DaVinci AI
One subscription across roughly 28 image models and six video models, plus voiceover, music, artwork, upscaling and an editor. An aggregator, not a model lab.
How we got access DaVinci supplied credits to a normal account so we could test the product, and every generation described here was made with those credits. No payment was made for this review or the score. DaVinci carries a ToolChase badge linking back to us.
What DaVinci AI is
DaVinci AI is an aggregator. It does not train the models it sells. What it sells is one login, one credit balance and one interface sitting in front of other companies' engines, and the breadth is larger than the video framing suggests.
On video the picker offers Veo, Kling V3 Turbo, Seedance 2.5 and Seedance 2.0 Fast, MiniMax H3 and Wan 3.0. On images it lists roughly 28 models, including GPT Image 1 and 2, Nano Banana, Nano Banana Pro and Nano Banana 2, Flux 2, Flux 2 Turbo and Flux Pro Kontext, Ideogram 3.0 and 4.0, Recraft V3, V4.1 and V4.1 Pro, Seedream 4.0 through 5.0 Pro, Qwen Image through Qwen Image 3, Grok and Grok Pro, Adobe Firefly, Krea 2 and Microsoft MAI Image 2.5. One house model, DaVinci Ultra, sits in the same list.
Beyond generation there are five more surfaces: an Artwork generator driven by named moodboard presets, a voiceover library with categories and named voices in several languages, a music generator with genre presets and auto lyrics, an Enhancer that the vendor says upscales to 22K, and an Editor for images and video capped at 75MB and 15 seconds. Video adds Edit Video and Motion Control tabs, start and end frame slots, and separate reference slots for image, character and clothing.
The trade-off is the one every aggregator carries. You are renting access rather than buying a relationship with any model vendor, so quotas, prices and model availability can move underneath you, and output quality is whatever the upstream model does that week. In exchange you skip six or seven separate subscriptions and get a comparison surface none of the individual vendors has a reason to build.
Pricing
DaVinci publishes its prices, which already puts it ahead of several competitors in this category. Buzzy AI, the closest structural rival we have reviewed, publishes none at all. The subscription is 1.99 dollars for a seven day trial, then 29.99 dollars a month unless you cancel, or 19.99 dollars a month billed quarterly. Generation consumes credits, and when the subscription allowance runs out you buy top-up packs.
| Credit pack | Price | Per credit | Advertised as |
|---|---|---|---|
| 25,000 | 149.99 (from 299.99) | 0.0060 | up to 12,500 images or 625 videos |
| 10,000 | 69.98 (from 139.99) | 0.0070 | up to 5,000 images or 250 videos |
| 5,000 | 39.99 (from 79.98) | 0.0080 | up to 2,500 images or 125 videos |
| 2,000 | 15.99 (from 31.99) | 0.0080 | up to 1,000 images or 50 videos |
| 1,000 | 8.99 (from 17.98) | 0.0090 | up to 500 images or 25 videos |
All figures in US dollars, read from the purchase modal in September 2026. Every pack is displayed as a discount against a higher struck-through price.
How to budget credits correctly
The pack descriptions and the model picker use different bases, and it is worth understanding which one to plan from. Those "up to" figures work out to two credits an image and 40 credits a video. The model picker, one screen away, prices every model with its own "from" figure. Those are the numbers your actual spend follows.
| What the modal implies | What the model picker charges | Result |
|---|---|---|
| 2 credits per image, so 12,500 images from 25,000 credits | The cheapest image model we could find is Flux 2 Turbo at 9 credits, and "from" means that rises with resolution and options | 25,000 divided by 9 is 2,777 images, not 12,500 |
| 40 credits per video, so 625 videos from 25,000 credits | Reachable only on the cheapest model at the shortest duration and lowest resolution, which is exactly what the video composer defaults to: Seedance 2.0 Fast, 4 seconds, 480p | Technically possible at the floor, unrepresentative of any real use |
Neither figure is a claim about our own usage. Both come from numbers DaVinci publishes itself, on adjacent screens, and anyone can check them in a minute. "Up to" is hedged language and the video figure is reachable at the floor, on the cheapest model at the shortest duration. The image figure assumes a two credit model, and the cheapest one on the list is nine, so plan from nine upward.
The practical takeaway: budget from the model picker, not from the pack description. Pick the models you actually intend to use, read their "from" figures, and treat those as floors that climb with duration and resolution. This is a clarity gap rather than a pricing problem, and one line on the purchase screen naming the assumed model and settings would close it. We were given test credits rather than buying them, so we cannot tell you how a paid allowance feels day to day, only how to read the published numbers.
