Leonardo AI

By Minel Gunesoglu, founder. I put every number on this card together from dated public evidence, and read the community record behind it, before I let a score go live. Last updated 18 August 2026. This page is re-checked monthly, and the score moves when the evidence does.

Leonardo AI is the tool in this set with the largest consumer record behind it: 1,388 reviews when every profile in the category was re-read on 12 August 2026, broadly agreeing with each other, where the cards beside it run to twenty-one, or twenty-five, or a hundred and thirty-five. Until 12 August 2026 that depth was half of a scored row, and until 18 August it was still most of another. It is inside neither now, and this page has twice had to be rewritten around that rather than renumbered through it. What replaced it is the vendor's own paperwork, read end to end on that date, and the paperwork says something about Leonardo's free tier that no review of this tool appears to have printed.

One caveat before the card, because the shape of this page's evidence just changed. The usability row rests on 48 reviews and the capability row on an arena entry whose vote count was never recorded, so neither is deep; the contract row is not a sample at all, and its strength is a question of whether the document is current rather than of how many people answered. Read each row against the thing it is actually made of. How each dimension gets built is public.

Leonardo AI review score: 75.7/100, and what each row is actually made of

Leonardo AI's composite is 75.7/100 (3.8/5) as of the 18 August 2026 re-run, with the arena figure retrieved 23 July 2026, the entry price read 11 August and the terms read 12 August. What the rows rest on matters more than a one-word grade ever could. Here is the full card.

The row worth reading first is the bottom one, and it reads 56.2 where the same row read 71.0 a day earlier. Nothing about Leonardo changed. What changed is what the row is made of: it used to blend that Trustpilot figure with a single yes-or-no read of one clause. The blend is gone; the row is now contract text and nothing but. Two of those clauses are the whole story of the number. §8.3 gives paid subscribers ownership outright ("ownership of all Intellectual Property Rights in any Content you … create while using the Platform will vest in you upon creation"), which is the top of our scale. §8.7 does the opposite for free subscribers, and that is the 25 dragging the cell down. The indemnity question comes back empty in a way worth stating precisely, because our own record had it wrong: we had published no indemnification offered, which reads as a refusal. The document takes no position at all. A search of all 38,524 rendered characters, split into 187 paragraphs and tested one by one, returns zero matches for the word in either direction, no indemnity from Leonardo, and no obligation on you either. A contract silent on indemnity grants none, so the rubric scores it where the industry default lands, but silent and refused are different facts and this desk had been printing the wrong one.

Now the honest part, and this row has changed twice. Usability reads 82.8 against a raw 90.8, and Leonardo did not get harder to use either time. The figure comes from 48 reviews across G2 and Capterra: a real sample but not a deep one, so the raw rating is pulled part of the way back toward the middle of the scale. What the row lost is unearned confidence, not user satisfaction.

What moved on 18 August 2026, and what the checking cost. This row published 91.5 until that day, and the composite 77.7. Both came down (82.8 and 75.7) and the reason is that somebody opened a page nobody was obliged to open. Leonardo's Trustpilot profile had been raising this row: 4.6 out of 5 converts to 92.0 against the 90.8 those 48 business-software reviews support, on a sample nearly thirty times as large, so averaging the two by head count pulled the row up and held it there. The profile was opened on 18 August 2026 to establish how its reviews had been gathered rather than what they said. What came back was Trustpilot's own record of no invitation activity from this company in recent months, with each review on the visible profile tagged as one nobody had asked for. A sample nobody invited describes the people who chose to write, and it cannot stand in for everyone who paid, so it is no longer averaged with a sample gathered another way. It is reported in the section below, where it always did its real work.

Nothing forced that check. The record was lifting the number, the number was published, and every mechanical gate on this site was green. The cost of looking landed on the card with the most to lose from being looked at, which is the only reason this movement is worth a reader's trust: the same rule, run the same day, lifted five other cards. Capability at 75.4 comes from arena preference data, ELO 1119 for the Lucid Origin Ultra model, retrieved 23 July 2026. A beautiful sample and a one-in-four failure rate can sit inside the same product, and nothing measured here would tell them apart. Separating them takes a controlled battery, which this desk does not run, so we cannot tell you which Leonardo you would be buying. The scoring rules explain why we would rather ship a labelled gap than a comforting invention.

