AI Video Generation Statistics 2026: 60+ Data Points

AI video generation statistics are the most polluted data category we track: the space moves so fast that most published roundups recycle each other's numbers, cite aggregators instead of primary sources, and quietly flatten three different market-size reports into one headline figure. This collection takes the opposite approach — every statistic below is dated, attributed to a named source, and graded by how it was disclosed.

The stakes of getting this right are not academic. Teams are budgeting real production spend against per-second API prices that are widely misquoted, and two labeling laws — the EU AI Act's Article 50 and California's SB 942 — come into force on August 2, 2026. A strategy built on a zombie stat is a strategy built on sand.

This is a statistics reference for the generation technology itself: model adoption, detection accuracy, first-party pricing, the Sora shutdown arc, funding, regulation, and market sizing. For survey data on how marketing teams use video generally, see our companion piece on video marketing statistics — the two collections deliberately do not overlap.

Most AI video statistics circulating in mid-2026 fail one of three ways. First, aggregator laundering: a stats-farm site invents or garbles a number, a dozen roundups cite the stats-farm, and within months the figure carries false authority through repetition. Second, scope flattening: at least three legitimate market-size reports with materially different scopes get quoted interchangeably as "the AI video market." Third, metric conflation: registered accounts become "users," theoretical peak costs become "actual burn," and a download count becomes a DAU figure.

Our fix is structural, not rhetorical. Every number in this collection carries a named source and a publication date, inline or in the table where it appears. Beyond that, we grade each figure by how it entered the public record — because an investor-relations disclosure from a listed company, a peer-reviewed study, and a vendor press release are not the same kind of evidence.

The single best-sourced adoption dataset in AI video belongs to Kling AI, because its parent Kuaishou is a Hong Kong-listed company that reports Kling metrics through investor-relations channels. Per Kuaishou's IR release of January 13, 2026: as of December 2025, Kling served 60 million+ creators, had generated 600 million+ videos, and counted 30,000+ enterprise partners.

The revenue trajectory is a clean three-point line. Kuaishou disclosed a $100M annualized revenue run-rate in March 2025 — ten months after launch — and $240M by December 2025. Then Kuaishou's Q1 2026 earnings, as reported by the South China Morning Post on May 27, 2026, showed Kling revenue up more than 300% year-over-year to roughly RMB 650M (~$90M) for the quarter, an annualized run-rate of roughly $500M. The same SCMP report notes Kling topped App Store charts in 42 markets during the quarter, including Brazil and Germany.

On the demand side, Wyzowl's 2026 State of Video Marketing report (the 12th annual edition, surveyed late 2025, n=266) found 63% of video marketers have used AI tools to create or edit marketing video — up from 51% the year before. But the multi-year line is volatile, not smooth: Wyzowl's separate 2024 AI-in-marketing report (n=128, surveyed March 2024) had reported 75%. Those are different surveys with different sample sizes, and the honest reading is that adoption measurement is still noisy — not that adoption fell by a third and recovered. Competitor roundups routinely smooth these three numbers into one trend line; we won't.

Our interpretation: the supply-side numbers (Kling's 600M videos, Veo's 70M in two months) are growing an order of magnitude faster than the demand-side survey numbers, which suggests most AI video volume in early 2026 was consumer experimentation and social content, not budgeted marketing production. That gap is exactly where operationalized content engine workflows are being built — turning casual tool usage into repeatable production. For the model-level detail behind these adoption numbers, see our hands-on guide to Kling's 4K/60fps output and our breakdown of ByteDance's Seedance model.

The most consequential AI video statistic of the past year is not a revenue number. It's a detection number. A peer-reviewed study published in Communications of the ACM on September 22, 2025 — titled, aptly, "As Good as a Coin Toss" — measured human accuracy at identifying AI-generated content at 51.2% across all media types, barely above the 50% guessing baseline. For video-only stimuli, accuracy was 50.7%.

