There is a version of this article on every website in the world right now, and almost all of them are copying each other's numbers. So I did the boring thing instead. I opened the actual pricing page for every paid tool on this list, read the actual licence file for every free one, and pulled the star counts and download sizes from the live repositories on the day I wrote this. Four of the claims that get repeated about these tools turned out to be wrong, including two I had to correct in my own video.
Here is what survived. Eight AI tools that do serious work. Four of them cost nothing at all, and I mean nothing: not a trial, not a free tier that expires, but open licences on your own machine that nobody can put a price on later. Between them they stand in for about $76 a month of subscriptions, which is $912 a year. And then I am going to tell you exactly how much of that $76 actually holds up, because it is less than the headline, and the honest number is the useful one.
The video counts these down and shows each one running. This article is the receipt: every price with the plan name and the billing basis attached, every licence named exactly, and the corrections I owe you where the video and the internet got something wrong. Read it, watch it, or both.
First, the number on that thumbnail. It says $224 a month and this page says $76, so let me square that before you read another line. The $224 was a headline figure, and it is not one I can stand behind, because the video's own description adds the same tools up at the same tiers and lands between $72 and $81. Nothing in the film ever justified the larger number. $76 is the figure I checked, tier by tier, on each company's own pricing page, monthly toggle confirmed. And by the end of this article even that comes down to about $46, because two of the eight swaps do not survive scrutiny. The smaller number is the honest one, and I would rather correct my own thumbnail than let it sell you a saving that is not there.
The 8 at a glance
| # | Tool | What it does | Stands in for | What it really costs |
|---|---|---|---|---|
| 1 | Voicebox | Clones your voice, speaks any text in 23 languages, locally | ElevenLabs Creator ($22/mo) | Free, MIT |
| 2 | Meetily | Records, transcribes and summarises meetings on your machine | Granola Business ($14/user/mo) | Free, MIT (speaker labels are paid) |
| 3 | Qwen-Image 2512 | Image generation with genuinely correct text in the picture | Midjourney Basic ($10/mo) | Free, Apache 2.0 |
| 4 | MiniMax H3 | Short video clips with dialogue, effects and room tone baked in | Runway / Veo-class credits | Free weights, or credits on the hosted app |
| 5 | Nemotron 3 Ultra | 550B open model for long-running coding and agent work | Cursor Pro ($20/mo) | Free weights, brutal hardware |
| 6 | Screencap | Records how you work and makes it searchable months later | Loom / Scribe-class tooling | Free personal, $9/mo business |
| 7 | TubeTutor | Turns a YouTube tutorial into written steps with timestamps | Your own hour | $9.90/mo Basic |
| 8 | Mubert | Generates original music with no copyright claim attached | Suno Pro ($10/mo) | Free non-commercial, $39/mo to monetise |
The honest math: what a free stack actually saves you
Start with the real number for the paid side. These are the current monthly-billed prices, taken from each company's own pricing page today: ElevenLabs Creator is $22 a month, Cursor Pro is $20, Granola Business is $14 per user, Midjourney Basic is $10, and Suno Pro is $10. That is $76 a month, or $912 a year.
Now the part most articles skip. Not all four free swaps work for a normal person.
Voicebox for ElevenLabs Creator saves $22. This one is clean. It runs on a laptop.
Qwen-Image for Midjourney Basic saves $10. Also clean, if you have the disk.
Meetily for Granola Business saves $14. Clean for recording, transcription and summaries. Speaker labels are the exception, and I cover that below.
Nemotron 3 Ultra for Cursor Pro does not save you $20. The weights are free. Running them is not. NVIDIA's own model card lists the minimum deployment as eight GB200, B200, GB300 or B300 GPUs, or sixteen H100s, or eight H200s, with roughly 1.5 TB of aggregate memory for a single-node BF16 run. That is a data centre, not a desk. You can rent it through a router instead, but then you are paying per token, which is a different bill, not no bill.
