Gamma

By Minel Gunesoglu, founder of Vouch. The work behind this page was documentary. I read Gamma's Terms of Use in a rendered browser and searched it for the sentence that assigns ownership of generated material, opened the pricing page in both billing views because it does not load on the one most people assume, hand-classified the twenty most recent Trustpilot reviews after a keyword classifier got the split wrong, and traced the one independent benchmark that seats this product back to the paper and the live board behind it. No deck of mine sits under any figure below, and no session with this product was run for this page. Published 18 August 2026, re-checked monthly.

Fifteen of the twenty most recent Trustpilot reviews of Gamma carry one star. Read by hand rather than counted by keyword, nine of those fifteen are about the slides (sameness, an editor that will not take a narrow instruction, layouts that will not reflow) and five are about billing. Separately, six researchers at Tsinghua University published a slide-generation benchmark in March 2026 whose per-dimension table gives Gamma 66.9 for presentation fundamentals, 54.1 for content fidelity, and 22.6 for the row headed Visual Design and Layout.

Two records, produced by completely different methods, neither one referring to the other, landing on the same part of the same product. That convergence is what this page is built around, and it is worth saying immediately what it is not. A position on a nine-system board is not a verdict, twenty reviews are one day's slice of 121, and neither of them touches the half of this card that is genuinely strong: Gamma is cheap, the free tier is a real one, and value is the highest row on the card at 80.0.

Gamma Tech, Inc. sells a prompt-to-deck tool that also produces documents, websites and social posts. Its card here reads 73.2 out of 100 on evidence collected 18 August 2026, with every row filled and the four rows sitting within 11.2 points of each other. Everything below names the document or the board each number came off, and the day it was opened.

What Gamma Is, and Which of the Products Called Gamma This Review Scores

Gamma is a generator you brief in a sentence. You describe a deck, paste text, or import a PDF or PowerPoint file, and it returns finished slides you can then edit, present from a link, or export. Its G2 profile, read in a rendered browser on 18 August 2026, opens its description with "Gamma AI is an advanced AI-powered content creation platform that enables users to effortlessly generate professional presentations, documents, websites, and social media content." The company was founded in 2020 in San Francisco. TechCrunch reported on 10 November 2025 a $68M Series B at a $2.1B valuation led by Andreessen Horowitz, following a $12M Series A led by Accel in 2024. The revenue and user figures that travel with that round (an ARR figure and a user count), are attributed by that outlet to the chief executive's own posts rather than to anything it verified, so they are recorded here as claims with a speaker attached and are not used as measurements anywhere on this page.

Three different things carry this name, and all three were checked before a single figure entered this card, because two of them have live web properties that a careless read would file as this product's.

That last row is the reason the pricing section below carries a retrieval date and a source in the same breath as the numbers: the lookalike ranks on searches for this product's pricing and publishes prices that are not these. The identity of the G2 listing needed settling for the same reason, since its URL slug reads like the security company's name. Three things on the page itself settle it: the page's own machine-readable breadcrumb files it under AI Presentation Tools, the description quoted above is unmistakably about slides, and the profile's overview credit names Gamma's co-founder. There is a Wikipedia article for the app as well, used here only to cross-check dates.

One scope line matters more than any of that. Gamma also builds websites, and the website output is a first-class product rather than a side feature. This card scores the presentation tool. Two of the one-star reviews in the sample read for this page are about the site builder (one about a noindex tag, one about internal links) and neither is counted as evidence about slide generation anywhere below, because they measure a different surface. If a hosted site is the thing you are shopping for, the website builders scored on this site are the cards to read instead.

Is Gamma AI Free? 400 Credits at Signup, and They Do Not Refresh

Yes, and the free plan is more usable than most in this category, with one structural catch that decides whether it is an evaluation tier or a permanent one.