The study's asymmetry finding matters even more for anyone thinking about misinformation risk: participants were much better at recognizing real content as real (64.6%) than at catching fake content as fake (38.8%). People default to believing. That asymmetry, combined with a coin-toss overall rate, is the empirical case for the machine-readable labeling laws covered in section 06 — if humans cannot do the detection, the metadata has to.

Four months later, Runway published its own detection study, "The Turing Reel" (January 22, 2026): 1,043 participants scored 57.1% accuracy against its Gen-4.5 model, and only 99 of 1,043 participants — 9.5% — achieved statistically significant accuracy (at least 15 of 20 correct). Runway's own framing: over 90% of participants could not reliably distinguish Gen-4.5 outputs from real video. The essential caveat is that this is a vendor-run study of the vendor's own model — a materially weaker evidentiary bar than the CACM study, which is why the table below labels every row by study type.

Per-second API pricing is where AI video misinformation gets expensive, because teams budget against it. The most recycled claim on the SERP is a "$0.02 per second" floor, usually attributed to Kling. We checked every major vendor's own pricing page, and that number does not exist: Kling's official floor is $0.084/second, and across Runway, Kling, and Google Veo the first-party floor is $0.05/second. The only official figure anywhere near two cents is ByteDance's Seedance 1.5 Pro at roughly $0.022/s on BytePlus ModelArk — a different vendor than the claim is usually attached to, and still not $0.02.

Two things in this table deserve explanation rather than flattening. First, the Veo range: $0.05–$0.75/s is not one product's price band, it's price compression across model generations — Veo 3.0 launched at $0.50/s (no audio) and $0.75/s (with audio) in May 2025, and the later Veo 3.1 generation cut the range to $0.05–$0.40/s across its Lite, Fast, and Quality tiers. Quoting "Veo costs $0.75/s" and "Veo costs $0.05/s" are both technically citable and both misleading without the generation label.

Second, the projection this trend supports: Veo's roughly 10x floor drop in under a year, alongside sub-$0.15 flagship pricing from Runway and Kling, points toward per-second pricing becoming a rounding error for short-form output by 2027 — at which point the competitive axis shifts from generation cost to editing control, rights, and distribution. Teams evaluating production stacks should weight workflow fit over today's per-second delta, because today's per-second delta keeps shrinking.

No dataset says more about the economics of consumer AI video than the Sora arc — seven months from the fastest launch in the category to a shutdown announcement. The verified timeline: the Sora app hit 1 million downloads in under 5 days after its late-September 2025 launch, faster than ChatGPT's launch pace, per OpenAI's Bill Peebles as reported by CNBC (October 9, 2025). OpenAI announced the discontinuation on March 24, 2026, closed the app and web experience on April 26, 2026, and scheduled the API to follow on September 24, 2026, per OpenAI's own Help Center notice.

The cost figures need careful separation, because two very different numbers get conflated. The Wall Street Journal (via TechCrunch, March 29, 2026) reported Sora was losing roughly $1M/day — a reported actual net-burn figure, after OpenAI had already throttled quality to control costs. Separately, Forbes (April 2, 2026) cited a Cantor Fitzgerald estimate that Sora's peak inference costs could have reached ~$15M/day — a theoretical peak-load estimate, not a measured burn rate. Against either number, Appfigures' estimate of ~$2.1M in lifetime in-app-purchase revenue (cited across shutdown coverage; not an OpenAI-confirmed figure) frames the unit-economics gap that ended the product.

The download-decline chart and the user-count contest are also a live methodology lesson: downloads, registered users, worldwide user counts, and DAU are four different metrics, and most secondary coverage swapped them freely. We've kept each figure attached to its own metric type above. For the full story of what happened and why, see our analysis of why Sora shut down, the deeper economics behind Sora's losses, and how the AI video market reshuffled after Sora.

While consumer AI video burned cash, the enterprise avatar segment quietly built real revenue. HeyGen announced on June 25, 2026 (via Business Wire) that it crossed $200M in annual recurring revenue, doubling in eight months — serving 30M+ users across 196 countries, with 85% of the Fortune 100 among its customers and 118M+ videos created on the platform. HeyGen also states it generates roughly $2.70 in ARR for every $1 of equity raised. All of these are company-reported figures from HeyGen's own press release — no independent audit is available — and we label them accordingly.