Mubert for Suno Pro does not save you $10 either, for reasons I will get to at number eight.
So the honest saving, for one person on one laptop, is about $46 a month. That is $552 a year, and it is real. It is just not the $912 the headline arithmetic implies, and you should know which one you are being sold.
1. Voicebox: voice cloning that never leaves your laptop
Voicebox is the one I would install first. You record a short sample of your own voice, and after that it will say anything you type, in 23 languages, still sounding like you. Nothing is uploaded. The project's own words are that models, voice data and captures never leave your machine, and it serves a local API on your own computer rather than calling out to anyone.
The clever part is underneath. It is one app sitting on top of seven different open speech engines, Qwen3-TTS, Qwen CustomVoice, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, HumeAI TADA and Kokoro, so you can swap the model without learning a new tool. It ships with over 50 preset voices if you would rather not use your own. It is MIT licensed, it went up on 25 January 2026, and it sits at 50,202 stars today.
The correction: you will see "clones your voice from three seconds of audio" everywhere, including in my own video. The project does not say that. Its actual wording is "a few seconds," and I could not find a three-second figure anywhere official. It is a small thing, but it is the kind of small thing that turns into a lie once ten blogs repeat it.
The catch: there are no pre-built Linux binaries, so Linux users build from source. Paralinguistic tags like [laugh] only work on Chatterbox Turbo; other engines will read them out loud as words. And no official RAM or disk figures are published, so budget generously.
What it replaces: ElevenLabs Creator at $22 a month, which is the tier that unlocks professional voice cloning. Alternatives worth naming if Voicebox does not fit: ElevenLabs itself if you need the polish and the studio tooling, Coqui-derived community forks if you want something older and simpler, or Kokoro on its own if all you need is a clean preset narrator rather than your own voice.
2. Meetily: meeting notes with no bot in the room
Meetily, from Zackriya Solutions, records your meeting, transcribes it as it happens, and writes the summary before you close the laptop. The thing that makes it different from every other notetaker is that no bot joins the call. There is no "Recording Bot" sitting in the participant list making everyone self-conscious. It captures your device's system audio the way a screen recorder does, which means it works with Google Meet, Zoom, Teams, Discord, Slack Huddles, Webex and anything else that makes sound.
It is written in Rust, MIT licensed, and at 29,009 stars today. All processing happens on your machine, transcription through Whisper or Parakeet, summaries through a local Ollama model. If you would rather have a cloud model write the summaries, you plug in your own API key for Claude, Groq, OpenRouter or OpenAI and pay for that part yourself.
The correction, and it is the important one: speaker labels are not free. I said in the video that it works out who is speaking. Diarisation, the feature that tags each voice in the transcript, is explicitly a Pro feature at $10 per user per month billed annually. The free MIT edition transcribes and summarises. It does not tell you who said what. If your notes depend on attributing quotes to people, price that in.
The catch: the published minimum is 8 GB of RAM, a four-core CPU and 10 GB of storage, plus Node.js 18+, Python 3.10+ and FFmpeg. macOS and Windows have builds; Linux is build-from-source.
What it replaces: Granola Business at $14 per user per month, or Fireflies. One honest note on Fireflies, because it is quoted wrong constantly: their $10 and $19 tiers are the annual-billed rates. Month to month, Fireflies Pro is $18 and Business is $29. Other alternatives: Otter if you want the long-established name, or Anarlog (formerly Hyprnote) if you want another open, local-first take on the same idea.
3. Qwen-Image 2512: the open model that spells
Ask most image models for a poster with a headline on it and the headline comes back as confident gibberish. Alibaba's Qwen-Image family is the exception. Words land correctly, in the right place, at the right size, which quietly makes it the most useful free image model for anyone making thumbnails, ads, covers or signage.
