is to pick one of these software that you are most comfortable with. To be honest, Claude and ChatGPT basically operate kind of the same. You go projects in there, you have got attachments, you can link documents, all the rest of it. The interface is pretty much the same. I would say at the moment, Claude is probably the stronger model. But watch the space for OpenAI because they have got quite a bit of money and quite a lot of clever people who work for them.

Image Generation Tools

For images, I am using a combination depending on what I need.

Primary setup is Google AI Studio. Nano Banana 2.5. Not the Pro one because that is crazy expensive. But the normal one, 2.5, basically gives me high-quality image generation, good control of style, and input things like temperature and stuff are good. It gives me high-quality image generation with good control of style and output. I use it for thumbnails, presentation graphics, and any visual content that needs to look professional. The quality is the best, to be honest. Like, you do not need the Pro one. Just get the normal one. It costs around about 10 pence an image. That is what I am working it out for.

For quick and dirty work, internal concepts, well, things that do not need to be polished, I am using Grok at the moment. I am not a huge fan of Grok. I think it is quite good. But the quality is pretty good, just for like whacking out standard images. And it is cheap and fast. You can do it all in, I think, it is an £8 a month plan. For generally Grok, it is quite good for that.

And just for low-quality, standard images you might need for blog posts or presentations or company media, I do not use Midjourney. To be honest, I tried to use it about six months ago, and I got really, really upset with the whole Discord channel thing. Now it might have changed since then, but it was just really clunky to use. But I do understand that it has got really good quality. So maybe I need to have a look at that.

The practical lesson here is you do not need to use Nano Banana for everything. There are other options. Like Grok is pretty good for low-quality stuff. So if you want something high quality, like a real final version, then definitely go and check out Google AI Studio. You just need to sign up and get a little API link and subscribe. It is not too difficult to set up, or there are plenty of other platforms out there where you can actually just link to them. But obviously, they are going to be more expensive because they make a bit of margin on it.

API Models for Custom Software

So this is where it gets interesting for anyone building custom tools.

In our ERP software, where I am not using Claude Code to generate things, we have got it connected up to OpenRouter for some of our tasks. And that is the API gateway. And then OpenRouter is really good. It gives you an opportunity to use multiple models through one API. And that basically connects into our ERP system.

So if you are building your own tools, you can connect them to different AI models through API. That gives you flexibility. You can switch between models depending on what you need the task to do, without paying for multiple subscriptions.

This is particularly useful for internal business tools where you want to integrate AI into your own workflows without relying on the public-facing chat interfaces.

Key Takeaways

• Stop paying for six subscriptions when you only use two. Audit your tools regularly. Every tool must earn its place or get dropped.

• Switching tools costs you 15-30% every time. Your team loses prompt mastery, integrations break, and you spend weeks reconfiguring. Pick one well-configured tool and stick with it.

• Use the right tool for the right job. Claude is better for building structure and front-end work. Codex with GPT is better for debugging and adding features to existing code.

• Keep sensitive data offline. Use open-source models for anything sensitive. Use frontier models like Claude Code for marketing, content, and general tasks.

• You do not need the most expensive options. The £90 a month Anthropic plan is enough. Nano Banana 2.5 is better than the Pro version. Grok is fine for quick, low-quality images.

• Always check AI-generated content before posting. There is too much AI slop online. Take the time to make it custom and relevant to your business.

• Build your own internal tools. It costs a fraction of commercial software. A custom ERP can cost you £10-20 a month in hosting instead of £300-500 a month.

Over to You

So there it is. My real AI tool stack in 2026. The tools I actually use to run multiple businesses, build my own software, and serve customers.

The biggest lesson here is that you do not need to pay for everything. You do not need the latest model. You do not need six different subscriptions. You need a few well-chosen tools that work together, and you need to stick with them.

What does your AI tool stack look like? Are you using similar tools, or have you found alternatives that work better for your business? I’d love to hear what’s working for you — drop a comment below.