Which AI tools I pay for, and what I run locally for free instead

2026-07-31 · Yusuf Gadelrab

I run a 30-billion-parameter model on my own laptop and it costs me nothing per token. That fact changes which AI subscriptions are worth paying for, and it eliminates more of them than most people expect. It does not eliminate all of them, and pretending otherwise is how you end up doing four hours of unpaid work to avoid a $22 charge.

Setup, so you can calibrate. I am a CS student. I run a quantized mixture-of-experts model locally through a small HTTP server, wired into my own automation scripts. It drafts, it classifies, it summarizes, it does first-pass code review. It runs overnight on a schedule while I sleep and costs me electricity.

So the question I have had to answer honestly, repeatedly, is: what is a paid AI tool doing that my free local one cannot?

The rule that decides it

Run it locally when the model is doing thinking. Pay for it when the product is doing something other than thinking.

That is the whole framework and it has held up for a year. The valuable part of most paid AI products is not the model, it is everything around it: proprietary data, hardware, real-time meeting capture, a rendering pipeline, a permissions system, a voice trained on ten thousand hours you do not have.

Text generation is thinking. A local model does it, badly at first, well enough after you learn to prompt it, and it never charges you or rate-limits you or trains on your half-finished ideas.

Voice synthesis that does not sound like a robot is not thinking. It is an enormous specialized model plus a training set. You cannot replicate it on a laptop, and the open-weights alternatives are, as of now, audibly worse in a way listeners notice within four seconds.

What I run locally, and what it genuinely costs

• First drafts of anything. Emails, outlines, documentation, commit messages. Quality is roughly a good intern. That is exactly right for a draft.
• Classification and extraction at volume. This is where local wins outright. Running thousands of items through an API costs real money and real latency. Locally it is a loop I start before bed.
• Anything touching private data. Financial records, visa paperwork, unpublished research. This is not a cost argument, it is a boundary. Some things do not leave the machine.
• Overnight batch work. Scheduled jobs at 4:30am. Zero marginal cost means I can afford to be wasteful with attempts, which is a genuine capability, not a saving.

The real costs nobody mentions: about 20GB of disk, a laptop that gets hot and loud, a context window smaller than the frontier hosted models, and a first weekend spent on setup. And the honest quality gap. On hard reasoning, long documents, and anything where being subtly wrong is expensive, the frontier hosted models are still clearly better. I use local for volume and for privacy, not because I think it is equivalent.

What survived, and why

Voice: ElevenLabs

This is the clearest case in the whole list. The gap between the best hosted voice synthesis and what I can run locally is not subtle, and audio is unforgiving: listeners detect wrongness in a second and stop trusting the content.

If you are making anything with narration, ElevenLabs is the one I would pay for first. If you are only doing occasional narration, the free tier is real and generous enough to test properly before you spend.

Video and podcast editing: Descript

What you are buying from Descript is not transcription. It is the interface idea that editing the transcript edits the video. That is a product insight, not a model, and no local setup gives it to you.

I built an ffmpeg-based auto-editing pipeline myself, so I know precisely what the free path costs: it is a weekend, and then it is brittle, and then every new edge case is another evening. My pipeline is worth it because it is a product I am building. If editing is not your product, do not build the pipeline.

Meetings: Fireflies or Otter, and probably neither

The model part of meeting notes is trivial. The hard part is being reliably present in the call, handling the calendar integration, and getting speaker separation right on bad audio.

Fireflies and Otter both do this well. But I want to be straight with you: as a student with a handful of club and lab meetings a week, I do not pay for either. I record and run a local transcription model. If you are in six external calls a day and the notes have consequences, the calculation flips completely, and that is who these are built for.

Marketing copy at volume: Copy.ai, with a caveat

Templated marketing copy is exactly the workload a local model handles well, so my default answer is do not pay. The case for Copy.ai is team workflow rather than generation quality: shared brand voice, approvals, non-technical teammates who will never touch a terminal. If you are one person with a laptop, run it locally.

The mistake I made, so you can skip it

For about two months I ran everything locally on principle, including things where the local model was plainly worse. I lost hours to re-prompting a model into an answer a better one would have given me first time. That is not thrift, it is paying yourself minimum wage to protect an identity.

The correction was to price my own time honestly. If a tool saves me three hours a month and costs $20, that is a trade I take without thinking. If it saves me twenty minutes, no, and the local model gets another prompt.

I write up the rest of my stack, including the parts that failed, on the stack page.