Autoresearch: best AI video generation tools (performance and cost)

Generated by /autoresearch on 2026-08-18. Synthesized across 3 rounds from 14 web pages, anchored by Grokipedia AI_video_generators. See Provenance. Treat as raw material — review before promoting into a project or thread. Context: vault/threads/ai-video-generation

Summary

There is no single “best” AI video tool as of 2026-08-18. On the independent Artificial Analysis Text to Video Arena (with audio), Google Gemini Omni Flash and MiniMax H3 sit at the top of blind preference Elo and are also among the cheapest frontier APIs (~$6–$7.80 per minute of 1080p). ByteDance Seedance 2.0 leads image-to-video. Official Google prices put Veo 3.1 Standard at $0.40/s — four times Omni Flash’s effective $0.10/s — while still ranking mid-pack on Elo. Alibaba’s Wan3.0 (public as of 2026-08) is the longest single-pass hosted clip (30 s) at $0.10/s 720p / $0.20/s 1080p; it is not yet on the AA tables this run fetched. For talking-head / performance-capture work already tracked in this thread, HeyGen and Runway Act-Two remain the commercial products; they are not ranked on AA’s general T2V/I2V boards.

This pass does not redo the May 2026 avatar-model survey. It answers the queue question: which tools currently win on performance and cost.

Findings

There is no universal winner — lanes matter

Grokipedia’s AI video generators page splits the field into generative models (text/image-to-video: Sora, Veo, Runway, Kling, Luma, Pika) and practical/template platforms (Synthesia, HeyGen, InVideo). Google’s own product page for Gemini Omni treats Omni as a conversational generate-and-edit model (text, image, video, audio in; video with audio out), not a talking-head SaaS. The question “best performing / most cost-effective” therefore has to be answered per job: short cinematic T2V, image-to-video, long narrative clips, local/open-weight, or scripted avatars.

Independent quality: Artificial Analysis Elo (fetched 2026-08-18)

The Text to Video leaderboard (with audio) ranks models from blind pairwise votes. API pricing on that page is “cost to generate 1 minute of 1080p video on the model creator’s API at default settings.” Top of the current table:

AA’s own FAQ on that page names Omni Flash as the current best T2V-with-audio model and MiniMax H3 as the best open-weights T2V-with-audio model (leaderboard).

The Image to Video leaderboard (with audio) reorders the same generation: Seedance 2.0 720p leads at Elo 1,197, then MiniMax H3 (1,189) and Omni Flash (1,187). Veo 3.1 is 10th (1,086, $24/min). Open-weights I2V leaders are MiniMax H3, then MAGI-2 Preview (Elo 1,104, pricing “Coming soon”), then LTX-2.3 Fast (959, $2.40/min).

Caveat: AA Elo is general video preference, not an avatar / identity-preservation / hand-fidelity benchmark. It does not replace independent-avatar-benchmarks. Runway Gen-4.5, Sora 2, HeyGen, and Act-Two do not appear in the top-28 T2V-with-audio table fetched today.

Official list prices (API / subscription)

Google — Omni Flash vs Veo 3.1. The Gemini API pricing page prices gemini-omni-flash-preview by tokens and states that output is billed at 5,792 tokens per second of 720p, “approximately $0.10 per second” on Standard — matching AA’s $6.00/min. Veo 3.1 on the same page (paid tier, per second, video with audio): Standard $0.40 (720p/1080p) / $0.60 (4K); Fast $0.10 (720p) / $0.12 (1080p) / $0.30 (4K); Lite $0.05 (720p) / $0.08 (1080p), 4K not supported. So Omni Flash is officially ~4× cheaper than Veo 3.1 Standard at 720p/1080p and sits ~150 Elo above it on AA T2V-with-audio.

OpenAI — Sora 2. Official OpenAI API pricing still lists video generation per second: sora-2 720p $0.10; sora-2-pro $0.30 (720p) / $0.50 (1024p) / $0.70 (1080p). At $6–$42 per minute that is at or above Omni Flash / Veo Fast without a corresponding AA top-28 listing on the tables fetched today. The consumer Sora discontinuation help article (help.openai.com/.../what-to-know-about-the-sora-discontinuation) timed out twice this run — do not treat search-snippet shutdown dates as fetched.