What Leonardo AI reviews actually measure: the 4.6 on Trustpilot and the support desk

Here is the finding that changed how I read Leonardo's headline rating. Its Trustpilot profile carries 4.6/5 over a sample of 1,388, counted on 12 August 2026, where 85% of the reviewers left five stars and 10% left one, a number most write-ups quote as proof the product is excellent. Read the actual reviews and a different picture forms: a disproportionate share of the five-star volume is people naming a specific support agent who fixed a billing or refund problem, not people praising the images.

The pattern is unmistakable once you see it, and one correction belongs inside it: until 12 August 2026 this paragraph opened with a review credited to a named German reviewer on 17 May 2026. Pages 7 to 10 of the profile with the all-languages filter cover that entire window, back to 2 May 2026, and neither the name nor the sentence is on any of them, so the quotation is withdrawn. The pattern it was illustrating is real and the profile carries plenty of it. "I'd like to sincerely thank Mark G for the outstanding support in resolving my membership issue on the site." (Trustpilot, Taraneh Kh, US, posted 17 May 2026.) "Mark Gomez was extremely professional, understanding, and helpful throughout the entire process." (Trustpilot, Franza Whyte, CA, posted 23 May 2026.) "Mark G. was incredible and took care of my problem promptly and with compassion." (Trustpilot, Hellena Banner, 7 May 2026.) A third-party analysis counted the same profile and drew a conclusion the pattern does not obviously suggest: “77 of the 90 positive reviews (86%) name a specific customer support agent. This is the highest agent-mention rate we have ever recorded and strongly indicates a review-solicitation process tied to support interactions.” (Rain AI Services, 100-review sample, published March 2026 and updated 27 March 2026 (this page called the post undated until 12 August 2026; it carries both dates on its own byline.) Read that carefully, because the figure cuts both ways. It is evidence that people who reach support are helped, and it is that analyst's reason for thinking the profile's positive reviews may be collected at the moment of that help rather than gathered independently. The same firm states its own limits), it did not test the product, billing system, content filters or cancellation process (so this is one reading of one sample, not a settled fact about how those reviews came to exist. A tension inside that is worth naming rather than resolving: the analyst reads the agent-mention rate as a sign that positive reviews are collected at the moment support helps somebody, while the profile itself, opened 18 August 2026, carries the platform's notice that the company has no recent history of asking for reviews. Those are not the same claim), the notice is about the platform's own invitation mechanism, and an agent suggesting a review at the end of a ticket is not that: so neither settles the other, and both are printed here.

This is not a scandal, and it is not fake reviews. It is a signal-composition problem, and it cuts in your favor to understand it. Where the review platforms actually grade the product (G2 and Capterra, whose reviewers are rating image quality and workflow rather than a support ticket), the scores sit a notch lower, in the mid-4s. So the honest read is: the support desk is genuinely good, good enough to lift the headline number, but if you are choosing Leonardo for its output, the 4.6 is partly measuring something you may never need.

Which brings up the thing this page has to say plainly rather than bury, and it begins with a correction to this page. Since 12 August 2026 this section has told you that the consumer record earns no number anywhere on our card. That was true of the row it was written about and false of the card as a whole: those 1,388 reviews were the larger half of the usability row until 18 August 2026, and they were holding it up rather than costing anything. As of that date the sentence is finally true. The record ships as prose with its sample size and its date attached (which is what this section is) and it buys the card nothing.

The mechanism this paragraph used to describe is gone as well, and it is worth retiring in writing rather than quietly. It said a consumer aggregate could clamp a composite from outside the table when a low rating sat on a deep sample. No such clamp exists; a record acts on a cell or it acts nowhere. What replaces it is simpler and harder to like: the deepest consumer evidence in this set sits outside every number on this card, because of how the sample was gathered rather than because of what it says. If you were choosing partly on that record, keep choosing on it: read it here, in words, rather than expecting to find it inside the score.