On the platform side, TikTok discloses that it has labeled over 1.3 billion videos as AI-generated to date (TikTok Newsroom) — the only platform-level disclosure of its kind; no comparable labeled-video count exists from YouTube, Instagram, or X as of this writing. TikTok's labeling stack combines creator self-disclosure, in-house detection models, and C2PA Content Credentials metadata, with invisible watermarking rolling out for TikTok-native AI tools. Brands publishing AI-assisted video into these feeds — a workflow we run inside our social media management practice — should assume platform-level AI labels are the default condition of distribution from here on. One caution: a widely recycled "TikTok enforcement up 340%" figure does not come from TikTok — it's covered in the refused-stats section below.

The regulatory numbers carry teeth. EU AI Act Article 50 requires systems generating synthetic video to mark outputs in a machine-readable, detectable format from August 2, 2026, with penalties for transparency breaches up to €15M or 3% of global annual turnover, whichever is higher; a draft Code of Practice (published March 2026) is expected to be finalized ahead of the deadline. California's SB 942 requires covered providers (1M+ monthly California users) to embed machine-readable provenance watermarks and offer a free public AI-content detection tool from the same date — text-only output is excluded, and a further platform-side C2PA provenance-detection obligation begins January 1, 2027.

Ask "how big is the AI video market?" and the SERP will hand you $3.44B, $42.29B, and $60.8B — presented interchangeably. They are not interchangeable. They come from three differently-scoped reports measuring three different things, and flattening them into one number is the single most common error in this category. The decoder:

For context against those forecasts: verified vendor run-rates alone — Kling at roughly $500M annualized by mid-2026, HeyGen at a company-reported $200M, Synthesia past $100M as of April 2025 — already approach the narrow Grand View 2025 base of $788.5M, which suggests the narrowest forecast is conservative about how fast revenue is actually compounding. That's an inference from the disclosed numbers above, not a published finding; we flag it as ours.

Every figure above survived primary-source verification. These five did not — and each one circulates widely enough that you will meet it in a pitch deck or a competitor's roundup this quarter.

"TikTok AI enforcement up 340%, with 51,618 removals in H2 2025"

Rejected: no primary source exists. These figures appear only on third-party stats-aggregator sites, which cite TikTok's Newsroom for the genuine 1.3B labeled-videos figure and then attach the 340%/51,618 numbers with no traceable citation. TikTok's own Q4 2025 Community Guidelines Enforcement Report gives different, non-comparable figures: 175M total videos removed platform-wide, of which 1.8% were flagged as "edited media and AI-generated content." No TikTok primary source states a 340% AI-enforcement increase.

Runway and Midjourney "monthly active user" counts

Rejected: no first-party MAU disclosure exists from either company. Secondary trackers conflate Midjourney's ~19.8M registered accounts with estimated daily actives of 1.2–2.5M — an 8–16x gap between the headline number and engagement estimates. Any specific Runway or Midjourney MAU figure you encounter is a triangulated guess presented as a disclosure.

"52% of B2B marketers call AI video their most-adopted new tech" and "LinkedIn AI video content up 310% in 2025"

Rejected: unsourceable. Both recur across content-mill roundups with no link to any LinkedIn Business or Marketing Solutions primary source, and LinkedIn's own B2B research hubs do not publish either figure as far as our verification could determine.

"The AI video market: $2.29B in 2024, $3.73B by 2033"

Rejected as an AI-video-generation figure: scope mismatch. This is Straits Research's Video Editing Software Market report — editing tools, not generative video — which mentions AI as a growth driver and gets routinely mis-filed under AI video generation. Not a fabrication, but wrong category.

"X% of AI videos are used for fraud or misinformation"

Rejected wholesale. We encountered multiple variants of this claim with specific percentages on aggregator sites during research, and none traced to a named primary report. Until a platform, regulator, or research institution publishes a methodology-backed figure, any specific fraud-rate percentage in this category is decorative.