Pick Your First AI Video Tool, and Learn to Read the Leaderboard

The news this week is mostly about what didn't ship. As of today, nothing brand new at the frontier of video generation landed in the last seven days. The top of the board got set back in the February-through-April stretch, and right now the field is in a settling phase, models getting wider availability rather than a fresh headline drop. So the live story is the reshuffle on the leaderboard, and two moves worth your attention: a stealth model that's now sitting at number one, and a tool a lot of you may still have wired into a pipeline that's officially going away.

Start with the model at the top. Alibaba's research group quietly put a model called HappyHorse on the Artificial Analysis Video Arena back in early April, with no name attached, and it went straight to the top of both the text-to-video and image-to-video blind tests before they revealed themselves, as Bloomberg reported on the tenth. The specs that matter to a shot: up to 1080p, fifteen seconds max, and fully synchronized native audio, meaning lip-synced dialogue, ambient, and music generated together with the picture in a single pass. Now the catch, because they pitched this as open weights. The license release of the weights was announced, but as of late May the download page is still locked behind a sign-in and there's no actual repository up yet, so for now it's something you call through an API, not something you run on your own machine, per reporting from WaveSpeed. If you want to try it, it's live on fal at fourteen cents a second for 720p and twice that for 1080p, so a full fifteen-second clip at the higher resolution runs you about four dollars and twenty cents. Keep an eye on Alibaba's Wan video repository for the day the real weights actually drop.

Second, and this one's a calendar item: OpenAI is winding Sora down. The Sora app and website were already discontinued back on the twenty-sixth of April, and the Sora 2 models and the video API behind them shut off on the twenty-fourth of September. Sora 2 still sits around tenth on the with-audio board, but it's a dependency with a hard expiration date now. If you've got it in a pipeline, the move is to port over to Kling, Veo, or Seedance before September. Bloomberg named those same generators as the replacements people are landing on.

On value, ByteDance's Seedance 2.0 is still the cheapest credible frontier option and sits at number two overall. It does up to 1080p, four to fifteen seconds, native synchronized audio, and at 720p a five-second clip lands around five cents through third parties, which is a different universe of pricing from where Sora was. ByteDance has reportedly confirmed a Seedance 2.5 for mid-year aiming at 4K and near-real-time generation, so that's the next thing to watch on the value tier.

The native-audio race is the real theme underneath all of this. Right now you've got at least four models generating synchronized dialogue and effects in the same pass, not dubbed on afterward: Kling 3.0, Veo 3.1, Seedance 2.0, and the new HappyHorse. Two newer names worth a bench if audio is your bottleneck: SkyReels V4 from Skywork claims tight audio-event alignment, under about a tenth of a second, and PixVerse V6 ships more than twenty cinematic lens controls, focal length, aperture, depth of field, the kind of camera language that actually changes a shot.

Now the leaderboard check, by job, with the usual caveat: this is a snapshot, it reshuffles within weeks, so verify it live before you quote it. For overall text-to-video prompt adherence, HappyHorse and Seedance 2.0 are in a dead heat at the top. For native audio and dialogue in a narrative, Veo 3.1 is still the one cited for true forty-eight-kilohertz dialogue, and Kling's Omni variant handles multi-shot sequences on a shared audio timeline in five languages. For image-to-video it splits by tab: Seedance 2.0 leads the with-audio board, HappyHorse leads without audio. And on value per second, Seedance at 720p is the floor, with HappyHorse the cheap frontier-quality pick. The thing that moved this cycle: HappyHorse's April debut knocked Veo and Kling down the with-audio table and now splits the top spot with Seedance. Your standing action, same as every week, open the split Arena tabs, text-to-video, image-to-video, and editing, each with and without audio, and don't trust a single ranking until you've checked the right one.

Pick Your First Tool and Learn to Read the Leaderboard

Welcome to episode one. Here's the whole show in one sentence: we're going to climb a single ladder, from typing a prompt into a website and taking whatever it hands you, all the way up to running a one-person studio where a brief comes in and a finished, on-brand cut comes out while you sit back and art-direct. That's the destination. Today is the first rung, and it's a small one on purpose. Pick a tool, get a clip out of it fast, and learn the one skill that'll still be useful a year from now: reading the leaderboard and testing tools on your own work.