Gamma's pricing page, read 18 August 2026, lists the Free plan at $0 with no card required and itemises it: "400 credits at signup", "Create up to 10 slides per prompt", import from PDF and PPTX, and "Export to PDF, PPTX, PNG & Google Slides". That export line is the unusual part. Plenty of free tiers hold the file formats back; this one does not, so a free account can produce a PowerPoint file and take it away. What it cannot do is remove the branding: the vendor's export article states the Made with Gamma badge appears on presentations exported from free accounts and that a paid plan removes it.

The catch is what happens after those 400 credits. Gamma's credits article states that on the free plan "all AI actions use credits, and credits do not refresh", and that once the initial allocation is spent a free user must earn more through referrals or upgrade. Referrals pay 200 credits each, and a free account can hold up to 2,000 credits at a time. So the free tier is a one-time allowance rather than a monthly one, which is a different shape from almost every subscription a buyer will have met, and it produces a specific moment: the product stops working for somebody who has never been charged anything. That mechanism is worth holding in mind when reading the consumer record further down, without treating it as an explanation for any particular review.

Credits are metered per slide and per image, and the rate moves. The same article states plainly that consumption rates and the actions that consume credits are subject to change, that longer generations use more, and that more advanced image models cost more than standard ones, with the cost shown before you generate. A separate article dated 24 October 2025 gives a ceiling of 50,000 tokens per initial generation on Free, against 100,000 on Plus and Pro. None of that is hidden, and all of it means the honest answer to how many decks 400 credits buys is that it depends on what you ask for.

How Gamma AI Credits Work on a Paid Plan, and What Cancelling Costs You

Paid plans behave conventionally where the free plan does not. Credits refill monthly on both monthly and annual billing (1,000 on Plus, 4,000 on Pro, 20,000 on Ultra) and unused credits roll over up to double the plan size, so a Plus subscriber can hold 2,000 at once. Running out does not lock the account: the vendor states you keep every non-AI feature and can still edit by hand. Additional credits can be bought on a paid plan.

Cancelling has a consequence the pricing page does not mention and the credits article does. You keep your plan and its credits to the end of the billing cycle, and then, in the vendor's words, "any remaining credits above 400 will be removed". A subscriber who cancels mid-year with a rolled-over balance loses the balance rather than carrying it into the free tier. That is stated in the vendor's own documentation on a page anybody can open, and it belongs in a buying decision because the rollover allowance is otherwise an argument for paying annually.

What Gamma Costs a Month: $12 for Plus, and the Toggle That Shows $9 First

The first thing to establish about Gamma's price list is which price you are looking at. The pricing page loads with "Pay annually" already selected and a banner above the switch reading "Save up to 28% with annual billing". Nothing is concealed (the switch is visible and labelled) but it means the figures a visitor meets first are the annual ones, and a shopper comparing this product against another in a browser tab is holding a discounted committed rate against whatever the other page happens to default to. Both views were opened in a rendered browser on 18 August 2026 and captured separately.

Every row above was read off gamma.app/en/pricing on 18 August 2026. The value cell on this card scores the monthly figure, $12, because that is what the product costs without committing to a year, and because a monthly basis is the one a tool with no annual option could still be measured on. One transcription note, because it is a thousandfold error waiting to happen: the annual Ultra total renders on the page as $1.080 per seat, with a full stop where an English-locale page would put a comma, while being served from an English URL. It is $1,080.

Two things are not on this table and should not be inferred onto it. The pricing page carries a second tab for business plans that these captures do not cover, and the vendor's help centre gives Team at $240 per seat with a two-seat minimum and Business at $480 per seat with a ten-seat minimum in a comparison table that never states the billing period. A neighbouring article says Team plans are annual only, which makes an annual reading likely, and likely is not read, so those two figures are reported here without a period attached rather than guessed at. The pricing page also carries the line "We're a SOC 2 Type II compliant organization." with a link to a trust centre, which is a compliance statement rather than anything this card measures.