There is a fork in the road here that trips people up. Qwen-Image 2.0, announced 10 February 2026, is API and chat only. You can use it, you cannot have it. The one you can actually download is Qwen-Image-2512, a 20-billion-parameter model under Apache 2.0. Editing an image you already have is a separate open model, Qwen-Image-Edit-2511, also Apache 2.0. The date codes are December 2025, not 2026, which is another number people get wrong.
Alibaba's own claim is that 2512 is "currently the strongest open-source model," based on more than 10,000 rounds of blind human evaluation on AI Arena. That is their claim on their own model card, and I am reporting it as their claim, not as an independent finding.
The catch: the full repository is about 57.7 GB. Community quantised GGUF builds bring that down to roughly 14 GB at q5, with smaller options down to 6.7 GB, but those are made by the community, not released by Alibaba. Do not expect an official small version.
What it replaces: Midjourney Basic at $10 a month monthly-billed, $8 if you pay yearly. Alternatives: Flux for a stronger open general-purpose look, Stable Diffusion through a local front end if you want the biggest ecosystem of community models, or Ideogram if text rendering is the only thing you care about and you are happy to pay.
4. MiniMax H3: short video that arrives with its own sound
Most video models hand you a silent clip and leave the audio to you. MiniMax H3, also branded Hailuo 3, generates the picture and the sound together in one pass: dialogue, effects and room tone, in native stereo at 32 kHz. Clips run 4 to 15 seconds at 24 frames per second. It is a 33-billion-parameter dense Omni Transformer. The hosted app launched at the end of July 2026 and the weights landed on Hugging Face a few days later.
The correction: "15 seconds at 2K" is generous. The model card puts base output at 768p on the shorter side. 2K comes from a separate regenerate module, not from the model's native generation pass. Say 768p natively, 2K with an extra step.
The catch, and it is a real one: the community licence carries a geographic condition. Users in the United States, the European Union, the United Kingdom and South Korea have to go through an application form. Everywhere else the licence covers local deployment directly. Also, the Hugging Face repository totals 498 GB across two checkpoints and their encoders, so there is no tidy single download; community quantised builds exist in the 9 to 22 GB range but they are third-party. VRAM requirements are not published officially, so ignore the specific numbers floating around blogs, including the reassuring ones.
What it costs if you use the hosted app: credits. The Hailuo Standard plan is $14.99 a month for 1,000 credits, up through Pro at $54.99 and beyond. Free accounts generate with a watermark and wait in a queue. Unused credits do not roll over. Alternatives: Runway or Google's Veo if you want polish and a support contract, or Wan and LTX-Video if you want other open weights to run locally.
5. NVIDIA Nemotron 3 Ultra: the biggest brain you can download, and cannot run
This is the one that sounds like science fiction and behaves like an electricity bill. Nemotron 3 Ultra is 550 billion total parameters with 55 billion active per token, a latent mixture-of-experts hybrid Mamba-Transformer built for long-running coding and agent work, with a context window up to one million tokens. NVIDIA released it under OpenMDW 1.1, cleared for commercial and non-commercial use, and they genuinely released the training data and recipes alongside the weights, not just the checkpoints. That last part is rarer than the parameter count and matters more.
The correction: it is not "the largest open model anyone has released." Moonshot's Kimi K2 has been publicly downloadable since July 2025 at roughly a trillion parameters and about 1.03 TB of weights. Nemotron 3 Ultra is NVIDIA's largest open model, which is still a big statement. Also worth knowing: NVIDIA's December 2025 press release previewed Ultra at "about 500 billion parameters" with "up to 50 billion active" and attached the 1M context window to the small Nano model instead. The final numbers come from the June 2026 model card. If you see 500B/50B quoted, that is someone citing the press release.
The catch: the BF16 weights are about 2.24 TB across 225 files, and the minimum deployment is eight B200-class GPUs or sixteen H100s. Nothing about this runs at home.
What it replaces, honestly: it does not replace your Cursor Pro subscription unless you rent serious hardware. What it does is give the people building on top of open models a frontier-class base with no licence trap. Alternatives that actually fit a laptop: Qwen3-Coder, DeepSeek's open coder models, or GLM's open releases, any of which you can run locally or route to cheaply.
