The AI video leaderboard rearranged itself this summer. Between July and September, a wave of new models — Seedance 2.5, MiniMax H3, Wan 3.0, FLUX 3 and Utopai X — pushed Seedance 2.0, the Arena leader through mid-year, down to #7 on text-to-video, and left Veo 3.1 at #19. The most useful number in this ranking is price: the cheapest model in the top five (MiniMax H3, about $4.80 per minute) costs a seventh of the dearest (Seedance 2.5, $34.12), for scores within 17 Elo points of each other. So the real question is no longer who wins, but what you are paying the premium for.
Six picks, by job — with per-minute math throughout.
How we picked
Same rules as every ranking on this site: sourced claims, dated pricing, models you can actually buy. Rank comes from Artificial Analysis’s blind human votes on both its text-to-video (v2.0) and image-to-video (v1.1) boards, each with audio. Gaps under about 20 Elo points sit inside the error bars, so we treat them as ties and break them on price and access. Two models sit out for that last reason: Utopai X ranks #2 on text-to-video but has no API, and Kling 4.0 is announced with only a Flash tier in limited early access.
The two boards use separate scales, each anchored at Kling 3.0 1080p Pro = 1000. Compare scores within a board, not across them. Prices below are Artificial Analysis’s cost to generate one minute of 1080p video at each model’s default settings, cross-checked against vendor price lists.
The short answer
| Your situation | Pick |
|---|---|
| Best all-round: quality and price — the default | MiniMax H3 |
| Best text-to-video quality | Wan 3.0 |
| 30-second single takes, many references | Seedance 2.5 |
| Animating images, editing by conversation | Gemini Omni 1.1 Flash |
| Faces, human motion, published rate card | Kling 3.0 |
| Free to start | PixVerse V6 |
All six, side by side
The seven-fold price spread, in one picture — Artificial Analysis’s price per minute of 1080p text-to-video (Veo 3.1 added for reference):
The full decision matrix — scores and prices from Artificial Analysis, specs from each model’s page:
| Model | Text-to-video Elo (rank) | Image-to-video Elo (rank) | Price / min | Max clip | Access |
|---|---|---|---|---|---|
| MiniMax H3 | 1141 (#4) | 1131 (#1) | $4.80 | 15s | Open weights + API |
| Wan 3.0 | 1158 (#1) | 1118 (#5) | $12.00 | 30s | Alibaba Cloud API (preview) |
| Seedance 2.5 | 1144 (#3) | 1125 (#4) | $34.12 | 30s | BytePlus / Volcano API (not US) |
| Gemini Omni 1.1 Flash | 1115 (#8) | 1130 (#3) | $9.12 | 10s (extendable) | Gemini API (paid) |
| Kling 3.0 | 1000 (#16) | 1000 (#19) | $10.08 | 15s | Kling API, published rates |
| PixVerse V6 | 985 (#18) | 1024 (#16) | $6.90 | — | Daily free credits + API |
Dashes mean the vendor doesn’t state it plainly.
The reasoning, pick by pick
1. MiniMax H3 — the best all-rounder
MiniMax H3 is one of only three models in the top five of both boards — with Wan 3.0 and Seedance 2.5 — and it is #1 on image-to-video (Elo 1131) and #4 on text-to-video (1141). What puts it first on this page is price: MiniMax lists $0.08 per second at 768p and $0.13 at 2K, which Artificial Analysis converts to $4.80 per minute — less than half of Wan 3.0 and about a seventh of Seedance 2.5. It is also the only one of the leaders with open weights (the 33B base model is on Hugging Face under a community license), so teams with GPUs can self-host.
The honest caveats: Artificial Analysis ranks the 768p output, so its 2K tier is not independently measured; clips stop at 15 seconds; and the module that lifts output to 2K stays API-only. Its community license also caps commercial use at $20M annual revenue without a separate agreement.
Strengths: top-five on both boards; cheapest of the leaders by a wide margin; open weights; omni-reference input (images, video, audio in one prompt). Watch out: 15-second ceiling; 2K is API-only and unbenchmarked; license caps commercial use at $20M revenue.
2. Wan 3.0 — the text-to-video leader
Wan 3.0 holds the highest text-to-video Elo (1158) and accepts a kind of input no other model here does: documents and web pages, which it turns into video directly. It generates 30-second clips in a single pass, with list prices of $0.05, $0.10 and $0.20 per second at 480p, 720p and 1080p on Alibaba Cloud. Alibaba’s docs also list a pricier “prime” tier and a limited-time discount on the standard one, so check your region’s console.