At $12 a month for a seat with a thousand credits, Gamma sits at the cheap end of the software this desk has priced, and the value row reflects that. It is the highest number on the card, though it now leads the next row by six points rather than by thirty, and it is also the row a reader should trust least over time: prices move, and this one carries the date it was read for exactly that reason.

Gamma Export to PowerPoint: the Formats, and the Fidelity Limits the Vendor Documents

Gamma exports to PDF, PNG and PowerPoint (PPTX), posts directly to LinkedIn, and reaches Google Slides by way of a PPTX upload rather than a direct route. Individual slides can be exported on their own. All of it is available on the free plan, with the watermark attached.

What makes this section worth reading is that the vendor documents its own export limits in unusual detail, in an article titled around the question users ask. Exports render Present Mode rather than Edit Mode, so some difference from the editor is expected by design. Beyond that baseline, the vendor's own troubleshooting article states that "Extremely long gammas (typically 150+ slides) can fail to export.", and advises splitting the deck, exporting the pieces and merging them outside Gamma. Heavy image content can also fail or render incorrectly. Its sibling article on export formats adds that "Gradient Headings: May render differently or default to a single colour in export" and that frosted backgrounds may fall back to substitute styles, attributing both to how PowerPoint and PDF readers handle those effects.

A vendor writing its own limits down is a stronger form of evidence than a pile of user reports, and it needs no adjective from me. It also settles one question that otherwise generates a lot of noise. On fonts, the same article says: "Google Slides ignores fonts embedded in an uploaded PowerPoint file and substitutes its own font, even though Gamma embeds fonts correctly on export. This is a Google Slides limitation, not a Gamma export issue." That is a property of the PowerPoint-into-Slides path generally rather than anything specific to this product, and a buyer who lands on it will meet it with any tool that exports PPTX.

The practical reading is about what your deliverable is. If you present from a link or hand over a PDF, none of this touches you. If the deliverable is a PowerPoint file that somebody else opens, edits and re-styles, the editor view is not what they will receive, and the gap is documented rather than disputed. One developer put the same point from the other end in the vendor's own Hacker News thread on 24 September 2025: "the most important thing for me is the system of record which is the CMS. and the secondary system of record: Canva, Google Slides, Figma. you don't zero shot a deck." That is one commenter's view of where a generated deck sits in a workflow, not a measurement of anything, and it is included because the same handoff shows up in the review record below.

What Gamma Reviews Actually Say: 1.7 on Trustpilot Across 121, 3.9 on G2 Across 29

Two consumer platforms carry enough reviews of this product to be worth converting, and they disagree by more than two stars. Trustpilot shows 1.7 out of 5 across 121 reviews, with the platform's own band word printed beside it: "Bad". G2 shows 3.9 out of 5 across 29. Both were read in a rendered browser on 18 August 2026, because all three of the big review platforms refuse automated retrieval and a search-result snippet is not a reading. The Trustpilot page is marked "Unclaimed profile" and carries no vendor replies. That badge is recorded here as a fact about the profile rather than used as evidence about the reviews on it: profiles this desk opened on 18 August 2026 that do carry the claim badge run from the bottom of the scale to the top, so the badge does not tell a reader how a sample was gathered.

That page also carries a summary block at the top which the platform labels as machine-generated from a subset of the reviews. Nothing from it is quoted anywhere on this page: it is a summary of users rather than any user's words, and attributing it to a person would be inventing a reviewer.

Why Gamma's G2 and Trustpilot Review Scores Disagree by More Than Two Stars

Neither number is reconciled into the other here, and neither is thrown away. What can be said from the records themselves is narrower and firmer than an average: the two populations are not the same people arriving for the same reason. G2's audience reaches a listing through a software-selection process and writes about a tool they chose for work. Trustpilot's audience is a general consumer one that typically arrives after something specific happened, a charge, a support ticket, a product that stopped doing what they wanted. A four-to-one difference in sample size sits on top of that.

