Boreal-H3 at Four Cents a Second, and the Launch Price Holding That Number Up

Creatify put Boreal-H3 live on September 30, 2026, and the figure its research post leads with is not the price. It is brief success on controlled advertising shots rising from 28% to 50% — measured by Creatify, on Creatify's own briefs, against the MiniMax H3 base the model was post-trained from. The other half of the launch is an agent: Creatify Ad Agent, the redesigned Creative Agent, which the company says runs on Claude Opus 5.5. Two models, two different meters — one bills per second of finished video, the other per token of reasoning — and that split is the most useful thing to understand about this release.

So is the asterisk on the headline price. The four cents is a launch offer, not a list rate, and it rests on a comparison the same post is careful to state outright.

What the model actually is

Boreal-H3 is the top of the Boreal family. Where Boreal was built to make advertising video cheap enough to generate at volume, Boreal-H3 is aimed at the shots that carry a campaign: the product has to keep its exact label, the creator has to stay the same person, the camera has to go where the brief says, and the action has to actually happen. Creatify describes it as a post-training job on MiniMax H3 rather than a new backbone — the same Recursive Self-Improvement loop it ran for Boreal, retargeted at reference fidelity, identity and brief execution.

That lineage is the structurally interesting part, because MiniMax H3 is a general omni-modal video model that reads text, image, video and audio references as one context and generates 4–15 second clips with native stereo audio. Boreal-H3 inherits that input surface and narrows the training target. It ships through three entries in Creatify's Model Playground:

• Text to video — prompt only, six aspect ratios, $0.04 per second at 768P, $0.12 at 1088P, $0.26 at 2K.

• Image to video — an opening frame plus an optional ending frame, same three rates.

• Reference to video — up to 9 images, 3 video clips and 3 audio clips combined in one shot, same three rates.

Creatify states the model is its own and that it serves the renders itself, which is what allows a per-second rate and what makes the reference workflow possible at all — a multi-reference shot is a different product from a prompt-and-pray clip.

The metrics, and who ran them

Everything below comes from Creatify's own Boreal-H3 page and its technical report, both dated September 2026. No third party has reproduced any of it. Treat the numbers as the vendor's description of its own model, which is what they are.

The central comparison is against the H3 base, on what Creatify calls difficult production briefs. Brief completion rises from 27.8% to 50.0% — the 28%-to-50% figure the launch leads with, rounded the same way on both ends. Subject consistency, which the page reports as identity match, goes from 83.3% to 94.4%. Visible defects fall about 70%, from 1.11 to 0.33 per clip. Against six frontier video models on a joint pass rate normalised so that MiniMax H3 equals 100, Boreal-H3 scores 133.3 on what the report calls an Ads quality index, with 85.3% reference fidelity. The underlying pass rates are given as 35.3% for Boreal-H3 and 26.5% for MiniMax H3, so the index is a ratio of two vendor-measured rates, not an absolute score.

Two things about that evidence are worth stating plainly. First, it measures Boreal-H3 against its own base and against comparators Creatify selected — it is not an independent bake-off. Second, the base it improves on is itself a model the same company post-trained, and the 83% starting point for identity match is the same figure Boreal's launch used a fortnight earlier. The two models are being measured against the same floor.

The price, and the launch offer holding it up

Four cents a second is real and it is what Creatify is charging today. It is also explicitly a launch offer: Creatify says the $0.04/s is 50% off MiniMax H3's $0.08/s 768P list price, and the pricing panel shows the $0.08 struck through beside it. That distinction matters more than the number does. At list, Boreal-H3 would be $0.08 a second — not four times Boreal's rate but eight, and level with the base model it was trained from. A team budgeting a campaign against the launch rate is budgeting against a discount whose end date is not published.