I want to start with that last part, because it's the thing most people get backwards. When you're new, the temptation is to find out which AI video model is the best one right now, learn it, and go. The trouble is the answer changes about once a month. To give you a sense of the pace: back in early February of 2026, Kling 3.0 and Seedance 2.0, two top-tier models, launched within a single week of each other, and whatever had been on top the week before got bumped. So if your skill is "I know the best model," that skill has a shelf life of about thirty days. If your skill is "I know how to read the board and test the top few on my own shot," that one survives every swap. We're building the second one. Everything else today serves it.

Before we touch a tool, one mental adjustment, because it's the difference between people who quietly burn fifty dollars in an afternoon and people who don't. Stop thinking about cost per generation. Start thinking about cost per finished clip. Here's what I mean. A tool tells you a clip costs fifty cents to make. Feels cheap. But AI video is non-deterministic, which is a fancy way of saying you run the same prompt twice and get two different clips. So you generate, it's not quite right, you run it again, and again, and eight tries later you've got one keeper. That keeper didn't cost fifty cents. It cost four dollars. The sticker price is per roll of the dice. Your wallet pays for the take you actually use. Hold onto that, because every habit I teach today is really about keeping that one number, cost per keeper, under control.

Okay. Let me get a handful of words out of the way, gently, because I'm assuming some of you are brand new and I don't want to lose anyone on vocabulary. We'll go deeper on all of these in later episodes. Right now you just need a working definition that lets you make a clip.

A prompt is just the text you type describing what you want to see. A good beginner video prompt has about five parts: the subject, what it's doing, where it is, the lighting, and the camera move. Something like, a barista slides a latte across a marble counter, warm morning light through a window, slow push-in. Subject, action, environment, light, camera. That's the whole recipe for now, and we'll spend a full episode on it soon.

There are two ways to make a clip, and this difference matters more than anything else today. Text-to-video means you type a prompt and the model invents the entire clip from nothing. Image-to-video means you start from a still picture, one you uploaded or generated, and your prompt only describes the motion. The image locks the look, the character, the colors, the framing, and the prompt just animates it. Here's the takeaway, and I'll repeat it all season: if you care what the subject actually looks like, start from an image. Text-to-video is rolling the dice on appearance and motion at the same time. Image-to-video pins the appearance and only gambles the motion. That's far more control, for free, on day one.

Let me make that concrete, because it's the lever everything else hangs on. Say you want a clip of your own product, a particular bottle with your label on it. With text-to-video, you describe the bottle in words and the model invents a bottle, close-ish, but the label's gibberish and the shape drifts every time you run it. With image-to-video, you hand it a clean photo of the actual bottle and just say, slow rotate, soft light coming up. Now the bottle is your bottle, the label is your label, and the only thing the model is deciding is how it moves. You've taken the hardest, most failure-prone part, getting the look right, off the table entirely. That's why nine times out of ten, when the look matters, you start from an image. We'll spend a whole episode on the keyframe trick that makes this sing, but the instinct starts today.

A seed is a number that sets the random starting point of a generation. Same prompt plus same seed gets you roughly the same clip back. Change the seed and you get a fresh take on the same idea. You'll reach for this the moment you get a clip that's almost right: lock the seed, change one word in the prompt, instead of re-rolling blind and praying. The mistake beginners make is treating every regeneration as a fresh coin flip. Once you've got something close, the seed is how you stop flipping and start steering.

Aspect ratio is just the shape of the frame. Widescreen, the wide horizontal shape, is for YouTube and web embeds. Vertical, the tall shape, is for TikTok, Reels, and Shorts. Square sits in between for some feed posts and certain ads. Pick the one your final destination wants. One rule will save you a lot of grief: pick your delivery shape first and generate straight to it. Don't make a widescreen clip and crop it vertical later, because you'll throw away half your frame, usually the half you cared about. And if you're doing image-to-video, make your source image the same shape as the video.

And credits. Almost no consumer tool charges you dollars per clip directly. They sell you credits, an in-app currency, and each generation spends some credits depending on resolution, length, and quality mode. The trap is that credits hide the real per-second cost, which is exactly why that cost-per-keeper habit matters so much.

A few more words you'll hear me use and then set aside for later. A keyframe is a pinned frame the motion moves toward. A LoRA is a small add-on file that teaches a model one specific character or style. A color grade is adjusting color and contrast so different shots feel like one world. And an NLE, a non-linear editor, is the timeline software where you assemble clips into a finished piece. We'll meet every one of those properly down the line. Episode one, you don't open an editor at all.