6. Screencap: your own work, made searchable
Screencap solves a problem you probably have not named. You do a job once, correctly, and three months later you cannot remember the seven small decisions that made it work. Screencap records the session, and an on-device model splits it into separate labelled tasks, indexing both the spoken words and everything that appeared on the screen. So months later you search for "payroll run" and land on the exact minute you did it.
It is local first. The recording stays on your machine until you choose to share one. Free for personal use, $9 a month for business after a seven-day trial. It requires macOS 14 or later on Apple silicon; Windows, a browser extension and Android are all listed as coming.
The nuance worth knowing: it is source-available, not open source. The licence is PolyForm Noncommercial 1.0.0, which means you can read the code and use it personally, but it is not an OSI open-source licence and commercial use is the paid path. That is a fair deal, it is just not the same thing as MIT. Alternatives: Loom if you want the polished sharing side, Scribe if you want automatic step-by-step documentation, or Guidde for a more enterprise-flavoured version of the same idea.
7. TubeTutor: the tutorial, without the tutorial
The one answer you need is buried somewhere in a two-hour video. TubeTutor watches it and hands back the part that applies to you. Paste a YouTube link and it returns the tutorial as text, reading the on-screen menus, settings and code as well as the narration, so the steps come out in order with timestamps you can jump straight to. You also get the full transcript, a mind map and the key takeaways.
The correction: the free tier is not "two videos a day." I could not find that anywhere on their site. What is actually published is one guest preview of up to five minutes with no card required, and 200 starter credits valid for 30 days when you sign up. Paid plans are Basic at $9.90 a month for 1,000 credits and Advanced at $19.90 for 2,500, with credit packs available separately.
What it replaces: your own hour, mostly. Alternatives: Claude or ChatGPT with the transcript pasted in if you only do this occasionally, or NotebookLM if you want to keep a library of sources and question all of them together.
8. Mubert: the one I have to walk back
Mubert generates original music that has never existed, which means there is no rights holder to file a claim against your video. It has been running since 2016, so the licensing side is properly built, and with music the licensing matters far more than the sound does. You pick a mood, a genre and a length, and it writes the track.
The correction, and it is the biggest one on this page. I said Mubert's Creator plan at $11.69 a month covers videos you make money from. It does not. Two things are wrong there.
First, $11.69 is the annual-billed rate. The monthly price of Creator is $14.
Second, and this is the one that matters, Mubert labels Creator as "for personal projects and non-monetized videos," explicitly non-commercial. Monetised channels need Pro, at $39 a month monthly-billed, or $32.49 a month billed annually. The free Ambassador tier is non-commercial with attribution, 25 generations and five MP3 downloads a month.
So for anyone monetising video, Mubert is not the cheap option, it is the expensive one. Suno Pro, the tool it was supposed to be replacing, is $10 a month. That swap is backwards, and I would rather say so here than let it sit. Alternatives for licensed music you can monetise: Suno Pro at $10 a month, Udio, Soundstripe or Epidemic Sound for human-made catalogue licensing, or Stable Audio Open if you want weights you actually own.
The four claims that did not survive a pricing page
Pulling these together, because this is the part you can use on any list like this one:
- "Nemotron 3 Ultra is the largest open model anyone has released." It is NVIDIA's largest. Kimi K2 is bigger and has been downloadable since July 2025.
- "Meetily works out who is speaking, for free." Speaker diarisation is a Pro feature at $10 per user per month billed annually.
- "Mubert Creator at $11.69 covers monetised video." Creator is non-commercial. Monetisation is Pro at $39 a month, and $11.69 was the annual rate for a $14 plan.
- "TubeTutor is free at two videos a day." The published free tier is one five-minute guest preview plus 200 starter credits for 30 days.