Did Canva buy Leonardo AI? The ownership behind the "declining every month" reviews

There is a countercurrent in the record that the 4.6 hides, and it is worth surfacing plainly. A dated, detailed Trustpilot review from a multi-year user frames Leonardo's 2026 service as a marked decline. It opens: "I've been using Leonardo AI for several years - and honestly, it's a huge disappointment now." Three paragraphs of specifics follow (glitches, unstable performance, token cost) and the review lands, in its own last two lines, on "It keeps getting worse and worse every month." and "Honestly - don't waste your money." (Trustpilot, Alex, GB, posted 27 April 2026, marked by the platform as an unprompted review.) Those are three separate sentences by one person, spaced out through one review, and they are printed that way here because until 12 August 2026 this page welded them into a single quotation with ellipses. Every word real, the shape not. Leonardo replied on 8 May 2026, thanking the reviewer and saying results vary by model, workflow, prompt and settings, rather than disputing the premise. One angry long-term user is an anecdote, not a trend, but the timing of the complaint is what makes it interesting, because it points at a corporate fact most reviews of this tool skip entirely.

Leonardo is owned by Canva. Canva's newsroom post announcing the acquisition is datelined 30 July 2024, and it frames the deal as preserving the platform while folding Leonardo's Phoenix foundational model into Canva's own suite. The sentence is written in Canva's first person and is quoted here as Canva wrote it: "we also plan to rapidly integrate Leonardo's leading technology and their Phoenix foundational model into our existing suite of Magic Studio products." (Canva Newsroom, 30 July 2024.) Until 12 August 2026 this page printed that sentence rewritten into the third person and dated a day early; both are corrected. That single fact reshapes the trust picture in a way no VC-funded rival's does, and it cuts both ways, which is why nobody else seems to write about it honestly.

On the reassuring side: a genuinely free tier is expensive to run, and a standalone startup burning venture money has every incentive to quietly gut it. A tool owned by a profitable design platform can subsidize free generations as a funnel into Canva's paid suite, which is the structural reason Leonardo can afford the Apprentice plan below while a subscription-only rival like Midjourney offers no free tier at all. On the worrying side: acquisitions redirect roadmaps, and the "declining every month" complaint is exactly the kind of thing that surfaces when engineering attention shifts from a standalone product toward integration work elsewhere. I cannot prove that connection from a single review, and I will not pretend to. What I can say is that the ownership is real, dated, and material to anyone betting a workflow on this tool for the next two years, and that it belongs in a review of Leonardo, not buried in a "did Canva buy Leonardo AI" search result.

Leonardo AI hand errors: the weakness the vendor admits, and the test I still owe you

Every review of an image generator eventually reaches hands, and Leonardo's are a known weak point: not a fringe gripe but a company-acknowledged limitation. Leonardo's own support has replied to a Trustpilot complaint about extra and missing fingers by conceding the point: "Faces and hands are indeed challenging for image models." (Leonardo.Ai's own reply to a Trustpilot review, 13 October 2024. The same reply goes on to say the tools that fix faces and hands "are premium features and require a paid subscription".) When the vendor itself says the anatomy is hard, you do not need a rival's marketing to tell you it is a real limitation.

The record also says it may be narrowing. A Product Hunt reviewer describes the newer pipeline favorably: "The Phoneix [sic] model performs well without major tweaks and even while generating single images." (Product Hunt, Arpan Samuel Ramtek. The misspelling is the reviewer's; this page silently corrected it until 12 August 2026, which is a quotation reporting a text that was never written. Read with two limits: Product Hunt stamps this review only as "2yr ago", so the listing places it somewhere between roughly mid-2023 and mid-2024 and prints no exact date, and the platform files all four reviews of this product under "Founder Reviews", so they come from people who built something on it rather than from a general user sample.) Both things can be true at once: a baseline that has genuinely improved since the older SDXL-era output, with fine-anatomy edge cases that still break. That is the most defensible summary of where hands stand in mid-2026.

Now the part where most reviews fake it, and I won't. I have not run a fresh hands-on anatomy battery on Leonardo, so this page shows you no before-and-after hand test, because I don't have one yet, and a fabricated screenshot would be worth less than an honest gap. Competitors assert the weakness in prose ("monstrosities," one calls the output) but not one of them dates a real, reproducible example. Publishing an invented one to match them would defeat the entire point of this site. So the honest state is: the weakness is real and vendor-acknowledged, it appears to be improving, and the dated proof is a to-do I owe you, not a claim I'll invent.