Until this morning the card averaged the two by how many people stood behind each, which is why the usability row had read 47.1 and why the larger, unhappier sample was the heavier half by a long way. It no longer does, and the reason is not a ruling about which platform is right. Averaging two ratings by head count asserts that both samples were gathered the same way and that one more review on either side means the same thing on both. These two are not recorded as the same kind of sample: our own record files the 121 Trustpilot reviews as self-selected, written by people who arrived and posted on their own, while how G2's 29 were gathered has never been read. So the card reports them separately instead of blending them, the usability row is built from the G2 reading, and Trustpilot's 1.7 across 121 is published beside it with its own count and date.

Being self-selected does not make a rating low, and nothing here should be read as saying it does: profiles carrying that same description that this desk opened on 18 August 2026 run from 1.6 out of 5 to 4.6 out of 5. What the description costs is representativeness. A sample of people who came on their own describes the people who came, not everybody who paid.

That change moved the row from 47.1 to 73.9, and the move costs something a reader should be told rather than left to work out. The row now stands on 29 readings where it drew on 150 before, so it went up and the evidence beneath it got narrower in the same step; the card marks that cell thin for exactly this reason, and thin is the honest word for it. Nothing was thrown away. Both source figures are still printed with their counts, so a reader who thinks the business-software audience is the relevant one can weigh 3.9 across 29, a reader who thinks the consumer record is the more honest one can weigh 1.7 across 121, and neither has to take 73.9 on trust.

What Gamma's One-Star Reviewers Report: 9 of 15 Are About the Output, Not the Bill

The twenty most recent reviews on that Trustpilot profile were read on 18 August 2026, sorted by recency and dated from the page's own machine-readable timestamps rather than from the relative stamps the interface prints. Fifteen carry one star, two carry two, one carries three, two carry five. Hand-classifying the fifteen one-star reviews: nine concern the product, five concern billing, one concerns support.

That hand count matters, and the reason is a correction to my own method rather than a finding about the vendor. A keyword classifier looking for cancel, refund, charge, billing, subscription, trial, card, renew and money returned nine billing complaints here. Reading them returned five. The classifier had counted two reviews that complain about the output and happen to mention the price, and a regular expression cannot tell what a sentence is about. The published split is the hand count.

What the nine actually say clusters tightly. "Every presentation looks the same", wrote one reviewer on 27 June 2026. "Every output looks the same; trying to get anything A4 is impossible; it just crams everything into the top 5th of each page and then blanks the rest.", wrote another on 8 June 2026. On the editor: "The initial design you get is not bad, but it is impossible to make changes. The AI agent simply does not listen." (14 June 2026), "I asked them to only change one slide, and they changed all" (4 May 2026), and "The AI doesn't understand simple instructions" (5 July 2026). A reviewer on 18 April 2026 summarised the experience as "too slow, heavy to adjust, unpredictable, and volatile."

The five billing reviews are a smaller and more familiar cluster: renewal dates, charges that arrived after a subscription had been forgotten. One from 16 June 2026 reads: "A year later, I completely forgot about the subscription until I was charged." That shape appears across subscription software generally rather than being particular to this vendor, and the refund window documented further down is the part of it a buyer can actually check in advance.

The sample also contains people who disagree with all of it, and leaving them out would be reporting a selection rather than a page. A five-star reviewer on 12 May 2026 wrote: "idk why are there too many negative reviews I personally had a very very good experience, the users probably were too lazy to edit or do some prompt engineering". That is worth quoting whole, because it names the same axis the critics do and draws the opposite conclusion from it: how much editing the first draft needs, and whose job that is.

Three platforms that would have widened this record could not be opened at all on 18 August 2026. Capterra returned 403 to every method tried, and its listing identity is unconfirmed against the page itself, which matters because Capterra carries more than one listing under this name. Product Hunt returned 403. Every Reddit path returned a wall, including three that returned a plausible-looking 200 carrying a login interstitial rather than the content. No login was attempted. Reddit is the largest missing surface here and the one where an unmoderated critical view of this product would live, so the community record on this page is short by exactly that much, and saying so is more useful than paraphrasing threads nobody read.