What we tested, and what we did not
Reviews of platforms this wide often imply a score covers everything. This one does not, so here is the boundary in plain terms.
Tested: video generation, using a nine-part protocol covering human realism and motion, complex prompt adherence, character consistency, camera control, physics, spatial reasoning, start and end frame control, audio and dialogue, and a 30 second narrative. Twelve generations across five models, analyzed frame by frame and, for the dialogue test, by audio spectrogram.
Not tested: the Artwork generator, voiceover, music, the Enhancer, the Editor, Motion Control, and the reference and start-and-end-frame system. Two of our runs were meant to exercise the reference system and both were generated without images attached, so it remains unmeasured. Image generation was touched twice, not benchmarked.
That means one of seven surfaces was measured properly. The score below weights what we verified and gives no credit, in either direction, for what we did not.
Test results by model
Because DaVinci resells other companies' engines, a single score for "DaVinci video" would be meaningless. Each model was scored on its own. Wan 3.0 received the fullest run.
| Model | Tests run | Score | What it is good at, and where it breaks |
|---|---|---|---|
| Wan 3.0 | 5 of 9 | 52/65 | The strongest all-rounder we ran. Held four correctly dressed people for a full 15 seconds with the right person performing the right timed action, kept a single walking figure coherent for 30 seconds with correct five-finger grips on two different objects, and held four faces across four shot changes. Breaks on camera language: it executes orbit and close-up on X, ignores locked off, and reads tilt up as move along. Weakest on physics, where a lemon grew past the width of the glass it had fallen into. |
| Seedance 2.5 | 1 of 9 | 4/5 | Given a 30 second script with six timed beats, it delivered all six in order, kept four characters correctly cast and dressed, and cut to black on cue. It missed one specific instruction, a clock that should have read 2:17 rendered as 12:00, and dropped characters in the final shot. |
| MiniMax H3 | 1 of 9 | 7/10 | Produced the only 2K output in the whole test. On the dialogue scene it placed exactly two speech events, attributed them to the right speakers, and the man's mouth was closed during silence and articulating during speech. The requested rain and room ambience were largely absent, and the second line runs to the last frame and is probably clipped. |
| Kling V3 Turbo | 1 of 9 | 6.5/10 | Held three chair colors and a locked camera cleanly for 10 seconds with convincing animal movement, but did not put the cat behind the red chair as instructed and rendered six chairs where three were specified. Separately, in an unscored run it produced a lead character closely resembling a well known actor in a scene reminiscent of a specific prestige series, which is a rights consideration for commercial work. |
Scores are out of the points each test carries in our video protocol, not out of 100. No model completed the full nine-test suite, so no model here has a publishable total. Full method and per-clip findings are in our test log.
Where DaVinci AI is the strongest pick
DaVinci is at its strongest for someone who has not settled on a model and does not want to find out the hard way. Running the same brief through Wan 3.0, Seedance 2.5, MiniMax H3 and Kling from one balance is genuinely faster and cheaper than holding four subscriptions, and our own testing is the argument: the four models diverged sharply and the right choice depended entirely on the job. If you need people and timed action, Wan 3.0. If you need a multi-beat narrative, Seedance 2.5. If you need dialogue and resolution, MiniMax H3. Nobody discovers that from marketing pages.
The second argument is scope. Image, artwork, video, voiceover, music, upscaling and editing behind one login is a different proposition from a model router, and for a solo creator or small team assembling short-form content it removes a genuine amount of tool sprawl.
Best for
Solo creators, short-form video producers and small marketing teams who want breadth across models and media types rather than depth in one, and who are comfortable working out real costs from a model price list rather than a headline. Not for teams that need predictable per-asset budgeting from the marketing copy, or that have already standardized on a single model and would be paying for breadth they will not use.