Its lead is thin: Artificial Analysis’s own rank range for it is 1–2, the #5 text-to-video entry is only 24 points behind, and its image-to-video score (1118, #5) trails H3 and Omni. It is closed-weight, a departure from the open-weight Wan 2.x line, and Alibaba’s API reference still labels it “in preview.”
Strengths: highest text-to-video Elo; 30-second clips; document and web-page inputs; $12.00 per minute. Watch out: lead is within the error bars; still labelled “in preview”; closed weights; model names, tiers and availability vary by Alibaba Cloud region.
3. Seedance 2.5 — long takes and reference control
Seedance 2.5 launched on July 31 and sits at #3 on text-to-video (1144) and #4 on image-to-video (1125). Its case is control: native 30-second single shots, up to 50 multimodal reference assets, phoneme-level lip-sync, and localized scene editing that fixes one detail without re-rolling the clip. If you are producing long, reference-heavy brand or e-commerce video, nothing else here matches that spec sheet.
You pay for it. Artificial Analysis lists $34.12 per minute for text-to-video, and BytePlus bills roughly $0.23 per second at 720p. And the “native 4K” announced in June is not on the API: BytePlus’s reference lists 480p, 720p and 1080p. BytePlus also says its services are not available in the United States.
Strengths: 30-second single shots; 50 reference assets; localized editing; top-four on both boards. Watch out: the priciest leader at $34.12 per minute; no 4K on the API; BytePlus not available in the US; real faces can’t be uploaded as references.
4. Gemini Omni 1.1 Flash — images in, edits by chat
Gemini Omni 1.1 Flash is Google’s current video model, released August 27. It ties for #1 on image-to-video (Elo 1130, within one point of H3 and Vidu Q4 Preview) at $9.12 per minute, and it is the only model here you refine by conversation: ask for a change and it edits the clip, extends it, or interpolates between a first and last frame. A 360p draft tier at about $0.034 per second makes iteration cheap.
Limits matter: each generation is 3–10 seconds, 1080p and 4K are upscaled from a 720p-class render, text-to-video is only #8 (1115), and there is no free API tier. Editing or extending uploaded videos is unavailable in the EEA, Switzerland and the UK.
Strengths: ties for #1 on image-to-video; conversational editing and extension; cheap 360p drafts; SynthID watermarking. Watch out: 10-second clips; upscaled high resolutions; mid-table text-to-video; no free API tier.
5. Kling 3.0 — the people specialist
Kling 3.0 has slipped in blind votes — it is the boards’ 1000-point anchor, roughly 140–160 Elo below the leaders — but its strengths are the ones those votes measure least: human motion and face consistency across a clip, plus native 4K, multilingual audio and a published rate card. Kling lists $0.084/sec at 720p and $0.112 at 1080p without native audio, $0.126 and $0.168 with it, and $0.42 at 4K. At $10.08 per minute it sits in the middle of the pack rather than undercutting it, so the case for it is faces and predictability, not value.
Kuaishou announced Kling 4.0 on September 28–30 (30-second clips, up to 10 keyframes, stereo audio), but only the lighter Flash tier is in limited early access and there is no API or price list yet. Treat 3.0 as the version to build on and revisit when 4.0 ships.
Strengths: faces and human motion; native 4K; multilingual audio; stable published rates. Watch out: well below the leaders on blind votes; 4K costs $0.42/sec; clips top out at 15 seconds; strict content filter.
6. PixVerse V6 — the free pick
PixVerse V6 pairs daily free credits with features that usually cost money: multi-shot character consistency, native audio and 20+ cinematic lens controls. The ceiling: free and Standard ($10/month) cap at 720p, 1080p/4K needs Pro and up, and the API bills separately by resolution and audio ($6.90 per minute on Artificial Analysis). At #18 on text-to-video (985) and #16 on image-to-video (1024) it is a way to start without paying, not a quality leader.
Strengths: genuinely useful free tier; multi-shot character consistency; 20+ lens controls; native audio. Watch out: 720p ceiling until Pro; ranks in the lower half of both boards; API bills resolution and audio separately.