How Gamma Scores 73.2 out of 100, and Which Row Holds It Down

Four rows, four sources, one collection date. The interesting thing about this card is not that it is full. It is that four rows built from four unrelated kinds of evidence have landed within 11.2 points of each other while still answering completely different questions. That closeness is a fact about this product on this date, not a property of the method, and it is new: as recently as this morning the same four rows spanned 32.9 points.

Multiply those four and take the fourth root and 73.2 falls out. That is worth thirty seconds with a calculator, and what the exercise shows changed today: add the same four and divide by four and the answer is 73.3, a tenth of a point away and in the same band. Multiplying rather than adding is what makes a weak row expensive, and this card no longer has a row weak enough for the choice between the two to matter. This morning it did, because the usability row stood at 47.1 and dragged the total 1.2 points below a flat average. The row holding the total down now is capability at 68.8, which is a benchmark position rather than anything a customer said. A buyer choosing on price and a buyer choosing on how much editing a draft will need are still reading opposite ends of this card; they are no longer reading opposite ends of the scale.

What moved on 18 August 2026, and why. This page went up this morning at 65.4, and its usability row had read 47.1; both figures came out of averaging the two consumer records by head count. The model stopped doing that today. Consumer records are now averaged into one cell only where the way each sample was gathered is recorded as the same; where the records disagree on that, the cell reads one of them and the other is reported beside it with its own count. Seven of the site's twenty-eight cards moved on the change: five rose and two fell, one fell because a record that had been lifting its cell turned out to be self-selected and left the pool, and one because a single sample had been counted twice on the same card. Gamma's is the widest of the seven moves, and that follows from its two records having been the furthest apart to begin with. Nothing Gamma did between yesterday and today is inside this number.

Capability here rests on a published benchmark rather than on a blind-vote arena, and the difference is worth one sentence because it is not like-for-like with every other card on this site. Where a category has an arena (thousands of people picking between two outputs with the names hidden), that is what the row converts. No such arena covers slide generation, so this row converts the closest independent thing that does exist: a rubric-based academic benchmark that scores commercial products by one method. Both answer the same buyer question, which is how good the output is, and they are not the same kind of evidence. What a row on one of these cards is allowed to rest on is the same rule for every tool measured here.

Gamma's Capability Score: 49.2 on PresentBench, Seventh of Nine, and What That Converts To

PresentBench is a slide-generation benchmark from six researchers at Tsinghua University, published as arXiv:2603.07244 on 7 March 2026 with a live leaderboard beside it. It clears the bar most sources in this category do not, for a reason that is checkable in one line: the authors ship no competing slide-generation product, and the top-scoring entry is not theirs. The paper says it evaluates "widely used slide generation products currently available on the market", and its reference list cites Gamma to gamma.app, which is how this figure was tied to this product before it was used.

The task is heavy. Across 238 instances spread over academia, advertising, economics, education and talks, each system is handed an average of 22.2 thousand input tokens and 34 pages of source material and asked to produce a deck at default settings with no further interaction. Each output is checked against instance-specific binary checklist items (54.1 of them per instance on average), across five dimensions, and the Total is the mean of those pass rates.

Gamma's Total is 49.2, seventh of the nine rows on the board as it stood on 18 August 2026. The field runs from 35.9 to 70.8 with its middle at 53.6. That spread is why the card reads 68.8 rather than 49.2: a mean of pass rates on this benchmark looks like a percentage out of a hundred, and in this field it is not one. The board's best row reaches 70.8, and the paper says as much in its own voice about its narrower eight-system field. "Even the best-performing system only reaches 62.5 overall, indicating that grounded, end-to-end slide authoring remains far from solved". Treating 49.2 as a score out of a hundred would spend the top thirty points of the scale on ground nobody occupies, so what the row converts is where Gamma sits inside the distribution the benchmark itself published.