There is a pattern behind all four, and it is worth more than the list. Three of them are billing-basis errors: a pricing page defaults to the annual toggle, someone reads the big number, and the annual rate gets published as the monthly one. Suno's page does this, Fireflies' page does this, Midjourney's does this. Before you repeat any AI tool's price, find the monthly/annual switch and check which side it is sitting on. The fourth is a tier error: the feature exists, it is just not in the free tier, which is the whole business model of open-core software. Free and open source rarely means all of it is free.
What "free" actually costs
Nobody puts this in the listicle, so here it is. Free tools cost you three things.
Disk. Qwen-Image is 57.7 GB for the full repository. MiniMax H3's repository totals 498 GB. Nemotron is 2.24 TB. You will run out of drive before you run out of enthusiasm.
Hardware. Voicebox and Meetily are laptop-friendly. Qwen-Image wants a decent GPU or a lot of patience. Nemotron wants a rack.
Time. Every one of these is an install, a model download, and a first run that does not work. If your hourly rate is real, a $10 subscription can be the cheaper option, and there is no shame in that. The reason to run the free ones is not always money. It is that nothing leaves your machine, and nobody can raise the price or change the terms on you later. For a lawyer on a client call or a clinician with a patient, that is not a saving, it is the whole requirement.
What I would actually install this week
If you make content: Voicebox first, then Qwen-Image. Those two are the clean wins, they run on a laptop, and together they take $32 a month off your bill. If you monetise video, keep paying for music. Suno Pro at $10 is cheaper and safer than Mubert Pro at $39.
If you sit in meetings: Meetily, today. Free, local, no bot in the room. Add the $10 Pro tier only if you genuinely need to know who said what.
If you build: skip Nemotron unless you have the hardware, and run a smaller open coder model locally instead. Point your agent at whatever is cheap, and spend the attention you save on what you are building rather than what you are building it with.
If you train or document anything: Screencap on a Mac, and be honest with yourself that the business tier at $9 is the real price once you use it for work.
The part that actually pays
Here is the thing I keep coming back to. A free stack does not make you money. It lowers your costs, which is a smaller and much slower lever than it feels like when you are cancelling subscriptions on a Sunday night. Saving $46 a month is one good afternoon of client work.
What these tools genuinely change is the floor. A voice, a face, an image, a soundtrack and a finished clip used to require either money or other people, and now they require neither. Which means the constraint has moved entirely to the thing it was always going to move to: knowing what to make, and knowing who pays for it.
That is the whole reason IdeasRepay exists. It takes you from "I have the tools" to knowing the exact business to point them at and actually having it built.
Come and build one of these
Everything at ideasrepay.com is a walkthrough rather than an article. You do not read about a business, you click through building one: the tool stack with current prices, the prompts in the order you actually run them, the templates and scripts as real downloads, what to charge, the conversation that closes the client, and a first-week checklist that ends at a paying customer instead of a folder of software.
The closest one to this page is The UGC Studio. Brands burn through ad creative faster than they can produce it, and the casual, authentic-feeling videos that actually convert are exactly what they cannot make in-house. You become the studio that makes them, on camera or completely faceless with tools like the ones above. Inside is every ad format they buy, the scripts, what to charge per video, where the clients are, and the pitch that turns one job into a monthly retainer.
One payment of $99 opens that walkthrough and every other one we have published, plus every one we publish after it, across all three verticals: online businesses, YouTube and content, and offline work in the real world. Downloads and templates included, no renewals, no upsells, and you can email us while you build. That is launch pricing, and it moves to $199 after the first 500 members.
You are also not doing it alone. There is a community building alongside you, a channel where I teach this every week, and me, mentoring through it personally. If you want to sharpen the one skill that makes every tool on this list better, the prompting, our free Claude prompting course is open to everyone. If you want the wider view of where this leads, The Only 13 Ways to Make Money With AI maps the whole field, and Claude Design Free covers the open-source design side of the same stack.
Cancel what you are not using. Then go make something with what is left.