🧮 Free Unlimited AI Video and Voice: The Math Says No

I wrote a 30-line script whose only job is multiplication: seconds of video times the provider's published price per second. No error handling, no tests, no dependencies. It still falsifies more marketing copy per minute than any other tool I own.

Point it at any page whose title tag contains "free unlimited AI video" and the outcome is the same. OpenAI's Sora 2 API bills $0.10 per generated second, with Sora 2 Pro between $0.30 and $0.70 depending on resolution. Google's Veo 3 on Vertex AI is published at $0.75 per second, with native audio adding roughly 50%. So a single eight-second clip costs whoever pays the bill somewhere between $0.80 and $6.00, before a single retry. Nobody absorbs that at unlimited volume for anonymous visitors who have not entered a card.

"Unlimited" is not a feature. It is a tell about the pricing model.

By the numbers#

• $0.10/sec for Sora 2, $0.30–$0.70/sec for Sora 2 Pro: an 8-second clip runs $0.80 at 720p to $5.60 for Pro at 1080p (CostGoat).
• $0.75/sec published for Veo 3 on Vertex AI, or $6.00 per standard 8-second clip; native audio adds roughly half again (Veo3AI).
• $1.49–$6.98/hr for an NVIDIA H100 across 15+ cloud providers, with the on-demand market average sitting around $3.00/hr (IntuitionLabs).
• 10,000 credits/month on the ElevenLabs free plan, about 10 minutes of Multilingual v2 speech, with no commercial rights and mandatory attribution (CostBench).
• 20,000 characters per week is the real cap at TTSMaker, one of the most-cited "free unlimited" TTS sites, with only a subset of voices genuinely uncapped.

Do the arithmetic yourself, in ten seconds#

Take the advertised unlimited service. Assume it is fronting a frontier model. Multiply your intended monthly output in seconds by $0.10 as a floor. Sixty clips of eight seconds is 480 seconds, or $48 a month at the cheapest published rate and $360 at Veo's. Now ask what the site earns from you. Display ads on a generator page clear a few dollars per thousand sessions on a good day. The gap is three orders of magnitude wide.

If the operator is running open weights on rented silicon instead, the floor is the H100 line: roughly $3.00 an hour of wall-clock GPU time, whether or not anyone is watching. A queue of free users is a meter that never stops.

So what: any page promising unlimited frontier-quality generation for nothing is either capping you somewhere it has not told you, or charging you in a currency other than money.

The four different things called "free"#

These have completely different failure modes, and no roundup separates them.

1. Venture-subsidised loss-leader. Real model, real quality, priced below cost to buy market share. The cap exists but sits on an axis the landing page does not mention: queue position, resolution, watermark, export rights. Failure mode is a pricing change with 30 days' notice.

2. Resale of an endpoint the operator does not pay for. Someone found an undocumented service and wrapped it. Failure mode is instant and total: the upstream closes the door and every downstream site dies at once.

3. On-device or in-browser inference. WebGPU running a small model on your own machine. Honestly unlimited, because your laptop's fan is the invoice.

4. Open-weight local models. Same principle with better quality and real licences.

The 30-second identification test: open devtools, hit generate, and watch the network tab. Zero outbound inference requests means category 3. A request to the vendor's own API means 1 or 2, and the difference is whether that vendor has a public price list for the same model. If the model has no documented price anywhere, assume 2.

So what: only categories 3 and 4 can use the word unlimited without lying.

Where the paywall actually falls, by modality#

The bait-and-switch is consistent enough to predict.

Voice cloning. Creating the clone is free. Generating with it is metered. That order is deliberate, because once your voice is uploaded you are far less likely to walk. ElevenLabs' free tier excludes instant voice cloning entirely.

Dubbing. The preview plays free. The clean, unwatermarked export in your target language is the gate. Compare per-minute rates in our AI dubbing and translation tools guide before committing a back catalogue.

Text to speech. Caps are per week or per session, not per month, which makes them feel unlimited during evaluation and bite in production. TTSMaker's actual number is 20,000 characters a week.

Video. The cap is seconds and resolution, never clip count. HeyGen's free plan allows 3 standard plus 3 Avatar IV videos a month, each up to one minute at 720p, each watermarked.