While we're naming things, let me point at the software you'll hear about across the season, so the ladder's visible even though we're not touching most of it today. DaVinci Resolve is a free, genuinely professional all-in-one editor, color tool, and effects suite, and it's our default for finishing later in the show. CapCut is a free, fast editor built for vertical social video, with some AI automation built in, and it also happens to be where ByteDance's video tools live. Premiere Pro is Adobe's paid, industry-standard editor, for those of you already in their world. For sound, ElevenLabs does voices, sound effects, and music, and Suno and Udio do music. File those away. Today, all we're doing is making one clip.

Getting your first clip out

So let's actually get a clip. Here's the copyable path, and it's deliberately boring.

The fastest free start right now is Google Flow, which is Google's creative front-end sitting on top of their Veo model. Every personal Google account gets somewhere around ten free generations a month, no credit card, plus a small monthly allotment of credits that refreshes on the first. Clips run up to eight seconds, and the free tier renders at 720p, with the sharper resolutions needing a paid Google plan, per the breakdowns of Flow's free access. That is more than enough to learn on. If Google's not your thing, the principle is identical on any of the front-ends I'll list later. Pick one. Don't agonize.

Step one once you're in: choose text-to-video or image-to-video. If you don't much care about the exact look, type a prompt and go, that's text-to-video. If you need a specific face, product, or composition, make or grab a still first and animate it, that's image-to-video, and it'll behave for you far more reliably.

Step two: write the prompt using those five slots. Subject, action, environment, lighting, camera move. Resist the urge to write a paragraph. One clear beat, described concretely.

Step three, and people skip this and regret it: set your aspect ratio and duration before you generate, not after. Pick vertical for social or widescreen for web, and pick the shortest length that tells the beat. Most tools default to somewhere between five and eight seconds. You can extend a clip later, but every extension is another generation and another credit hit, so don't reach for length you don't need.

Step four: generate, and then watch the result like a director, not like a proud parent. The first clip back always has a little dopamine glow to it, it moved, it looks like a movie, and that glow will talk you into accepting junk. So score four things, fast and cold. One, the motion: is it natural, or does it warp and melt? Watch for things sliding when they should be planted, or a limb that smears between frames. Two, prompt adherence: did it actually include what you asked for, or did it quietly drop the marble counter and the push-in? Three, the danger zones, faces, hands, and any text in frame, which are where AI video still falls apart most often. Count the fingers. Read the sign on the wall. If a hand has six fingers or the label is alien glyphs, that's a reshoot, not a keeper. And four, value per second: was that worth the credits it cost, and would a cheaper tool have gotten you there? Run that four-point check in about ten seconds and you'll make better calls than someone agonizing for ten minutes.

Step five: iterate on purpose, not on reflex. If it's close, this is where you lock the seed and change one variable, or nudge one phrase in the prompt. And cap yourself. Three rolls per shot, then you either take the best one or you change your approach. That cap is the single most important habit in this whole episode, and I'll come back to why in a minute.

How to actually read the leaderboard

Now the durable skill, the reason you'll outlast every model fad. Let me teach you to read the Artificial Analysis Video Arena, which is the live ranking the whole field watches.

Here's how it works under the hood. Someone submits a prompt, two models generate from that same prompt, and the two clips are shown side by side with no labels, so you can't tell which model made which. You vote for the better one. Thousands of those blind votes feed an Elo rating, the same rating math chess uses, where a model gains or loses points based on whether people prefer its output head to head. So the number next to a model isn't a spec sheet. It's an aggregate of human preference across a lot of matchups.

The most important thing to learn is that it's not one list, it's several, and reading the wrong one gives you a wrong shortlist. There's a text-to-video board and a separate image-to-video board, because those are different jobs. And then each of those splits again, into a "with audio" arena and a "without audio" arena. Those are genuinely different rankings with different leaders. That split exists because of the big shift in 2026: older models hand you a silent clip and you add sound in post, while newer ones generate native synchronized audio, dialogue and footsteps and ambience baked in and lip-synced, in the same pass. Google's Veo generates its audio natively as part of the model, and Seedance 2.0 ships synchronized audio too. So if you need a character who talks, read the with-audio board. If you're going to score the clip yourself in an editor anyway, the without-audio board is the fairer read on pure motion and picture quality. Same model, very different rank depending on which tab you're standing on.

One more reading skill, then we'll use it. The board shows a confidence interval next to each score, a margin of error. Two models a handful of points apart are a statistical tie, not a winner and a loser. So don't agonize over rank three versus rank four. Treat the whole top cluster as your candidate pool, and let your own testing break the tie. And know that the board moves constantly. New models get added, the top reshuffles. Any standings table, including the one I'm about to read you, is perishable.