Four caveats travel with that number, all of them from the source's own text. The checklist items are verified by a multimodal language model, and the paper's limitations section says the evaluation "relies on multimodal LLMs as verifiers, and their capability may limit the reliability of the resulting scores". Its agreement with human rankings is reported honestly and is moderate: Spearman 0.532 against a human-agreement ceiling of 0.664, better than the two baselines it is measured against, on a 24-sample user study. Gamma's row carries a "Tested at" stamp of 2026-01, seven months before this reading, and these products ship continuously: whether anything changed since is an open question rather than an answered one, since the vendor's blog refused retrieval. And the board accepts submissions: it has already gained a row above every original entry, so it is re-read on a schedule rather than treated as fixed.

Where Gamma's Benchmark Row and Its Review Record Point at the Same Thing

The board publishes a second table breaking each system down by dimension, and Gamma's row is not flat. Presentation Fundamentals reads 66.9. Content Completeness 54.3, Content Correctness 47.7. Content Fidelity reads 54.1, which is the highest figure in that column among the eight rows tested in 2026-01, above NotebookLM's 45.1, the material that goes in comes out intact more faithfully than anything else measured in that group. Visual Design and Layout reads 22.6, second-lowest of the nine.

Set that beside the review record and the two are describing one thing from two directions. A benchmark built from 238 generation tasks, which references no review platform anywhere in it, gives this product its highest relative mark for keeping the source material intact and its lowest by a distance for how the result looks and is laid out. Trustpilot reviewers, writing on a platform that paper never mentions, say "Every presentation looks the same" and describe a page whose content is crammed into the top of the sheet. Neither of those is proof of the other. Together they are two independent records, produced seven months apart by completely different methods, landing on the same half of the product, and that is a firmer basis for a decision than either would be alone.

The honest boundary on this section: it says nothing about whether Gamma's output would look right to you, on your material, at your standard. Every figure above was read off a document or a board somebody else published; none of it came from putting a deck through Gamma and judging what came out.

Is Gamma Safe for Commercial Work? The Contract Defines Results and Never Says Who Owns Them

This row is a reading of paperwork. It reports the positions Gamma's contract takes on a stated date, not how the company behaves when a customer invokes them, and not what any customer thought of the experience. Gamma's Terms of Use Agreement, effective 9 September 2025, was read in full on 18 August 2026.

Indemnification runs one way. Section 10 opens "You agree to indemnify and hold Gamma, its parents, subsidiaries, affiliates, officers, employees, agents, partners, suppliers, and licensors" harmless from losses arising out of your content and your use of the services, and no clause anywhere in the document runs the other way. That is the ordinary position across the vendor contracts read for this site rather than an unusual one, and it is worth saying so in the same breath: the notable contracts in this corpus are the handful that undertake anything back to the customer. If a client agreement requires the vendor to stand behind generated output, this document does not do it, and that is a fact about the industry's default as much as about this company.

Ownership of what the tool generates is the finding on this page that took the longest to establish and is the most useful to a buyer. The contract never uses the word Output. It defines the AI-generated material at section 1.4 as Results (the service may use third-party models "to generate certain responses, templates, suggestions, drafts, materials and results") and then does two things with that definition. It disclaims every warranty about them: "and Gamma makes no representations or warranties of any kind or nature with respect to any Results". And it never states who owns them. That absence was searched for rather than assumed. Results appears five times in the whole agreement, all five inside section 1.4, and none of them near a word about ownership, title, rights or licence. Section 4.3 says "Gamma does not claim ownership of Your Content." Section 4.1 carves out of Gamma's own intellectual property only "Your Content" and "User Content". Results are named in neither carve-out.

