Rankings in AI video move fast enough that a "best model" claim from three months ago is often already wrong. This page tracks the current standing on the industry's most-cited independent leaderboard, then goes model by model through launch dates, capabilities and verified pricing, sourced directly from each company's own announcements and pricing pages.
The gap at the top is closer than the marketing suggests: the four leading models are separated by under 50 Elo points on the benchmark below, meaning the practical difference between "best" and "fourth-best" is often smaller than the difference between two prompts on the same model.
Part of our AI Video Generation Statistics 2026 series.
1,247
Runway Gen-4.5's Elo score on the Artificial Analysis Video Arena, currently #1.
Google Veo 3 (1,226) and Kling 2.5 (1,225) are essentially tied for second.
Our take: First place by 21 points out of 1,200+. A lead this thin will probably flip by the time you're reading this.
Runway Research01 / 3
The current leaderboard
Artificial Analysis runs blind, crowd-voted comparisons across every major model and is the closest thing the industry has to an independent ranking, no vendor self-reporting, no cherry-picked demo clips. The headline number is how close the top four are: 41 Elo points separate #1 from #4, which in practice means the "best" model for your specific prompt might not be whichever one currently sits on top.
Runway Gen-4.5 leads at 1,247 Elo, its December 2025 release.
Our take: The number to watch here isn't who's #1, it's how often the #1 changes.
Runway ResearchGoogle Veo 3 sits second at 1,226 Elo.
Our take: 19 points behind Runway, a gap most viewers couldn't perceive in a blind side-by-side.
Runway ResearchKling 2.5 is essentially tied for second at 1,225 Elo.
Our take: Essentially tied with Veo, from a company charging noticeably less per generation.
Runway ResearchOpenAI Sora 2 Pro rounds out the top four at 1,206 Elo.
Our take: Fourth on the leaderboard the same month the consumer Sora app got discontinued, ranking and business viability are different questions.
Runway Research
02 / 3
Model by model
Launch history and verified pricing for the models above, plus the rest of the field, pulled directly from each company's own announcements and pricing pages, not a reseller's markup. Pricing is the number that changes fastest: Google cut Veo 3 by nearly half within four months of launch, and Kling shipped native 4K before any competitor priced it in at all, so treat the figures below as a snapshot and check the linked source before you budget against them.
OpenAI Sora 2 (Sep 30, 2025): synchronized audio, enhanced steerability. API pricing was $0.10/sec for sora-2, up to $0.70/sec for sora-2-pro at 1080p. The consumer Sora product was discontinued April 26, 2026.
Our take: Discontinued five months after its own leaderboard debut, proof a model's ranking doesn't guarantee the product survives.
OpenAIGoogle Veo 3 (May 2025): 8-second clips with native audio and lip-sync. Google cut pricing 47% in September 2025, from $0.75 to $0.40/sec.
Our take: A 47% cut within four months of launch, Google's playing a volume game here, not a margin game.
Google Developers BlogRunway pricing: Free (125 credits), Standard $12-15/mo, Pro $28-35/mo, Max $76-95/mo.
Our take: The free tier's 125 credits go fast, budget for Standard if you're doing more than a handful of test generations.
RunwayKling 3.0 (Feb 2026): 15-second max clips, 2K/4K output. Kling became the first model with native 4K output per Kuaishou's Q2 2026 earnings.
Our take: First to 4K, but check whether your output pipeline, social specs especially, can even use it before paying for the upgrade.
Kuaishou TechnologyPika pricing: Free (80 credits), Standard $8/mo, Pro $28/mo, Fancy $76/mo, all paid tiers on Pika 2.5.
Our take: The cheapest entry point among the models we checked pricing for, worth a look if budget matters more than leaderboard rank.
PikaLuma Dream Machine's Ray2 (Jan 2025): "10x the compute of Ray1," 5-10s clips extendable to 30s, up to 4K upscale for subscribers.
Our take: "10x the compute" is Luma's own framing, we can't independently verify the multiplier, only that the output quality jump was real.
Luma AIMiniMax's Hailuo 02 (Jun 2025): claims 2nd place globally on the Artificial Analysis Video Arena at launch. Hailuo 2.3's "Fast" mode cuts batch-creation costs up to 50%.
Our take: A launch-day claim from the company itself, worth checking the current leaderboard rather than taking it at face value now.
MiniMaxByteDance's Seedance 1.0 generates a 5-second 1080p video in 41.4 seconds, claiming a 10x inference speedup. Seedance 2.5 extends single generations to 30 seconds.
Our take: 41.4 seconds to generate 5 seconds of video is a genuinely fast turnaround, useful if you iterate through a lot of variations.
ByteDance Seed
03 / 3
Enterprise & avatar tools
Adobe Firefly, Synthesia and HeyGen sit in a different category from the leaderboard above: enterprise-focused avatar and editing tools built for training videos and corporate marketing, not open-ended scene generation. They don't compete on the same benchmark, but they're growing faster than almost anything on it, HeyGen alone crossed 85% Fortune 100 penetration and 118 million videos created, numbers that dwarf most consumer-facing model launches.
Adobe Firefly Video launched Oct 2025, "the first publicly available video model designed to be commercially safe." Pricing runs $9.99-$199.99/mo by tier.
Our take: "Commercially safe" is Adobe's real pitch here, aimed at brands who got burned by copyright uncertainty on other models.
AdobeSynthesia crossed $100M ARR in April 2025, then a $200M Series E at a $4B valuation in Jan 2026, with 1M+ users and 90%+ of the Fortune 100 as customers.
Our take: 90%+ Fortune 100 penetration usually takes a decade in enterprise software, Synthesia got there in about three years.
SynthesiaHeyGen reached $200M ARR in June 2026, having doubled in 8 months, with 30M+ users, 118M+ videos created, and 85% of the Fortune 100 as customers.
Our take: Doubling ARR in 8 months while adding tens of millions of users, growth and monetization scaling together, not one at the other's expense.
HeyGen
Frequently asked questions
Which AI video generator is the best in 2026?
On the Artificial Analysis Video Arena, Runway's Gen-4.5 currently ranks #1 at 1,247 Elo, narrowly ahead of Google Veo 3 (1,226) and Kling 2.5 (1,225). The gap is small enough that the right model often depends more on your specific use case, price, and control features, than the leaderboard rank.
What's the cheapest AI video generator?
Free tiers exist on Runway, Pika and others, but for paid use, Google's Veo 3 Fast at $0.10-0.15/sec and Kling's lower tiers are among the cheapest per-second options among the top-ranked models. Check our AI video cost calculator for exact per-attempt pricing.
How often do AI video model rankings change?
Frequently. Every model in the current top four (Runway, Veo, Kling, Sora) shipped a major version update within the past year, and the leaderboard order has shifted with nearly every release. Treat any "best model" claim as a snapshot, not a permanent verdict.
What's the difference between a video generation model and a tool like Synthesia or HeyGen?
Sora, Veo, Runway, Kling and similar models generate video from a text or image prompt. Synthesia and HeyGen are enterprise avatar platforms built on top of similar underlying technology, but focused on talking-head training and marketing videos rather than open-ended scene generation.