How I used AI to decide which £1,000 computer to buy

My son starts secondary school in September, so I’ve been looking for a desktop computer.

I wanted something he could use for homework, Minecraft and Roblox now, but that wouldn’t stop him experimenting with Roblox Studio, Python, Blender or whatever he gets interested in next.

I wanted to use it too: writing, coding, AI experiments, creative work and the occasional game.

And I had a fairly simple limit. About £1,000 felt reasonable. £1,200 started to feel like too much.

This should have been a straightforward shopping decision.

Instead, I accidentally turned it into a month-long experiment in how I use AI to buy things.

The specification started determining the budget

I started where most people probably would: what is the best PC I can get for roughly £1,000?

ChatGPT helped me compare prebuilts. Very quickly I was discussing Ryzen versus Intel, RTX 5060 versus 5060 Ti, 8GB versus 16GB of VRAM, DDR4 versus DDR5, AM5 upgrade paths, power supplies, motherboards and 1440p performance.

All useful information.

But something else was happening.

The £1,000 computer became a £1,080 computer because the graphics card was better. Then £1,200 started looking defensible because the platform was newer. Then a 16GB graphics card looked attractive for local AI. At various points we were looking at machines approaching £1,500.

This is one of the dangers of very good comparison tools: the specification can start determining the budget instead of the budget determining the specification.

So I changed the question.

Not: what is the best computer?

But: what am I actually buying this computer to do?

What I actually need

For my son, the progression I care about is fairly simple:

play → modify → build → code

Minecraft and Roblox are not separate from the educational justification. They’re probably the route into it. If Roblox becomes Roblox Studio, or Minecraft shaders become curiosity about how graphics work, or he decides to try Python or Blender, I don’t want the computer to be the thing that stops him.

But I don’t need to buy today for every hypothetical thing he might become interested in three years from now.

For me, it needs to be a useful shared workstation too. Most of that means writing, browser-based AI tools, building small products and ordinary creative work. I kept letting local AI and high-end gaming influence the specification, even though neither is a central reason for buying it.

I also wanted one large monitor rather than a dual-screen setup, and a decent keyboard I would enjoy working on. That mattered because I was buying a complete working environment, not just comparing towers.

That brought the budget back under control.

The computer I thought I would buy

The first version of this article ended with a CyberPower desktop at Costco for £999.99:

AMD Ryzen 5 7500F
NVIDIA GeForce RTX 5060 8GB
16GB DDR5 RAM
1TB NVMe SSD
Windows 11

It was a sensible recommendation. The newer AM5 platform gave it a better upgrade path, while the NVIDIA card suited Blender and local AI experimentation.

But it was also still hypothetical. I hadn’t bought it, and the machine later became unavailable.

That distinction matters because recommendations are easy to optimise in theory. A real purchase has stock, delivery, tax, accessories, warranty and a moment when somebody has to decide that the search is over.

Both models got the easiest fact wrong

During the earlier search, I had been keeping the decision in a GitHub document called secondary-school-desktop.md: budget, shortlist, rejected options, reasons, intended uses and open questions. Instead of restarting the research every time I came back to it, the document accumulated the thinking.

Eventually the document itself contained a prompt asking Claude to challenge the recommendation ChatGPT and I had reached.

Claude did.

It questioned whether I was overvaluing the more expensive RTX 5060 Ti. It looked at the 8GB memory limitation. It found evidence that the Costco machines had previously been discounted much more aggressively than the prices I’d been comparing.

Claude also said it couldn’t reliably verify the live price. That limitation was useful because it told me exactly what I still needed to do.

Then I took its findings back to ChatGPT. It investigated the price and initially agreed that £899.98 was current. On that basis, it changed its recommendation and told me to buy.

I opened Costco on my phone.

£999.99.

The models had spent a lot of time correctly discussing relatively complicated things — GPU performance, VRAM limitations, CPU platforms and upgradeability — and still got the simplest decision-critical fact wrong: what I would actually have to pay.

That didn’t make the AI research useless. It showed that different sources were answering different questions.

A historical deal price tells me that a product gets discounted. A search snippet tells me what a search engine has indexed. A metadata field tells me what a page exposes to a machine. The customer-facing product page tells me what the retailer is asking me to pay.

I thought that was the end of the lesson. It wasn’t.

The cart found another truth

When I finally chose a different CyberPower machine, I pasted the configuration into ChatGPT for one last check.

The displayed total was £921.50. It included the PC, a 34-inch AOC ultrawide monitor and a Keychron mechanical keyboard. ChatGPT treated that as the final price and called it an unusually strong deal.

Then I opened the cart summary.

The £921.50 was before VAT. The checkout added £184.30, taking the amount I would actually pay to £1,105.80.