Also considered
These were in earlier versions of this ranking or are close contenders. None made the six:
- Seedance 2.0 — the Arena leader through mid-2026, now #7 on both boards (1120 text-to-video, 1114 image-to-video) at $22.45 per minute. Still the documented 4K option on BytePlus’s API, but H3 and Wan 3.0 beat it for less.
- Veo 3.1 — #19 on text-to-video (965) and #13 on image-to-video (1034) at $24.00 per minute. Still polished and widely available; Veo 3.1 Fast ($7.20) and Lite ($4.80) score lower still.
- FLUX 3 Video — #6 on text-to-video (1131) and #8 on image-to-video (1084) at $17.40 per minute; up to 20 seconds with native audio. BFL’s own tests ranked it first, but independent votes put it behind the six above on price and rank.
- Runway Gen-4.5 — still the filmmaker’s workflow: motion brush, camera moves and an editorial suite, credit-based at roughly $0.10–0.25/sec. It is not on Artificial Analysis’s current boards, so we cannot rank its raw output; you are buying the toolset.
- Hailuo 2.3 — MiniMax’s earlier budget model, now filed under “Legacy Models” on its price list: $0.19–0.28 per 6-second 768P clip (about $0.032–0.047/sec) and $0.33–0.49 at 1080P. Not on Artificial Analysis’s boards, so unranked, and MiniMax H3 is the newer pick for the same buyer.
- Grok Imagine Video 1.5 — #11 on text-to-video (1047) and #9 on image-to-video (1065), $15.00 per minute; native 1080p and voice references.
- Sora 2 — physics-realistic with synced audio, but access keeps shifting and it is not on the boards.
- MAGI-2 Preview — Sand.ai’s Apache-2.0 114B MoE; #20 on text-to-video (962) and #11 on image-to-video (1044). The strongest fully permissive license, but H3 is the stronger open-weights model.
- Wan 2.7 — Alibaba’s open-weight 1080p model, the community self-hosting base; superseded by Wan 3.0 for hosted use.
On the horizon
- Kling 4.0 — Kuaishou’s next generation, announced September 28–30, 2026 (preview): native 30-second clips, up to 10 keyframes and two-channel stereo audio. Only Kling 4.0 Flash is in limited early access; the full model is due “in October” with no API, price list or independent benchmark yet.
- Vidu Q4 Preview — the October 2026 preview sits at #2 on Artificial Analysis image-to-video (Elo 1130) at $5.04 per minute. Preview status and no full vendor documentation yet, so it stays unranked until it ships.
- Utopai X — #2 on text-to-video (1149), but Artificial Analysis lists no API, so there is nothing to buy.
Full specs for every model live in the video models directory. The full ranking was last reviewed on October 13, 2026.
Changelog
- 2026-10-13 — Full re-review against Artificial Analysis’s text-to-video v2.0 and image-to-video v1.1 boards. New ranking: MiniMax H3, Wan 3.0, Seedance 2.5, Gemini Omni 1.1 Flash, Kling 3.0, PixVerse V6. Moved Seedance 2.0, Veo 3.1, Runway Gen-4.5 and Hailuo 2.3 to “Also considered”; Seedance 2.5, MiniMax H3 and Wan 3.0 promoted from “On the horizon”. Corrected Kling 3.0 pricing to Kling’s official list rates and added Kling 4.0 (preview). Price comparisons now use Artificial Analysis’s price per minute.
- 2026-08-11 — August update 2: added Wan 3.0 (30-second clips, document-to-video) and MAGI-2 Preview (Apache 2.0, 114B MoE with native audio) to “On the horizon”; updated Wan 2.7 entry to note Wan 3.0 succession.
- 2026-08-05 — August update: FLUX 3 Video GA with pricing; added MiniMax H3 (first major open-weight video model) and Grok Imagine Video 1.5 (native 1080p) to “On the horizon”; confirmed ranking unchanged pending independent Arena results for the new entrants.
- 2026-07-29 — July update: added FLUX 3 to “On the horizon” as a new image+video unified entrant; updated Seedance 2.5 status.
- 2026-07-02 — Initial ranking published: Seedance 2.0 · Veo 3.1 · Kling 3.0 · Runway Gen-4.5 · PixVerse V6 · Hailuo 2.3, with Seedance’s double Arena lead (Elo 1,269 T2V / 1,351 I2V) weighed above its provider-dependent access.
Frequently asked questions
What is the best AI video generator right now?