The Apprentice tier is the real free plan Midjourney never gave you

This is Leonardo's clearest advantage over the biggest name in the category, and it is worth stating without hedging: Leonardo has a genuinely free tier, no card required, and Midjourney has none. Leonardo calls it the Apprentice plan (a daily token allowance you can use to actually generate images without paying, refreshed each day. The community treats this as a real differentiator rather than a trial-ware tease. "The free version is quite generous, and the paid version is worth the money if you need higher quality generations." (Capterra, Diana B., founder, computer software, 23 December 2024; the listing prints that date on the card and labels the review non-incentivized. This page previously dated it 2026 and said the listing showed no date), both wrong, corrected 12 August 2026.)

If "is Leonardo AI free" is the question that brought you here, the answer is yes, with a real daily allowance and no trial clock counting down to a paywall, which is the exact inverse of Midjourney, where the closest thing to free is a refund window that closes almost immediately. There is also a paid free trial on the subscription tiers for users who want to sample the premium models before committing, distinct from the always-on Apprentice allowance.

The friction is real too, and it is a value story, not a marketing one. The generous impression holds until you move past the default fast setting. Token consumption on the premium paths (Quality mode (the setting formerly branded Alchemy), video, and heavier models), is opaque, and it can drain a full day's allowance in a handful of generations. That ambiguous-token-pricing complaint (this page's own label for it, not a phrase anyone wrote), recurs across Trustpilot and Capterra, and it is why the free tier is a superb way to evaluate Leonardo and a frustrating way to do sustained serious work. The plan is real; the ceiling is lower than the word "generous" suggests once you turn the quality up.

And there is a second ceiling on the Apprentice plan that has nothing to do with tokens, which the terms read on 12 August 2026 state in a single clause. §8.7 covers free subscribers, and it assigns them out of the picture entirely: ownership of the output "will vest in us upon creation, and … you hereby assign your right, title and interest in the Output (if any) to us." No licence back to the free user appears in that clause. Set it beside §8.3, which vests ownership in paid subscribers on creation, and the free plan stops being a smaller version of the paid one and becomes a different arrangement: on the paid tier you own what you make, and on the free tier the document says Leonardo does. Nothing on this desk held that finding before 12 August 2026, and it is the clause that pulls the contract row down to 65.9. The practical consequence is narrow and sharp. Evaluate on Apprentice all you like; if a specific image is going to a client, into a product, or onto anything you intend to hold rights in, generate it on a paid tier, and check the current wording of §8.7 yourself before you rely on any of this: terms change without notice, and this one is a single paragraph away from meaning something else.

Leonardo AI pricing: what $12 a month and a single token actually buy

Leonardo lists a free Apprentice tier plus paid plans, and paid entry is $12/month on the Essential plan, a figure with a property none of its rivals here share. Leonardo's pricing page carries a yearly/monthly toggle whose own label advertises up to 20% off, and switching it leaves Essential at $12 either way; the yearly discount only reaches the tiers above it, moving Premium from $30 to $29 and Ultimate from $60 to $57 (leonardo.ai/pricing, SOLO tab, both toggle positions, read 11 August 2026). So this is the one entry price in the category where there is no monthly-versus-annual trap to fall into. But the number that governs your actual experience is not the monthly dollar figure: it is the token, and specifically how fast the quality settings eat it. On the default fast path, the daily free allowance stretches. Turn on Quality mode, upscaling, or video, and the same allowance can vanish in a few jobs, which is the recurring "opaque token pricing" theme from the section above wearing a price tag. Budget in tokens-per-usable-image on the settings you actually intend to use, not in dollars-per-month, or the plan will feel more expensive than it looks.