Alibaba — Wan3.0 (Aug 2026, not yet on the AA tables fetched today). Official Model Studio pricing on the Wan3.0 product post: $0.05/s 480P, $0.10/s 720P, $0.20/s 1080P. A 30-second 1080P generation is $6.00; a 30-second 480P draft is $1.50. The 7 Aug 2026 announcement says public beta on Model Studio / Qwen Cloud, 30 s vs Wan2.7’s 15 s max, document-to-video (PPT/PDF/DOC), and character/product reference consistency. Wan2.7 itself (April 2026) is a four-model suite — t2v / i2v / r2v / videoedit — 2–15 s, 720p/1080p (Alibaba Cloud Community, 7 Apr 2026). AA still ranks Wan 2.7 / Wan2.7-260612, not Wan3.0.

Runway — credits, not a flat per-second API. Runway Academy model rates: Gen-4.5 12 credits/sec; Gen-4 Turbo 5/sec; Act-Two 5/sec; Gemini Omni Flash on Runway 10/sec; Veo 3.1 with audio 40/sec; Kling 3.0 Pro with audio 17/sec; Seedance 2.0 720p 36/sec. Search snippets for runwayml.com/pricing quoted Standard $12/mo (annual) / 625 credits (~52 s of Gen-4.5) and Pro $28 / 2,250 credits; the live pricing page timed out, so treat those plan dollar figures as unfetched. Runway’s 1 Dec 2025 Gen-4.5 post claimed #1 on AA T2V at Elo 1,247 vs Veo 3 / Kling 2.5 / Sora 2 Pro. That claim is stale relative to the 2026-08-18 AA table, which neither lists Gen-4.5 in the top 28 nor shows 1,247 as the current leader (Omni Flash is 1,239).

HeyGen — avatar SaaS credits. Official pricing: Free (up to 3 videos/mo); Creator $29/mo / 600 credits; Pro from $49/mo / 1,000 credits (tiers to 100,000 credits at $4,300); Business $149/mo / 1,500 credits + $20/seat. The credit explainer matches those pools. The per-feature rate card is more expensive than the marketing FAQ on the pricing page: Avatar III photo 7 cr/min / video look 4 cr/min; Avatar IV photo 16 / video 31; Avatar V 48 cr/min (photo or video); Custom Expressive Motion (Avatar IV only) +40 cr/min. On Creator (600 credits, $29), Avatar V is 12.5 minutes/month ($2.32/min of output); Avatar III video-look is 150 minutes/month (~$0.19/min). The pricing-page FAQ still quotes a simpler “Avatar III 3 / Avatar IV–V 20 credits/min” — internal contradiction; the dedicated rate-card article is the more specific source.

Kling official pricing was not fetched: klingai.com/blog/kling-video-3-0-credit-cost-guide returned HTTP 446. Use AA’s $15.12–$20.16/min and Runway’s 13–17 credits/sec as third-party meters only.

Best-performing (quality) as of this fetch

• General T2V with audio: Gemini Omni Flash, statistically tied with MiniMax H3 (AA T2V). Google’s Gemini Omni page adds vendor-side human evals (MovieGenBench overall preference / instruction following; I2V VBench “tied” with Grok-Imagine-Video and Kling — Google’s numbers, not AA).
• General I2V with audio: Seedance 2.0 720p, then H3 / Omni Flash (AA I2V).
• Open weights: MiniMax H3. Official open-source note: 4–15 s, up to 2K via hosted H3-Regenerate-2K, 24 fps, 32 kHz stereo, 11 dialogue languages; H3-Omni-Transformer is a 33B dense model (~13B AdaLN cacheable). Weights on Hugging Face MiniMaxAI/MiniMax-H3 under the MiniMax H3 Community License. H3-Context-IR (prompt orchestration) and H3-Regenerate-2K are not in the open drop — local quality will lag the hosted stack unless you rebuild that preprocessing.
• Long clip / document-to-video: Wan3.0, 30 s single pass, $0.10–$0.20/s (Wan3.0 post). Not Elo-ranked on the AA pages fetched today. Authors flag weak audio texture and on-screen text.
• Control-heavy editor suite: Runway still hosts Gen-4.5 plus third-party Veo / Kling / Seedance / Omni on one credit meter (Academy). Quality leadership claimed in Dec 2025 is not supported by the current AA table.