The information was live this time. The arithmetic was easy. The mistake was in understanding what the number represented.

That is a more interesting failure than an outdated search result. “Current” is not the same as “complete”. A visible price might exclude tax, delivery, required storage, a warranty or the accessories that make two offers comparable. The closer I moved towards the transaction, the more authoritative the evidence became.

The product listing helped me find the machine. The configurator told me what I had selected. The cart told me what I would pay.

Verification wasn’t a single final check. It had layers.

What I actually bought

I bought the CyberPower Ultra 55 with:

AMD Ryzen 5 5500
AMD Radeon RX 9060 8GB
16GB DDR4 memory
1TB NVMe SSD
Windows 11
AOC 34-inch 3440×1440 curved monitor
Keychron C2 Pro mechanical keyboard
five years’ labour and two years’ parts warranty

See the CyberPower Ultra 55

The tower worked out at £849 including VAT. The monitor was £198 and the keyboard £58.80. Delivery was free, bringing the complete setup to £1,105.80.

There are better computers. I found a Costco machine with a much stronger Intel i5-14400F processor and an RTX 5060. Once I included the same monitor, keyboard and the £42 Costco membership I would need, the real comparison was about £1,250.

For another £144, I would have bought meaningful additional processing power. But that did not mean I needed it.

The cheaper machine is more than capable of Minecraft, Roblox, schoolwork, writing, browser-based AI tools and the kind of building either of us is likely to do next. Its processor is older and its 8GB graphics card won’t make full use of a 180Hz ultrawide monitor in demanding games at maximum settings. Those are genuine limits, not reasons to keep spending.

The decision came back to the requirement: nothing we currently want to do becomes possible at £1,250 that is impossible at £1,105.80.

My AI shopping workflow now looks like this

The useful pattern I’ve ended up with is:

Ask → compare → challenge → narrow → price the whole requirement → verify at checkout → decide

AI was extremely useful in the middle of that process.

It turned a vague requirement into criteria. It explained technical differences in normal language. It searched a much wider market than I would manually. It challenged my assumptions. A second model challenged the first. A persistent document meant neither had to start from scratch.

What I couldn’t outsource was the small number of volatile facts capable of changing the decision, or the judgement that enough research had been done.

For this purchase, the volatile facts were price, stock and what the displayed total included. For something else it might be availability, a warranty condition, a cancellation clause or the denominator behind a statistic.

The answer isn’t to manually verify everything AI tells me. That rather defeats the point.

It’s to know which facts deserve the expensive check, and how close to the underlying action that check needs to happen.

That’s an editorial judgement problem as much as an AI problem.

The document mattered more than either model

I’ve written before about context as capital and Portable Intelligence. This shopping exercise made both ideas much more tangible to me.

The important thing wasn’t whether ChatGPT or Claude was better.

The decision existed somewhere outside either conversation.

ChatGPT could build on it. Claude could challenge it. I could correct both. New prices and rejected options could go back into the same record instead of disappearing inside another chat.

The file got smarter even when the models were wrong.

The models also failed differently. Claude was useful when it admitted it could not verify the price. ChatGPT was useful when it searched current alternatives, but it was also too confident about both an old Costco price and the pre-VAT CyberPower total. The human role wasn’t to stand outside the process and distrust everything. It was to notice what kind of claim was being made, decide what evidence could support it and keep responsibility for the commitment.

That’s a much more useful way to think about AI than choosing a favourite chatbot.

Buying was part of the research

I could have kept going. There will always be a newer platform, a stronger processor or a deal that appears the week after purchase.

But comparison has diminishing returns. Past a certain point, more options stop improving the decision and start postponing it.

The final computer is not the theoretical best one I found. It is the one that meets the actual requirement at a price I was willing to pay, with the monitor and keyboard I wanted, available when I was ready to buy.

My son still doesn’t know how much thought went into it. He just wants to play Minecraft with shaders on.

The purchase finally gave this article its ending. More importantly, it exposed the part of AI-assisted decision-making that no comparison table can complete: somebody still has to look at the final number, accept the trade-offs and press buy.

A note on how this article was made

The first version of this article was written by Claude after it reviewed the PC decision and noticed that the multi-model workflow itself was interesting.

ChatGPT then challenged Claude’s findings. I caught a pricing error by opening the retailer page myself. The first published version still ended before the purchase, with the Costco machine as the likely choice.

Nine days later, the actual buying process produced a different machine and another pricing error. I extended the article with ChatGPT around that completed journey and edited it before republication.

That process is part of the point: AI made it possible to capture and develop an observation I probably wouldn’t have spent an evening writing from scratch. But the evidence kept changing, and so did the article. The judgement about what survived, what changed and what I was willing to buy remained mine.