Anyone Can Build Software Now. So What Are You Paying For?

On 30 September a developer put a free copy of Photoshop on GitHub.

Thirteen days later it has more than 57,000 stars. It was built with Claude, written in Rust, and it’s one of seven Adobe apps the same developer has copied. And by 9 October the studio behind it was showing copies of Word, Excel and PowerPoint too.

So is this the end of Adobe and Microsoft?

Not this year. The builds are rough. HotHardware’s write-up talks about noticeable glitches, unpredictable tool behaviour and performance hiccups, and nobody running a creative team is swapping Creative Cloud for an alpha on GitHub.

The more useful question is closer to home. If one developer can put out a free, working copy of Photoshop, what are you actually paying for when you rent software?

I’ve got skin in this. At G3NR8 we build software on our clients’ own data, and we train their teams to use AI at work. So every week I see what’s suddenly easy to build and what’s still hard to own.

Here’s what the last few weeks tell you, where it’s heading, and what to do about it before your next renewal.

What Actually Got Copied

The Photoshop copy is called PhotoCraft, part of a set its studio calls ArtCraft. It’s open source under the Apache 2.0 licence, which means anyone can use it, change it and build a business on top of it.

On 9 October Paul Thurrott reported that the same studio is going after Office. By its own count, the Word copy covers 87% of Word’s ribbon commands, the Excel copy covers 89% of Excel’s ribbon commands and about 93% of its worksheet functions, and the PowerPoint copy covers 79%.

Treat those percentages with care. They’re the studio’s own figures, the Office apps are pre-alpha, and the last few per cent of any product tend to take the longest.

Even so, copying most of the best-known office software on the planet is now a side project for a small team. That used to take a big team and a lot of money.

Why now? Because the engine underneath got very cheap.

The Engine Got Cheap, and a Lot of It Is Chinese

Anything built this way leans on an AI model, and models are priced per token, the chunks of text a model reads in and writes out. Those prices have dropped hard, and a lot of the pressure comes from Chinese labs releasing open models anyone can download.

DeepSeek’s top model writes text for about a fifth of the price of Anthropic’s top model. Its Flash model is cheaper again, and off-peak it’s half price.

Are they any good? Stanford’s 2026 AI Index says the performance gap between US and Chinese models has effectively closed, with Anthropic’s top model ahead by just 2.7%. Epoch AI puts the best open models about four months behind the best closed ones.

And people are using them. Alibaba’s Qwen models were downloaded more than 2 billion times on Hugging Face in the first seven months of this year, roughly five times Google’s figure.

On OpenRouter, a service that sends requests to hundreds of different models, Chinese open models averaged about 13% of all usage in a study of 100 trillion tokens, and hit nearly 30% in some weeks.

Airbnb’s Brian Chesky said it plainly last year: “We’re relying a lot on Alibaba’s Qwen model. It’s very good. It’s also fast and cheap.”

There’s a catch, and it’s the one your IT team will raise first. Australia banned DeepSeek on government devices in early 2025, Taiwan and South Korea’s trade ministry restricted it, and Italy’s privacy regulator moved to limit how it handled Italian users’ data.

Downloading an open model and running it on your own servers is one thing. Sending customer data to an API hosted in China is something else entirely. Plenty of people building tools at their desks won’t know which of the two they’re doing.

Building Software Is Now Normal

Cheap engines and good models mean building has gone mainstream.

In April, Sundar Pichai said 75% of Google’s new code is now AI-generated.

Lovable, one of the tools where you describe an app in plain English and get a working one back, says more than 60 million projects have been built on it since November 2024, and those apps get more than 900 million visits a month. Replit made $240 million in revenue last year and has more than 50 million users.

Inside companies, Retool’s 2026 Build vs Buy report found 35% of teams have already replaced at least one software tool with something they built themselves, and 78% expect to build more this year. Those are Retool’s own customers, so the sample leans towards builders. It’s still a third.

So is everyone ripping out their software? Hardly. Menlo Ventures found 76% of enterprise AI use cases were bought in 2025, up from 53% the year before. Most companies still buy, and that share went up.

What’s changed is the option. A year ago, telling a supplier “we could just build it” was mostly bluster. Now it’s a real choice for a lot of the smaller tools sitting around the edges of a business.

Building Got Cheap. Owning It Didn’t.

Here’s the bit the free Photoshop headlines skip.

The expensive part of software has always been keeping it running. Someone has to fix it when it breaks at 3am, patch it when a security hole turns up, change it when the rules change, and understand it after the person who built it leaves.

AI makes the building much faster. The upkeep is still there, and on some of it AI makes life harder.

Veracode tested AI-written code from the big models again this summer. It still fails security tests 44% of the time when nobody asks for security, and that figure hasn’t improved in a year, even as the models got smarter.