Most cost-effective (usable quality per dollar)

Using AA’s $/min × Elo as a coarse quality-per-dollar screen (Elo is not linear dollars, but the clustering is informative):

• Best frontier $/quality: Omni Flash at $6.00/min and Elo 1,239; MiniMax H3 at $7.80/min and Elo 1,237 (AA T2V; Gemini pricing).
• Cheapest Western hosted that is still mid-pack: Veo 3.1 Lite $4.80/min Elo 1,087; grok-imagine-video $4.20/min Elo 1,062 (AA T2V).
• Cheapest open-weight API on the board: LTX-2.3 Fast $2.40/min Elo 976 — a large quality drop vs the top three (AA T2V).
• Long-form draft-then-finish: Wan3.0 480P $0.05/s then 1080P $0.20/s (Wan3.0).
• Avoid as a default “quality” buy: Veo 3.1 Standard at $24/min for Elo 1,089 — same Elo band as Veo Fast ($9/min) and Lite ($4.80/min) (AA T2V; Gemini pricing). Sora 2 Pro at $0.50–$0.70/s (OpenAI pricing) is the most expensive official Western list price seen this run.
• Avatar minutes: HeyGen Avatar V at 48 credits/min on a $29 / 600-credit Creator plan is ~$2.32 per output minute — cheaper than Veo Standard per minute of finished talking-head, but a different product (script + look, not T2V) (HeyGen rate card; pricing). Runway Act-Two at 5 credits/sec is 2.4× cheaper in credits than Gen-4.5 (12/sec) for performance-driven character animation (Academy).

Avatar / character lane (one-line update only)

This thread’s May 2026 survey already covers Wan-Animate, Runway Act-Two, HeyGen Avatar V, MimicMotion, etc. New, cited facts only:

• HeyGen still ships Avatar IV/V as the flagship talking-head engines, now on a unified credit pool (HeyGen credits).
• Runway still sells Act-Two at 5 credits/sec and also meters Kling 3.0 Motion Control at 13–17 credits/sec (Academy).
• Wan’s general line moved from Wan 2.2 (this thread’s wan-2-2) to Wan2.7 (Apr 2026, five-character consistency claim) and Wan3.0 (Aug 2026, 30 s + document input) (Wan2.7; Wan3.0). That is not a re-benchmark of Wan-Animate.

Practical pick list (honest, dated)

As of 2026-08-18, for a reader who wants a default stack:

1. Default generate + iterate: Gemini Omni Flash — top AA T2V Elo, official ~$0.10/s, conversational edit (AA; pricing; DeepMind).
2. Open-weight / self-host (partial): MiniMax H3 — near-tied Elo, weights public; hosted IR + 2K regen still closed (MiniMax; HF).
3. Still-image animation: Seedance 2.0 720p — AA I2V #1 at $9.07/min (AA I2V).
4. Longest hosted shot / doc-to-video: Wan3.0 — 30 s, $0.10–$0.20/s; no AA rank yet (Wan3.0).
5. Scripted talking head: HeyGen Avatar IV/V on Creator/Pro credits (HeyGen).
6. Drive a character from a performance video: Runway Act-Two at 5 credits/sec (Academy).
7. Budget mid-pack: Veo 3.1 Lite or grok-imagine-video (AA T2V).

Contradictions and open questions

• Runway Gen-4.5 “#1 Elo 1,247” (Dec 2025) vs current AA table. Vendor page vs live leaderboard disagree. Either Gen-4.5 left the “current models” filter, Elo was rescaled, or the model fell out of the top 28. Unresolved without AA’s historical series.
• HeyGen credit rates. Pricing-page FAQ (3 / 20 cr/min) vs rate-card article (4–48 cr/min by engine and look). Prefer the rate card; confirm in-product before budgeting.
• Wan3.0 vs AA. Official 30 s / $0.10–$0.20/s is real; independent Elo is missing. Do not rank it above Omni Flash / H3 / Seedance on quality.
• Sora consumer/API sunset. Official pricing page still sells Sora 2; the discontinuation article did not fetch. Status of the consumer app is unverified in this file.
• Kling list prices. Official Kling blog 446; only AA + Runway credit meters.
• Local H3 hardware cost. Official docs do not publish a VRAM floor. Third-party “RTX 3060” claims on X news aggregates were not backed by permalinks and are not cited as fact.
• Avatar-specific independent benches remain open (independent-avatar-benchmarks, hand-fidelity-comparison-across-avatar-models).
• SEO comparison blogs (kursvideoai, technerdo, frankx, neuroselect, tech-insider) disagree with each other on clip length and sticker prices and were not fetched as sources.