So what you're actually doing on the board is building a shortlist, not finding an answer. The motion goes like this. Decide your job first: am I making a talking character or a silent shot, and am I starting from text or from an image. That picks your tab, one of four. Then look at the top cluster on that tab, the models bunched together at the top inside each other's margins of error, and grab the top two or three. That's it, that's all the board is for, handing you a short candidate list for the job you're doing. The board does not tell you which one wins for you. It tells you which two or three are worth your ten-minute test. People treat it like an oracle and get burned. Treat it like a filter and it's the most useful page in the field.

A snapshot, which is already going stale

So here's roughly where things stand in early June of 2026, and I'm telling you the numbers mostly so you can practice reading them, not so you'll memorize them. By the time you hear this, expect it to have moved.

On the text-to-video board with audio, HappyHorse from Alibaba and Dreamina's Seedance 2.0 are essentially tied at the top, both around an Elo of twelve hundred, with SkyReels V4 and a couple of Kling 3.0 variants filling out the next slots. Flip to the without-audio version of the same board and HappyHorse jumps to around thirteen-sixty, a very different-looking number for the same model, which is exactly the point about tabs. On image-to-video with audio, Seedance 2.0 leads, HappyHorse is just behind it, then Veo 3.1, xAI's Grok video model, and PixVerse V6. And notice the open-weight options, like Lightricks' LTX, sit lower but they're on the board, which matters later when you want something you can run yourself.

Here's a detail that makes the lesson for me. Even within the same month, separate trackers disagree, one general explainer had Veo on top for text-to-video while the Arena had HappyHorse. That disagreement isn't a bug. It's the whole reason you don't outsource the decision to a single table.

The pitfall you will actually hit

Now the trap, and I've got a perfect story for it. Remember HappyHorse, the model that's been topping the board. When it first showed up in early April, it appeared anonymously, just a mystery name, "HappyHorse," with no lab attached, and it rocketed to number one on both the text-to-video and image-to-video boards, beating Seedance outright. People went a little wild. And then, within days, both versions vanished from the public leaderboard, with no official explanation, while everyone argued about whose model it secretly was, as one analysis of the episode laid out. Now, it turned out to be Alibaba, and it came back. But sit with the lesson: a stealth, unattributed model can take the top of the board with zero accountability, and a name sitting at number one today might be unbuyable, or simply gone, next week. If you'd torn up your workflow that week to chase the new leader, you'd have rebuilt around a ghost.

That's the first half of the pitfall, chasing the monthly leader. The second half is quieter and it'll cost you actual money: blind regeneration. Because the output is random every time, it is genuinely easy to sit there re-rolling the same prompt, hunting for the perfect take, and burn through twenty dollars of credits without noticing. The fix is the cap I mentioned, and now you know why. Three rolls per shot, then stop. When you're close, change the prompt or lock the seed and change one thing, don't reroll into the void. And make peace with over-generation as a normal cost, not a failure: for a three-minute short you might generate three hundred to five hundred seconds of raw footage to keep about a hundred and eighty. That ratio is fine. Spending your whole balance to make one shot perfect is not.

And the third piece, the one that sets the whole trap: those gorgeous launch trailers. A vendor's launch reel is the best handful of clips out of thousands, hand-picked. Here's how you catch it: if your first three honest attempts look nothing like the trailer, the trailer was curated, not typical. Which is the perfect cue for the antidote.

How to bench, which is the real job

The antidote to all of it is benching, and it's simple. Don't ask the internet which model is best. Ask your own work.

Pick one representative test shot, a single shot that looks like the work you actually do or sell. If you do product video, that's something like, product on a turntable, soft studio light, five seconds, vertical. If you do narrative, it's two characters talking. The leaderboard averages over thousands of prompts that aren't yours, and the Arena's own people will tell you the rankings should guide your decision, not replace testing on your specific workflow.

Then take that exact shot, the same prompt, same aspect ratio, same duration, and run it across the current top two or three from the correct tab, the one that matches your audio need and your text-versus-image choice. Hold everything constant except the model. And score them on the jobs that matter to you: motion quality, did it follow the prompt, did the faces and hands hold up, and value per second, which is just cost divided by the seconds you'd actually use.

Let me walk one through, because it's faster to do than to describe. Say I make short product spots, so my test shot is, a perfume bottle on a slow turntable, soft studio light, five seconds, vertical. I open the image-to-video with-audio board, take the top two, let's call them the current leader and the one in third, and I generate that identical shot on both, same source photo, same words, same five seconds. Tool A gives me beautiful light but the bottle subtly warps as it turns, which on a product shot is fatal. Tool B is a touch flatter but the bottle stays rock solid, and it cost me half as much. For my work, Tool B wins, even though it sat lower on the public board. If I were doing dialogue scenes instead, the warp wouldn't matter and the lighting might, and Tool A would win. Same two tools, opposite call, decided by my shot and not by a stranger's average.