So what: read the pricing table for the word "per" and check what noun follows it.

What you are actually licensed to publish#

This is the question the unlimited pages never answer, and it is the one that costs money later. ElevenLabs' free plan carries no commercial usage rights and requires ElevenLabs attribution on every public file. HeyGen's free exports carry a visible watermark with no workaround at that tier. Neither of those is a bug. They are the product working correctly.

Open weights invert it. Kokoro-82M ships under Apache 2.0, Chatterbox under MIT, Wan 2.2 under Apache 2.0. Commercial use is permitted, no attribution is owed, and nobody can retroactively decide your back catalogue needs a logo. Our ElevenLabs v3 guide covers where the paid tiers still earn their keep.

Free is a payment method#

Voice.ai's terms state that by using the services, users consent to the company utilizing their hardware for metamodel training and computational purposes. Your machine becomes part of the fee. GetVoices' terms require uploaders to grant a worldwide, non-exclusive, royalty-free, sublicensable and transferable licence to use, reproduce, modify, distribute and create derivative works from that content, plus a waiver of moral rights and rights of publicity and privacy. On Google's Gemini API free tier, submitted data may be used to improve Google's products; the paid tier excludes that. Pro models also left free-tier access on 1 April 2026, leaving Flash and Flash-Lite.

Before uploading a voice sample anywhere, search the terms for four strings: "sublicensable", "derivative works", "moral rights", "improve our services".

So what: the price of free voice cloning is sometimes a perpetual licence to your own voice.

The edge-tts loophole, and why it expires#

A large share of "unlimited free TTS" sites, browser extensions and Python wrappers are front ends for Microsoft Edge's Read Aloud service, reached through the edge-tts library. It needs no API key, no Windows, and no Edge install, and it exposes hundreds of voices across dozens of languages. The unlimited claim is technically honest, because the operator meters nothing they pay for.

It is also an undocumented, unlicensed dependency with no SLA and no deprecation policy. I could not get a straight answer about rate limits from any of the three wrapper sites I tested; my script stopped receiving audio from two of them after a few hundred sequential requests, with no documented error and no header explaining why, so I cannot tell you whether that was an upstream limit, the wrapper's own throttle, or my IP being quietly dropped.

The path that is genuinely unlimited#

Local generation costs time instead of money, and the honest numbers matter more than the enthusiasm.

TTS: Kokoro-82M, 82 million parameters, Apache 2.0, 54 voices across 9 language and accent groups, running at or near real time on an ordinary CPU with no GPU and no cloud call.

Voice cloning: Chatterbox from Resemble AI, MIT licensed, zero-shot cloning from roughly 5 seconds of reference audio, 23 languages, PerTh watermarking embedded by default. In published blind testing, 63.75% of evaluators preferred it to ElevenLabs.

Video: Wan 2.2, Apache 2.0, the most permissive licence in open video. The TI2V-5B variant renders a 5-second 720p clip at 24fps on a single 24GB consumer card such as an RTX 4090 in under about nine minutes.

Nine minutes per five seconds is the real constraint. Ten finished clips is an evening. Card selection matters more than model selection here, which is covered in the best GPU for local AI guide, and the rent-versus-buy comparison sits in local vs cloud AI.

Three readers, three correct answers#

Occasional (under ~10 clips or 10 minutes of speech a month). Genuine free tiers are sufficient. Stay inside them, accept the watermark on drafts, and buy credits for the two files that get published. Start from the free AI video generators guide.

Semi-pro (weekly output, some revenue). Paid credits beat hardware until your render volume is predictable. Run the numbers on the cost calculator with your actual monthly seconds rather than an optimistic estimate.

Production (daily output). Local for the repetitive volume, paid API for the hero shots that need frontier quality. The AI video cost breakdown shows where that crossover lands per model.

Assume every endpoint is temporary#

The Sora 2 and Sora 2 Pro APIs are live and scheduled to sunset on 24 September 2026. That is a first-party, paid, well-documented endpoint from a well-funded lab, with a published expiry date. Free wrappers get no such courtesy; they simply stop returning audio.

Keep your scripts, reference audio, prompt files and source stills in your own storage, in plain formats, with the model and settings recorded next to each output. Do that and a shutdown costs you a re-render. Skip it and a shutdown costs you the archive.