Provenance

Rounds run: 3 of 3

Sub-questions by round:

Round 1 (broad survey):

1. What does an independent leaderboard say about current T2V quality and API $/min?
2. What are official Google / OpenAI / Runway list prices?
3. Which models are the current frontier (Omni Flash, H3, Seedance, Wan) vs the May 2026 avatar survey?
4. What open-weight options exist and what is actually open?

Round 2 (drill-down):

1. Confirm Gemini Omni Flash official price and product positioning — targeting the AA #1 that Grokipedia’s older Veo/Sora/Runway framing missed.
2. Image-to-video + avatar-lane prices (AA I2V, HeyGen, Runway Act-Two) — targeting the thread’s motion-mimic scope without re-surveying 2025 papers.
3. MiniMax H3 open-source boundary (what is not released).

Round 3 (resolve remaining uncertainty):

1. Official Wan 2.7 / Wan3.0 specs and Model Studio prices — targeting the Alibaba line already in this wiki.
2. HeyGen per-engine credit card vs marketing FAQ.
3. X discovery for primary claims — targeting --include-x.

Anchor source (Grokipedia, fetched before round 1):

• AI video generators — 25,017 chars extracted via _lib/grokipedia.py — taxonomy (generative vs template/avatar) and a 2025-era model list (Sora 2, Veo 3.x, Runway Gen-4.5, Kling, Luma, Pika, Synthesia, HeyGen). Used as vocabulary, not as the 2026-08 ranking.

X sources (--include-x; x-fetch skill + X MCP):

• 0 citable posts. WebSearch site:x.com returned no permalinks (three queries). X MCP search_posts_all returned HTTP 403 (“OAuth 2.0 User Context is forbidden… Application-Only”). X MCP search_news returned five story aggregates (x.com/i/trending/...) without individual status URLs — recorded here as discourse context only, not used as citations. Matches the SOURCE_RELIABILITY pattern from 2026-08-17 runs.

URLs fetched (14 successful, 5 failed):

Anchor:

• AI video generators — encyclopedia — field taxonomy.

Round 1:

• Artificial Analysis Text to Video leaderboard — independent Elo + API $/min.
• OpenAI API pricing — official Sora 2 / Sora 2 Pro per-second rates.
• Gemini Omni — Google DeepMind — official Omni positioning + vendor evals.
• [Failed: https://ai.google.dev/gemini-api/docs/pricing] — timeout on first attempt (succeeded in round 2).
• [Failed: https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation] — timeout (retried; still failed).
• [Failed: https://runwayml.com/pricing] — timeout.
• [Failed: https://help.runwayml.com/hc/en-us/articles/15124877443219-How-do-credits-work] — Cloudflare challenge interstitial.

Round 2:

• Gemini API pricing — official Omni Flash ~$0.10/s and Veo 3.1 per-second table (retry succeeded).
• MiniMax H3 open source — official architecture, duration, what is closed.
• Artificial Analysis Image to Video leaderboard — I2V Elo + $/min.
• HeyGen credit-based plans — official plan pools.
• HeyGen pricing — official sticker prices + FAQ credit rates.
• [Failed: https://klingai.com/blog/kling-video-3-0-credit-cost-guide] — HTTP 446.

Round 3:

• Alibaba Unveils Wan2.7-Video — official Apr 2026 suite.
• Alibaba Unveils Wan3.0 — official 7 Aug 2026 beta, 30 s.
• Wan3.0: 30-Second AI Video Generation — official Model Studio prices.
• How to use credits on HeyGen — official per-engine credit card.
• Runway Academy: Credits & Available Models — official in-app credit rates including Act-Two and third-party models.
• Runway Gen-4.5 announcement — vendor Elo claim (1 Dec 2025).
• [Failed: https://ai.google.dev/gemini-api/docs/video] — timeout (DeepMind + pricing pages already cover Omni vs Veo).

Tools used: WebSearch, WebFetch, grokipedia-fetch (skill / _lib/grokipedia.py), x-fetch (skill) + X MCP search_posts_all / search_news. Generated: 2026-08-18 13:20 UTC