Agnes Video 2.5 launched on the Pavo creative platform on August 26. The Flash variant in the lineup is entirely free — generation doesn't cost any credits.
The free tier's limits are drawn tightly: 720p resolution, 4-to-12-second clips, and 3 requests per minute. The paid Pro version runs on a credit system — the platform gives out 200 free credits a day, generation costs 2 credits per second, clips can run up to 100 seconds, and resolution goes up to 2K. Do the math and the free daily allowance works out to roughly 100 seconds of footage (200 ÷ 2, a rough estimate) — about enough for one short clip. The reference API price is $1.50 per minute.
What a 165-Point Elo Jump Actually Buys You
Agnes's own numbers show its score on the Artificial Analysis blind-test leaderboard climbing from an Elo of 917 to 1082 — a gain of 165 points.
It's easy to skim past an Elo number as just "higher is better," but the scale exists specifically to calculate win rates. Run the standard formula and a 165-point gap works out to roughly a 72% win rate — in a blind test, that means when the old and new versions generate clips from the same prompt and people pick a winner, the new version gets chosen about seven times out of ten. That's not a small margin; a jump that size usually signals a model generation change, not a patch-level tweak.
But leaderboard scores and real-world usability are two different things. Elo measures viewers' preference for a single clip — it doesn't capture how fast prompt adherence degrades over longer shots, or whether characters stay consistent across successive generations. Those are the two things that actually trip up anyone trying to make a short drama.
The Platform Takes Over Storyboarding Too
Pavo's real product logic isn't the model itself — it's that the whole production pipeline gets broken into steps. A user provides a one-line initial brief, and the platform automatically generates the fleshed-out requirements, an outline, character/scene/prop designs, the corresponding concept art, storyboards, and keyframes, with the actual shot-by-shot video coming last. The user confirms each step before the platform moves to the next.
The practical payoff is cost savings. Generate a finished clip in one shot and any misstep anywhere in the process means redoing the whole thing — with the cost landing entirely on the most expensive step, video generation itself. Break it into checkpoints instead, and a problem only sends you back to that specific step for a rerun, while everything already confirmed upstream stays untouched. For a free tier capped at 100 seconds a day, that difference determines whether you can actually finish something or not.
How Long Can the Free Tier Last
Video generation is the most expensive generative modality per unit of output — the number of frames needed for one second of footage isn't remotely in the same league as generating a bit of text. "Completely free" under those economics is almost always a customer-acquisition window.
Taken together, the three limits lock the cost ceiling down tightly: 720p caps the per-frame compute, 4-to-12 seconds caps the length of each clip, and 3 requests per minute caps concurrency. Multiply them out and a single user's theoretical maximum output is about 180 short clips an hour — all of them low-resolution. The platform clearly knows exactly what it's subsidizing, and exactly where that subsidy's ceiling sits.
The capability gap between Flash and the paid version already marks exactly where the payment wall sits: 720p versus 2K, 12 seconds versus 100. Anyone who wants to make something actually releasable will eventually have to cross it. Pricing like this usually plays out in three stages — free to drive activity first, then tighten the free allowance, then move the genuinely useful capability behind the paywall. Right now it's still at stage one.
One piece of information isn't available yet: 3 RPM is the API's rate limit, but actual queue times depend on how much concurrency headroom the platform's backend has left, and no official figures have been given — users will have to find out by testing it themselves.
Sources: Pavo platform announcement, QbitAI, CocoLoop, KuCoin News; free-tier limits, credit costs, and API pricing were cross-checked across sources, and the win-rate conversion is an editorial estimate.