What Creatify's Boreal-H3 reveals about measuring AI in production

What Creatify actually shipped

On 15 September 2026, Creatify — a company whose business is making video ads — launched Boreal, an AI video model trained specifically for advertising. The version now circulating as Boreal-H3 is post-trained on top of MiniMax’s open H3 model using real advertising projects, and it ships alongside an “Ad Agent” that turns a single instruction into a finished spot.

The capability list is deliberately narrow. The product in the shot keeps its exact label, shape, and text as the camera moves around it. The person on screen reads as one continuous creator, including two-person dialogue. The operator directs camera moves and keyframes. None of this is a general-purpose video demo; every claim is tied to something an ad has to get right.

The metrics are the story

The launch numbers are unusually specific. Creatify reports brief success — the share of jobs that produced a usable ad — rising from 28% to 50%. Identity match, the share of clips where the creator and the product survive intact, moved from 83% to 94%. Failure severity fell 71%. In blind side-by-side review, evaluators preferred the post-trained model to its untouched base in 81% of decisive comparisons across 40 production cases.

Read those figures together and a different picture emerges than “a better model shipped.” The company first defined what failure means for its own use case — a face swapped mid-shot, a label melted into nonsense, a product that changes shape halfway through — then built an evaluation loop that measures those specific defects and feeds the result back into training. That is the discipline enterprise teams keep asking about and rarely implement: not a leaderboard score, but a named failure taxonomy, a baseline, and a measured delta.

The economics are the second story

Boreal is priced per second of finished video at roughly one cent — about $0.05 for a five-second clip — and it renders in realtime, so the clip is ready in roughly the time it takes to watch it. Against Seedance 2.0, Creatify reports a median of about 198 seconds and $0.60 for the same five seconds. Against Seedance 2.5’s early-September list price, it claims to be roughly 47 times cheaper. A storyboard of twenty five-second shots lands near a dollar and finishes in under two minutes.

For decision-makers, that is the more consequential shift. When generation cost and latency fall by one or two orders of magnitude, the binding constraint on adoption stops being “can the model do it” and becomes “how many variants can we afford to test.” Cheap generation that still takes minutes changes nothing about how a team works. Cheap generation that returns in seconds changes the workflow itself.

What the numbers do not settle

Three caveats belong next to any launch-day benchmark. First, the comparisons are the vendor’s own testing on identical inputs; independent reproduction is the test that matters, and it has not happened yet. Second, on raw fidelity the gap narrows — a 1080p flagship still renders those inputs at higher resolution than Boreal’s 720p-class output, so resolution-sensitive measures read in the competitor’s favor. Creatify says so plainly, which is itself a signal worth noting. Third, the model is strongest on short clips and simple scenes; on single-person talking clips it improves least over its base, at 60%.

None of that disqualifies the release. It does mean the honest read is “a strong quality-per-cost point in a specific niche,” not “the best video model.” Vendors who publish their own weak spots are easier to verify and easier to trust.

Four questions for your own AI program

The Creatify launch is a useful mirror. Four questions are worth asking of any AI initiative:

1. What does failure mean here, specifically? Can you name the defects that would make output unusable — the way Creatify named label drift and identity swap — before you go looking for a score?
2. What is your baseline, and what is your measured delta? A number without a “before” is a demo, not evidence.
3. Is the cost model predictable before the job runs, or reconstructed afterward from a credit balance?
4. Does the latency change the workflow, or only the invoice? Realtime generation lets a team test more variants; batch generation only reduces the spend.

Where this leaves enterprise adopters

The interesting part of Boreal-H3 is not the video. It is that a domain-specific model beat its general-purpose base only after the team codified what “good” and “broken” meant for one narrow task, then measured both. That pattern travels — finance, healthcare, legal, operations. The teams getting durable returns from AI are the ones that turned their own judgment about quality into something a model can be scored against.

The work is unglamorous. It is taxonomy, baselines, and review loops — the same governance discipline that decides whether a pilot survives its first quarter in production. It rarely makes the launch post. It is usually the reason the launch works.

If your organization is moving an AI initiative from pilot to production, the right time to define what failure looks like is before the deployment, not after the first incident review. Book a strategy consultation to map the evaluation and governance model your program needs.