Key features
- Roughly 28 image models behind one login, including GPT Image 2, Nano Banana Pro, Flux 2, Ideogram 4.0, Recraft V4.1 Pro, Seedream 5.0 Pro, Adobe Firefly and Qwen Image 3
- Six video models: Veo, Kling V3 Turbo, Seedance 2.5, Seedance 2.0 Fast, MiniMax H3 and Wan 3.0
- Per-generation controls for model, aspect ratio, resolution, duration and sound, all stored and visible on every past generation
- Start image and end image slots, plus separate reference slots for image, character and clothing
- Edit Video and Motion Control tabs alongside Create Video
- Artwork generator driven by named moodboard presets such as Classical Surreal Collage and Oneiric Ink Surrealism
- Voiceover library with categories, named voices and multiple languages, seen in English, Spanish and French
- Music generator with genre presets, auto lyrics and an inline player
- Enhancer, which the vendor says upscales to 22K, and an Editor for images and video, both capped at 75MB
- Reuse action that reloads any past generation's prompt and settings
Best-fit use cases
- Running one brief through four video models to find out which handles your subject before committing a campaign to it
- Producing short-form social video where the model that wins changes shot to shot
- Replacing several separate image, voice and music subscriptions for a solo creator or two-person team
- Prototyping with expensive frontier image models without holding a subscription to each
Pros
- Genuine breadth: roughly 28 image models and six video models, plus voice, music, artwork, upscaling and editing, on one balance
- Prices are published, unlike several competing aggregators
- Credits are refunded when a provider rejects a prompt, which we confirmed on a fictional horror brief that tripped a filter
- Every generation stores and displays its model, aspect, resolution and duration, which makes results reproducible and comparisons honest
- Wan 3.0 through this platform was the strongest model we tested on people, wardrobe and timed action
- Resolution runs from 480p to 2K depending on model, and the Reuse action makes iteration quick
Cons
- Credit pack descriptions are floor figures rather than capacity: the up to 12,500 images headline works out to two credits an image, while the cheapest image model on the platform starts at nine
- Duration and resolution written into a prompt are silently ignored; only the interface controls apply, so a 10 second script becomes a 15 or 30 second clip without warning
- The seven day trial at 1.99 dollars converts to 29.99 dollars a month unless cancelled
- The upgrade modal uses a live counter of how many people took the offer today, and every credit pack is framed as a discount off a higher price
- Camera direction is poorly followed by the strongest video model, so anyone storyboarding specific moves will fight it
- Provider moderation rejections do not explain which rule was hit, only that the generation failed
What to watch out for
Three things, in order of how much money they cost you.
Budget from the model picker. The pack descriptions are floor-of-the-floor figures. Open the model list, note the "from" credit cost of the specific models you plan to use, and treat those as minimums that climb with duration and resolution. A reader who budgets from "625 videos" and then works at 720p for 15 seconds will be surprised.
Set duration and resolution in the interface, not the prompt. We wrote explicit durations and 1080p into every test prompt and none of it applied. The controls next to the Generate button are the only ones that count. This is a reasonable design, because each model has its own fixed output specification, but nothing tells you that, and a prompt that says 10 seconds returning a 30 second clip looks like a bug rather than a setting.
Read every "up to" figure as a ceiling, not a norm. The Enhancer advertises upscaling to 22K. We did not test it, and since the credit descriptions are written against best-case settings, we are reporting that number as a vendor claim rather than as a verified specification.
Is DaVinci AI worth it?
Yes, for the buyer it is actually built for, and our own testing is the strongest argument. Four video models given the same briefs produced sharply different results, and the right one changed with the job: Wan 3.0 for people and timed action, Seedance 2.5 for multi-beat narrative, MiniMax H3 for dialogue and resolution. Discovering that across four separate subscriptions would cost far more than 29.99 dollars a month, and the platform layers voice, music, artwork, upscaling and editing on top of it.
It scores 4.1 rather than higher for one reason, and it is not output quality. Publishing prices at all puts DaVinci well ahead of aggregators that hide them, and refunding credits when a provider rejects a prompt is a real consumer protection that we verified rather than took on trust. What holds the score down is that the credit pack descriptions are written against the cheapest possible model and settings, so the headline capacity is roughly four and a half times what a buyer working at normal settings will get. The product is not overpriced; the purchase screen is just optimistic, and that is the screen people decide on.
Worth saying plainly: the reference and start-and-end-frame system, which is the feature most likely to matter for anyone doing character-consistent work, is untested here because both of our attempts ran without images attached. That is our gap, not the vendor's, and the score can move in either direction once someone measures it.