The other operational fact worth knowing before you pay is cancellation, because it generates the most consistent negative cluster in the record. Leonardo's stated policy is that prepaid Premium points expire on cancellation, including unused ones you already paid for. In the company's own words, replying to a Trustpilot review: "when purchasing Premium, your Premium points expire after cancellation." (Trustpilot, reviewer Alfred Stahl, Germany, 14 July 2025.) A smaller but recurring group reports being charged after clicking cancel, missing cancel buttons, or year-long Apple Pay charges from what they thought was a monthly signup. The support desk (the same one lifting the Trustpilot score), does make exceptions. A five-star review describes a six-month auto-billing problem with the cancel button missing from the subscription screen, and credits a named agent who made an exception "to process refunds for December 2025 and January 2026 charges - something their policy normally doesn't allow!" (Trustpilot, reviewer RaiQueen k, 7 February 2026; the exclamation mark is the reviewer's, and exception is the reviewer's own word, emphasised in the original). Until 12 August 2026 this paragraph reported that sentence without quoting it, on the stated ground that Trustpilot's public pagination ends at page 10 and a February review therefore sat outside what we could show. That was wrong about the source: the profile's own keyword search reaches past the pagination floor, and the review reads exactly as quoted here. That is genuinely good service. It is also a policy generous humans are compensating for. If you subscribe, spend your tokens before you cancel, not after, and set a renewal reminder the day you pay.

One more paper fact the consumer reviews bury, because it comes from a different crowd: Leonardo has a real API, and developers rate it well independently of the image-quality debate. "Leonardo's API is great, and I could integrate it within a day." (Product Hunt, Megham Garg, listed as "2 years ago" on 3 July 2026.) The praise there is narrow but consistent: fast integration and a fit for asset-production pipelines (batch iteration, background removal, upscaling) rather than one-off art. If you are evaluating Leonardo as a component rather than a canvas, that thread is the relevant one, and it is a lane Midjourney, with no public API, does not contest at all. Leonardo also runs a video generator and gets used for concrete jobs like interior-design mockups, a broader surface than the pure text-to-image tools, though this review scores the image side only and treats the rest as unmeasured.

Where Leonardo sits among the image generators we've scored

Leonardo is scored under the same rules as every tool in the image-generators category, where the full comparison table lives, and every composite below was rebuilt on 12 August 2026 from the same four dimensions; what travels with each one here is the clause or the sample it now turns on. The one-line differences against its neighbors:

• Midjourney, at 77.8/100, has the more recognizable aesthetic but no free tier, and its own consumer record (366 reviews at 1.6 out of 5), left its usability row on 18 August 2026 for the same reason Leonardo's 1,388 left this one, which is why that card rose nine points on the day this one fell two. On output rights the two documents agree with each other on the paid tier: both vest ownership in the subscriber.
• Adobe Firefly, at 76.3/100 and now six tenths of a point clear of this card, is the pick when a team wants a rights position it can quote, its own terms permit commercial use of non-beta free-plan output, which is the exact question Leonardo's §8.7 answers the other way. On the indemnity itself Firefly's consumer document runs one way, user toward vendor; Leonardo's says nothing in either direction.
• Krea, at 76.3/100, is built around real-time generation and editing, and its slim lead sits on a card whose paid-tier ownership clause our reading could not resolve at all, against the plain assignment in Leonardo's §8.3.

That last comparison is the one to sit with, because it is what is left of this page's old thesis once the consumer record stopped scoring. Krea is 1.1 points above Leonardo. In June the published gap was 6.2, then 2.2, now 1.1, and not one point of that closing came from either product doing anything. It came from a missing sample size being read, and then from two legal documents being opened. A number this close is not a ranking of the tools; it is a reminder that a composite is a summary of what we could measure, and the third column of your own shortlist is where the actual decision lives.

So, is Leonardo AI worth it? A conditional answer, dated 12 August 2026. If you want the easiest genuinely-free on-ramp among the major image tools, and can work within a daily token ceiling that tightens the moment you raise the quality, Leonardo is a strong and honest pick, with one condition the terms make non-negotiable: anything you intend to own gets generated on a paid tier, because §8.7 assigns free-tier output the other way. If you need a measured failure rate before you commit, that number does not exist here and I will not invent it; if you need dated proof that hands are fixed, that test is a to-do I owe you, not a claim I'll fake. And if the Canva-ownership trajectory worries you, the "declining every month" review is in the record for a reason: weigh it, don't dismiss it. For a deeper head-to-head, a Leonardo-versus-Midjourney breakdown and a Krea-versus-Leonardo comparison are their own pages, coming as the category fills out.