Do this and you'll routinely find that the winner for your work is sitting at rank four on the public board. That's not the board being wrong. That's you having a board of one prompt, the only one that's actually about you. And here's the quiet payoff: you only have to bench when the board moves, which is monthly, and re-running one shot across two tools costs you maybe a dollar and ten minutes. That's the cheapest insurance in this whole business.

Where to start, and what it costs

Let me give you the quick tour of entry points, so you can pick one today and know roughly what you're spending. Two kinds: direct front-ends, one model family each, and aggregators, many models behind one login.

On the direct side: Google Flow, which I already walked you through, is the fastest free start, around ten free generations a month and eight-second clips at 720p. Runway has the strongest creative-control tooling and gives you a one-time pot of free credits to try, then plans start around fifteen dollars a month. Luma's Dream Machine offers a free monthly batch, watermarked and at draft quality. Kling, from Kuaishou, hands you a refreshing daily allotment of free credits, enough for a few short clips a day, but the free output is watermarked, so it's for learning, not delivery, with paid plans starting around seven dollars a month. Hailuo, from MiniMax, is similar, a few free watermarked clips a day. And Seedance 2.0 lives inside Dreamina, which is part of CapCut, so if you're already editing there it's right next to your timeline.

One important currency note that doubles as today's lesson in miniature: you can no longer just go to the Sora website. OpenAI shut the Sora app and site down in late April, and only the developer API is hanging on into September. So a tool that was the household name a few months ago isn't a place a beginner can visit anymore. That's exactly why we don't anchor our skills to any one tool.

On the aggregator side, these are your benching multitool, because you test many models under one credit pool instead of making three separate accounts. Higgsfield bundles fifteen-plus models, Sora, Veo, Kling, Seedance, and more, from around five dollars a month. Krea is a creative suite with dozens of models plus image and upscaling tools under one subscription. And fal and Replicate are pure pay-as-you-go, no subscription, priced per output or per GPU-second, which leans a little more technical but is the cleanest way to pay only for what you generate. My beginner guidance: use a direct free tier like Flow or Kling to get your very first clip, then use one aggregator to bench the top two or three without signing up three times.

Let me put real numbers on the cost-per-keeper idea, because it changes how you budget a job. Imagine a thirty-second social spot, which is maybe six shots of five seconds each. On a mid-tier tool at roughly a dime a second, each five-second generation is about fifty cents. If you nail every shot in one go, that's three dollars, lovely. But you won't. With my three-roll cap, figure two or three rolls per shot on average, so call it eight to twelve dollars of generation for the finished thirty seconds. Without the cap, chasing perfection, that same spot quietly becomes thirty or forty dollars, and you didn't feel it happen because each roll was only fifty cents. The discipline isn't about being cheap. It's about knowing, before you start, that a thirty-second spot costs you about ten dollars and an hour, so you can quote a client and not lose your shirt.

Two money facts to leave you with. First, watch the gap between sticker price and real cost. A clip might list at fifty cents to five dollars depending on model and resolution, but amortized over your regenerations and a subscription, the effective cost per keeper on something like Runway can land closer to a dime or two. The space between those two numbers is set entirely by your regeneration discipline, not by the price tag. That's the cost-per-finished-clip idea, made concrete. Second, watermarks are a hidden tax: most free tiers stamp the output, which makes those clips useless for client delivery. So treat free tiers as where you learn and bench, and budget a cheap paid tier for the work you actually hand over.

Where that leaves you

So that's rung one. One topic: pick a tool, get a clip, learn to read the board. One copyable workflow: open a free front-end like Google Flow, choose text or image to video, set your aspect ratio and duration first, write a five-slot prompt, generate, score motion and adherence and faces and value, and cap yourself at three rolls. And one pitfall you'll absolutely hit, the double trap of chasing this month's leader and torching credits on blind re-rolls, which you now recognize the moment you've spent more than three tries on a shot or you're switching tools over a leaderboard you never tested yourself.

Next time we slow down and pull apart that prompt, the real anatomy of a shot description, subject and action and camera and lens and light, so the clip you get is the one you pictured instead of a happy accident. For now, go make one clip, badly, today. The whole ladder starts with that.