How we arrived at 4.1
The platform score is separate from the model scores above, because the models are not DaVinci's work. It is weighted across ten categories.
| Category | Weight | Score | Reasoning |
|---|---|---|---|
| Model selection | 15 | 15 | Roughly 28 image models, six video models, plus voice, music, artwork, upscaling and editing |
| Video generation controls | 15 | 12 | Full control set, but prompt-level duration and resolution are silently inert |
| Image and reference support | 10 | 7 | Style, character, clothing and object slots exist and look well designed, but we did not exercise them |
| Start and end frame control | 10 | 6 | Present, unverified, both attempts ran without images attached |
| Generation speed | 10 | 7 | Under a minute for short clips, over a minute at 30 seconds, one notable outlier, not precisely timed |
| Credit value | 10 | 4 | Pack descriptions are written against best-case settings, so headline capacity runs well ahead of normal use |
| Output quality options | 5 | 5 | 480p to 2K per model, plus a dedicated upscaling surface |
| Interface and workflow | 10 | 8 | Clean, per-generation settings stored and visible, Reuse is a good iteration loop |
| Reliability | 10 | 8 | One late generation across twelve, nothing lost |
| Moderation transparency | 5 | 4 | Credits refunded on rejection, but no reason given for the rejection |
| Total | 100 | 76 | Twenty of those points sit on systems we did not exercise |
The 5 point score is a product rating rather than a linear map of that total, and it reflects a capable, unusually broad platform held back by one number on the purchase screen that sets the wrong expectation.
FAQ
Is DaVinci AI the same as DaVinci Resolve?
No, and they have nothing to do with each other. DaVinci Resolve is Blackmagic Design's professional video editing and color grading suite, a desktop application you install. DaVinci AI at davinci.ai is a browser-based generative platform that resells other companies' image, video, voice and music models from one subscription. The shared name is a coincidence and the two are not related products, companies or codebases.
How much does DaVinci AI cost?
The subscription runs 1.99 dollars for a seven day trial, then 29.99 dollars a month unless you cancel, or 19.99 dollars a month billed quarterly. Credits are consumed per generation and top-up packs are sold separately: 1,000 credits for 8.99 dollars, 2,000 for 15.99, 5,000 for 39.99, 10,000 for 69.98 and 25,000 for 149.99. That works out between 0.0060 and 0.0090 dollars a credit depending on pack size. Every pack is displayed as a discount against a higher struck-through price. Verified September 2026.
How many images or videos does 25,000 credits actually get you?
Fewer than the pack description suggests, so budget from the model picker instead. The purchase screen advertises 25,000 credits as up to 12,500 images or 625 videos, which works out to two credits an image and 40 a video. Those are best-case figures. DaVinci prices each model separately and the cheapest image model on the list, Flux 2 Turbo, starts at nine credits, with that figure rising as you raise resolution. At nine credits the same pack covers about 2,777 images. The video number is reachable on the cheapest model at the shortest duration and lowest resolution, which is what the composer defaults to. Pick the models you actually plan to use, read their from figures, and plan on those.
Which video model in DaVinci AI is best?
It depends on the job, and our testing separates them. Wan 3.0 was the strongest all-rounder, scoring 52 of 65 across five tests, and it is the one to pick for people, wardrobe, timed action and identity that has to hold over 15 to 30 seconds. Its weakness is camera language: it executes orbit and close-up on X but ignores locked off and reads tilt up as move along. Seedance 2.5 handled a six-beat 30 second narrative better than anything else we ran. MiniMax H3 produced the only 2K output in the test and got two-speaker dialogue timing and lip sync right. Kling V3 Turbo was the weakest of the four on our spatial reasoning test.
Does DaVinci AI refund credits if a prompt is rejected?
Yes. One of our fictional horror prompts was rejected at provider level despite being clearly fictional, and the credits were returned automatically. That is worth knowing because it is not universal on platforms that resell third-party models, and it means an experimental prompt that trips a provider filter does not cost you anything. What the platform does not surface is why a given provider rejected a prompt, so you are left guessing which clause you hit.
Pricing, model catalogue and feature claims verified September 2026 from davinci.ai, its model picker, its credit purchase modal and its subscription modal. Testing ran across twelve video generations on five models between September 2 and September 3, 2026, using credits supplied by the vendor. ToolChase is not an affiliate of DaVinci AI and received no payment for this review. Scores, rankings and list positions are never sold. See our methodology for how we score tools.
Quick info
1.99 for a 7 day trial, then 29.99/mo · 19.99/mo billed quarterly · credit packs from 8.99
No, the entry point is a paid trial
Roughly 28 image models and six video models, plus voice, music and editing, on one balance
Credit pack capacity claims overstate what the platform's own model prices deliver
Wan 3.0 for people and timed action, Seedance 2.5 for narrative, MiniMax H3 for dialogue
Try DaVinci AI
Visit DaVinci AI →Is DaVinci AI the right pick for your use case?
Answer five questions and get a ranked shortlist for the job you actually have, scored against verified feature grades, with the reasoning shown for every tool. No signup. Sponsorship never changes the order.
Find my AI video stack →