MiniMax H3 is our best overall pick: it is #1 on Artificial Analysis image-to-video and #4 on text-to-video, and at about $4.80 per minute it costs a fraction of Wan 3.0 ($12.00) and Seedance 2.5 ($34.12), the other two models in the top five of both boards. If you only care about text-to-video quality, Wan 3.0 holds the top Elo — but its lead is within the error bars and Alibaba still labels the model "in preview."
Is Wan 3.0 better than Seedance 2.5?
On text-to-video, Wan 3.0 scores slightly higher (Elo 1158 vs 1144) and costs about a third as much ($12.00 vs $34.12 per minute); each score carries a ±9 interval, so treat it as close. Seedance 2.5 edges Wan 3.0 on image-to-video (1125 vs 1118) and offers a larger reference limit. Both produce 30-second clips.
What happened to Seedance 2.0 and Veo 3.1?
Newer models passed them in blind votes. Seedance 2.0, which led the Arena through mid-2026, is now #7 on both Artificial Analysis boards (Elo 1120 text-to-video, 1114 image-to-video) at $22.45 per minute. Veo 3.1 is #19 on text-to-video (965) and #13 on image-to-video (1034) at $24.00 per minute. Both are still capable and widely available; they just no longer justify a top-six slot on price and rank.
Does Seedance 2.5 generate 4K video?
Not through the API. BytePlus's reference lists 480p, 720p and 1080p for Seedance 2.5 and keeps 4K for Seedance 2.0. Dreamina and CapCut advertise 4K in their apps without explaining how it is produced, so treat it as unconfirmed.
What is the best free AI video generator?
PixVerse V6 — daily free credits, multi-shot character consistency and native audio, with the caveat that free output caps at 720p. It ranks #18 on Artificial Analysis text-to-video, so expect good-enough shorts rather than top-tier quality. Gemini Omni 1.1 Flash has no free API tier.
Which AI video model is cheapest?
Among the top-ranked models, MiniMax H3 at about $4.80 per minute ($0.08/sec at 768p). Cheaper models exist but rank lower: Artificial Analysis lists Agnes-Video-2.5 at $2.40 per minute (#10 on text-to-video, Elo 1058), and MiniMax's own price list files Hailuo 2.3 under "Legacy Models" at $0.19–0.28 per 6-second 768P clip (about $0.032–0.047/sec), though Hailuo is not on Artificial Analysis's boards. At the other end, Seedance 2.5 costs $34.12 per minute.
What about Sora 2?
Sora 2 produces physics-realistic clips with synced audio, but its access model has shifted repeatedly in 2026 — it is available through ChatGPT and third-party API routes (~$0.15/sec) rather than a straightforward first-party API. It is not on Artificial Analysis's current boards, so we cannot rank it against the others.
Should I wait for Kling 4.0?
Only if you need 30-second clips, timed keyframes or stereo audio. Kling 4.0 was announced on September 28–30, 2026, but only the lighter Kling 4.0 Flash is in limited early access; the full model is due "in October" with no API, price list or independent benchmark yet. Kling 3.0 is the version you can use at scale today.
How are these models ranked?
By Artificial Analysis's blind human-preference votes on both text-to-video (v2.0) and image-to-video (v1.1), each with audio. We treat gaps under about 20 Elo points as ties and break them on price and access. The two boards use separate scales, each anchored at Kling 3.0 1080p Pro = 1000, so compare scores within a board, not across them.
How often is this ranking updated?
We re-check picks when major models ship. The full ranking was re-reviewed on October 13, 2026 against Artificial Analysis's text-to-video v2.0 and image-to-video v1.1 boards, vendor price lists and official documentation.
Sources
- Artificial Analysis — AA-Video-T2V v2.0 leaderboard
- Artificial Analysis — AA-Video-I2V v1.1 leaderboard
- MiniMax API Docs — pay-as-you-go pricing
- Alibaba Cloud Model Studio — model pricing
- Alibaba Cloud Model Studio — Wan3.0 video generation API reference
- Alibaba — Wan 3.0
- ByteDance Seed — Introducing Seedance 2.5
- BytePlus ModelArk — model pricing
- Google — Build with Gemini Omni 1.1 Flash
- Gemini API pricing
- Kling AI — API pricing
- Kling AI — Kling 4.0 vs 3.0: What's Different in the New Model?
- PixVerse — official site
- Runway — official site
- MiniMax — Hailuo
- Sand.ai — MAGI-2 Preview