AI Wins › Replaced paid software

A tech writer built his own stand-ins for $400 a year of app subscriptions, using a $20 AI plan

Verification: High Replaced paid software AI used: Claude (US$20/month plan), plus free open-source tools and a local AI model · October 2026

High verification: A first-person account by a named How-To Geek journalist, and the open-source tools he built on are public and checkable. How we check stories

What happened

In a first-person piece published on October 10, 2026, Dibakar Ghosh wrote for How-To Geek¹ that he set out to test the "AI will replace software subscriptions" claim on his own bills. He used his Claude subscription to "vibe code" (describe what you want and let the AI write the code) three swaps.

• Grammar checking. Instead of a Grammarly subscription, he pastes his text into Claude and asks it to build an Artifact (a small interactive panel inside the chat) that shows the passage with errors struck through and the fixes beside them, so he can accept or reject each one. He notes that the same idea works with Gemini Canvas and ChatGPT Canvas.
• Dictation. In place of the Wispr Flow dictation app, Claude suggested pairing two free open-source projects, RealtimeSTT² and faster-whisper³, then wrote him a Windows script that starts with his PC. One key records and pastes a transcript; another types as he talks.
• Time tracking. Rather than pay about US$120 a year for Rize's basic plan, he installed the free, open-source ActivityWatch⁴ tracker. A small AI model running on his own computer sorts the data, and Claude built the web-page dashboard that charts his day.

His point is that the AI plan is a workshop, not another subscription to lean on: once the scripts and pages exist, they keep working on his machine. In his words, the real value is in "building offline tools for yourself."

What's the catch?

Readers pushed back in the comments. One asked whether a US$240-a-year Claude plan really counts as saving US$400. Ghosh replied that he keeps paying for Claude because he uses it for work and writes about it, but that all three tools could be built in a weekend, and wouldn't need Claude again afterwards.

Some of the "intelligence" also moved onto his own hardware. He runs the dictation model on an Nvidia RTX 3060 graphics card with 12GB of memory (he found 6GB enough), and he uses a local model, Gemma 4 12B in LM Studio, to tag his activity so the data never leaves his PC.

Why this matters if you're choosing an AI plan

• One US$20 plan can do more than chat: used well, it can build small tools that replace paid apps.
• The savings come after the build. You can keep the tools even if you cancel the plan later.
• Hardware and skill matter: some swaps need a decent graphics card and a willingness to tinker.

Credits. Original first-person reporting by Dibakar Ghosh for How-To Geek. RealtimeSTT, faster-whisper and ActivityWatch are open-source projects by their own developers. Grammarly, Wispr Flow and Rize are trademarks of their owners, and none of them is connected with this story or with us. Up to Speed with AI has no relationship with the people or companies named, and nothing here suggests they endorse us or any AI plan.

Spotted an error, or are you in this story and want something changed? Tell us and we'll fix or remove it.