GitClear looked at 211 million changed lines of code and found that, for the first time, copy-and-pasted code outnumbers code that’s been tidied up and reused. That’s the stuff that makes software hard to maintain later.

Speed can be an illusion too. In METR’s 2025 study, experienced developers using AI took 19% longer on real tasks while believing they’d been 20% faster. The tools have moved on since, but the gap between how fast it feels and how fast it is should worry anyone signing off a build.

And when it goes wrong, it goes wrong quickly. In July 2025 Replit’s AI agent deleted a live database during a code freeze, holding records on more than 1,200 executives. Replit’s own CEO called it “unacceptable and should never be possible”.

Gartner expects half of enterprises to face delayed AI upgrades or rising maintenance costs by 2030 because of unmanaged technical debt from generative AI.

Remember Klarna “replacing Salesforce with AI”? Its CEO later set the record straight: “no, we did not replace SaaS with an LLM”. Klarna built its own internal data stack. That’s a serious engineering team taking on a serious system, with every bit of upkeep that comes with it.

So when you pay a software company, a big chunk of what you’re buying is ownership. Somebody else takes the 3am call.

Your Staff Are Already Building

Meanwhile, inside your business, people have worked out they don’t need IT to build a tool.

In September, Dataiku and the Harris Poll asked 685 CIOs across eight countries, including the UK. 84% said employees are creating AI agents and apps faster than IT can govern them. 81% said they don’t have complete oversight of agents built outside approved channels. And 72% can’t consistently measure whether their agents deliver anything for the business.

Picture what that looks like on the ground. Someone in finance builds a little app to match invoices. It works, so the team starts relying on it. It runs on their personal account, uses whichever model was cheapest that week, and nobody else knows how it works. Then they go on holiday.

That’s shadow IT with an engine. Gartner predicts more than 40% of organisations will suffer security or compliance incidents from shadow AI by 2030.

I’m all for people building. It’s how the best ideas get found. But a tool the business relies on needs rules and an owner, like anything else the business runs on.

What You’re Actually Paying For

Now look at the other side of the table.

Gartner forecasts worldwide software spending of $1.44 trillion this year, up 15.1%. Cheaper building hasn’t made anyone’s bill smaller yet.

Microsoft 365 went up on 1 July. E3 went from $36 to $39 a user a month and Business Standard from $12.50 to $14. Salesforce has rebuilt its editions around agents at $195, $395 and $550 per user per month.

And a lot of what you pay for, nobody uses. Zylo tracked 40 million licences and $75 billion of spend for its 2026 index. On average 36% of licences sit unused. That’s $19.8 million a year at the average organisation, and $3.8 million even at companies with up to 500 staff.

The way you’re charged is shifting as well. When Bain looked at how 30-plus software firms priced in AI, about two thirds bolted a usage meter on top of seats and about a third simply raised the seat price. Intercom’s Fin charges $0.99 per resolution. Salesforce prices Agentforce at $2 a conversation.

Gartner reckons up to $234 billion of application software spend is at risk by 2030 as agents start doing the work people used to log in for. Gartner’s George Brocklehurst summed it up: “you are increasingly buying it for agents”.

So what does that mean for you? Seats for people who never log in, and tools you rent for a single feature, are the first places to look. The big systems your data lives in will stay, and they’ll probably cost more.

Where This Goes Next

The markets have already had a wobble. In February, software stocks lost about $300 billion of value in a single session after Anthropic launched its Cowork tools and Palantir reported results.

Gartner expects 40% of enterprise apps to have task-specific AI agents built in by the end of this year, up from under 5% in 2025. In its best case, agentic AI drives about 30% of application software revenue by 2035, more than $450 billion.

It also calls generative AI and agents the first real challenge to mainstream productivity tools in 30 years. Last week’s Word and Excel copies are what that looks like up close.

Here’s my view. The software you rent is splitting in two. The core systems your business runs on, the ones holding your customers, your money and your stock, get stickier and dearer, with agents built in. Everything around the edges, the single-purpose tools and the spreadsheets held together with copy and paste, becomes something you can build, buy cheaply or pick up as an open-source copy.

The businesses that come out ahead will know exactly what they’re paying for, and they’ll own what they build.

That’s how we work at G3NR8. We build on top of the systems clients already have, using their own data. At a European industrial distributor we deployed operational intelligence across 30+ data sources and identified EUR 45m of at-risk revenue in 6 weeks. It’s a production system the sales team uses every day.

What to Do Before Your Next Renewal

The free Photoshop makes a great headline. The story that matters for your business is sitting in your own software bill.

So when did you last check what you’re actually paying for?

Sources & References

Frequently Asked Questions

Roy Murphy is co-founder of G3NR8. This post draws on 2025 to 2026 research on AI-built software, model pricing, software spend and how staff build with AI. Published 13 October 2026.