For most uses this is academic. For somebody generating an investor deck, a client deliverable, or anything that will carry a copyright line, it is a real gap: the document defines the generated material as its own category and then leaves that category unassigned. This is reported as ambiguity and nothing more. It is not a legal conclusion, it is not a claim about what anybody meant, and a reader who needs certainty should ask the vendor for it in writing before the question becomes expensive rather than after. The same search found no per-tier language at all, the phrases free plan, free tier and free version return zero occurrences and the agreement does not distinguish paid accounts from free ones anywhere, so there is no free-tier commercial restriction to report and equally no explicit permission. Permitted by omission is not a stated permission, and that position is what holds this row at 70.4 rather than higher.

Does Gamma Train Its AI on Your Content? The Answer Differs by Workspace Type

The vendor answers this in a help-centre article carrying its own absolute date of April 29, 2026, and the answer splits by the kind of workspace you are in. On individual workspaces (Free, Plus, Pro and Ultra), the article states the default setting: "Data is allowed to be used to improve Gamma's AI features", with an opt-out available in account settings at any time. On team and business workspaces the position inverts and is fixed: "For all Team and Business workspaces, data is automatically excluded from being used for training", and the article states the setting is locked and cannot be changed.

Two things follow for a buyer. If you are on an individual plan and your material is confidential, the control exists but it is off by default, so it is a setting to change rather than a state to assume. And if that guarantee needs to be structural rather than a toggle somebody might flip, the tiers that provide it are the team ones, which sit behind the seat minimums the pricing section could not date.

Gamma's Refund Policy: a Published 3-Day Window, and a Contract That Says Fees Are Non-Refundable

Two vendor documents answer this question differently, and both are reported here without being reconciled, because both are the vendor's own words on the same day.

The help-centre article publishes an entitlement on objective conditions. From 1 December 2025, refunds are offered only within the first three days of subscribing, on both monthly and annual billing, and only on contacting support. Cancelling is not enough on its own: "You will not automatically receive a refund if you cancel within 3 days of subscribing." Eligibility is additionally capped by credit usage, and the thresholds are published per plan, under 200 credits used on Plus, 600 on Pro, 900 on Team, 1,000 on Business, 2,000 on Ultra. The article's wording is an entitlement rather than a discretion: "Eligible for a full refund within the first 3 days after subscribing, if credit usage does not exceed the threshold for the plan". No refunds are issued after three days except where local law requires, and the article says eligibility is subject to local consumer-protection law, which is the vendor's own carve-out rather than my gloss.

The binding contract says something narrower. Section 9.2 of the Terms of Use reads: "Except as set forth in this Agreement, all Fees for the Service are non-refundable." Its carve-out points inside the agreement, not at the help centre. Section 9.3 adds that Gamma reserves the right to change subscription pricing at any time. A buyer relying on the three-day window should know that the document with legal force does not contain it, and that the practical answer is therefore to test on the free tier before subscribing rather than to subscribe and rely on getting out.

Gamma Alternatives: What One Independent Board Covers, and What Nobody Has Measured

An alternatives search for this product returns a great many rankings and almost no measurements, and the reason is structural rather than editorial. Most of the "best AI presentation tool" lists a search surfaces are published by companies that sell a presentation tool of their own. That is checkable in about ten seconds by opening the publisher's homepage and reading its title, and it is why no ranking of that kind is used as evidence anywhere on this page. It is also why this section names what was actually measured instead of ordering a field.

On 18 August 2026 this desk pulled twenty-two slide-generation benchmark papers from the arXiv API and searched each one for the products a buyer in this category is choosing between. PresentBench is the source that came out of it, for the reasons above. Beautiful.ai appears in two of those twenty-two, in both cases as a passing mention in a related-work paragraph rather than as a system under test. Plus AI appears in none of them. That is an absence of measurement rather than a measurement of absence, and the cause is visible in the task: these benchmarks all pose end-to-end deck generation from source material, which Gamma enters and a template-fill or add-in product does not. Nobody benchmarks tools that are not in the task, so the comparison a reader most wants, whose slides are better, cannot be evidenced today for most of the field.

What can be reported is the field Gamma's own figure sits in, straight off the board it was read from. The nine rows ran from 35.9 to 70.8 on 18 August 2026, with the middle at 53.6. The highest row, at 70.8, is a submission from a Huawei Cloud multimodal team tested in 2026-07; Google's NotebookLM reads 62.5 tested in 2026-01. Those are research-board entries rather than cards on this site: several are general assistants rather than presentation products, none of them has been scored here, a Total is not a composite, and nothing in that paragraph ranks anything against the number on this page. It is the distribution that gives 49.2 its meaning, printed so a reader can see the shape rather than take the conversion on faith.

On 18 August 2026, when this page published, it was the only scored card in the category. That stopped being true the next day and is further from true now: other products in this category carry cards of their own, and the presentation makers hub lists them in an order the site computes from the scores rather than one written here. What would change that is not mysterious: the board is open to submissions and has already taken one, a neutral publisher could start printing numbers instead of prose, or an independent export-fidelity study could appear, the metric is objective and the demand is obvious, and the reason none is used here is that every version this desk could find by 18 August 2026 was published by a company selling a competing product. Each of those cards carries its own dimensions, its own dates and its own gaps, and they are readable beside one another without any of them being crowned.

Who Should Buy Gamma at $12 a Month, and Who Should Stop at the Free Credits

Start on the free plan. That is not a hedge, it is the specific recommendation this evidence supports: 400 credits, no card required, and every export format present, which means you can generate the deck you actually need, export it in the format you actually hand over, and find out whether the output suits you before any money moves. Given a three-day refund window that the binding contract does not itself contain, testing first is worth more here than it is on most subscriptions.

Pay the $12 if the free run went well and the shape of your work matches what the record rewards. Gamma fits volume: many decks, quickly, where the draft is the point and the deadline is closer than the design standard. It fits a presenter who shares a link or a PDF rather than a file somebody else edits. It fits somebody who accepts that a first draft needs a pass, which is exactly what the five-star reviewer quoted above argues, and exactly what the one-star reviewers say they could not get the editor to accept. And at a thousand credits a month it is genuinely cheap for what it does, which is why value is the highest row on this card.

Do not buy it, on this evidence, if any of these describes you. Your deliverable is a pixel-exact PowerPoint file for somebody else to open, the vendor documents that exports render the presentation view rather than the editing view, that gradients and frosted effects may fall back, and that decks past roughly 150 slides can fail to export at all. You need to iterate slide by slide with the AI honouring narrow instructions, because that is the most repeated complaint in the sample read for this page and it is the same axis the benchmark scores lowest. Design is the deliverable rather than the vehicle, since 22.6 on Visual Design and Layout is the weakest figure on Gamma's benchmark row and reviewers describe the same thing in plainer words. Or you need written certainty about who owns what you generate, in which case the contract as it stands on 9 September 2025 does not give it to you and no amount of product quality substitutes for that.

What the 73.2 does not settle is anything about your material. Four rows convert a benchmark's field position, one review platform's reading with a second printed beside it, a price list and a contract. Not one of them samples your prompts, your source documents or your audience, and the two most useful sentences on this page are not on the scale at all: the free tier exports, and the contract never says who owns the Results.

Card assembled and page written by Minel Gunesoglu, founder of Vouch. LinkedIn · what this desk can and cannot verify. Every source behind this page was opened on 18 August 2026: the PresentBench board and its paper, gamma.app's pricing in both billing views, the Terms of Use, six help-centre articles, the G2 and Trustpilot profiles, and the twenty most recent reviews on the latter. Capterra, Product Hunt and Reddit refused every method attempted and are recorded as gaps rather than papered over. Disclosure: no affiliate relationship exists with Gamma, so every link above is a plain link; where affiliate links appear elsewhere on this site they are marked, and a paid submission buys review speed only, never a listing and never a score.