The rate card is otherwise clean and per-second, which removes the credit arithmetic that per-credit video APIs force on you. The technical report also puts a latency figure next to each price: Boreal-H3 is measured at 1.10 generation-seconds per video-second at 768P, against 32.29 for MiniMax H3 and 46.50 for Seedance 2.5, with the comparator timings drawn from recorded API measurements and provider-published speed reports. A ten-second clip at that rate is roughly eleven seconds of generation. As always with a vendor latency claim, the interesting question is not the ratio but the queue behind it.

For contrast, the report's own comparator set at 720p or 768p: Seedance 2.5 at roughly $0.47 per second, Wan 3.0 Prime at $0.14, Omni 1.1 Flash at about $0.10, and MiniMax H3 at $0.08. Boreal-H3's $0.04 is the cheapest line on that table by a factor of two and, at list, it would not be.

The half of the launch that bills per token

Creatify's technical report says the same release revamps its Creative Agent into Creatify Ad Agent, connecting the stronger video model to the whole ad workflow: develop an idea, produce variations, review the outputs, turn more creative directions into experiments. The model report is specific about what drives it — Claude Opus 5.5 — and that is the half of this launch whose meter is tokens rather than seconds.

That distinction is not pedantry. A campaign's cost is not the cost of its clips. It is the clips plus the brief expansion, the hook and script variants, the compliance pass, the caption set, and the routing logic that decides which cut runs where — and that second pool grows with the number of variants, not the number of finished videos. On an agentic pipeline where one brief fans out into dozens of small model calls, per-token billing is the line that moves.

The useful fact here is that Opus 5.5 is a model we do serve. It is live in our catalogue at $4.00 per million input tokens and $20.00 per million output tokens, with cached reads at $0.20, served on both the Anthropic Messages API and the OpenAI-compatible Chat Completions endpoint. If you want to run the agent-style half of an advertising pipeline — the scripts, the variants, the review logic — that is a call you can make through us today at provider list price with 0% markup, on one key, behind automatic failover. Boreal-H3 is not: OrcaRouter does not serve video generation models, and this article is not going to imply otherwise.

What the head-to-head does and does not show

Boreal-H3's page runs matched-brief comparisons against Seedance 2.5 and Kling 3.0, with the same reference images. Creatify is unusually explicit about method: the two comparators were run through their public APIs, Boreal-H3 is shown from Creatify's own served renders at 1088P, and no clip was edited. Resolution parity is therefore not held, and the page says so rather than burying it.

The render-time example it gives is concrete and checkable in one direction only: a fifteen-second image-to-video clip of a squishy toy took 84 seconds on Boreal-H3 and 411 seconds on Seedance 2.5. That is a vendor measurement of a vendor pipeline against a competitor's API, and it lines up with the 46.50 s/s figure in the technical report. It is also the kind of claim that only an independent run can settle, and none exists yet.

The qualitative pairs are worth more than the timings. Creatify's stated failure modes are the ones an ad team actually loses money on: a bottle that changes shape, a label that turns into gibberish, a creator who becomes someone else between frames, a product demonstration where the action never happens. Those are also the four things the model's own six-point features list claims to fix — product fidelity, first-take usability, complex spoken lines, and directable camera. A reader should notice that the failure list and the feature list are the same list. That is coherence in the pitch, not evidence of it.

What would settle it

Three things, none of them another vendor comparison. The first is an independent evaluation — Boreal-H3's entire evidence base is a vendor-run internal test on production briefs plus a matched-brief showreel, and no third party has run the model. The second is whether the launch rate holds: the $0.04/s is defined as half of the base model's list price, so the honest question for anyone planning a quarter of spend is what the rate is when the offer ends. The third is whether the reference workflow survives at scale — nine images, three videos and three audio clips in one shot is a genuinely unusual input budget, and multi-reference conditioning is exactly the capability that tends to degrade when it is pushed outside the curated demo set.

The defensible summary: the interesting claim here is not that advertising video got cheaper — Boreal already did that a fortnight ago at a cent a second. It is that the same lab shipped a second, four-times-pricier tier and pointed it at a specific class of shot, with a control surface the cheap tier does not have. Whether that tier is worth its rate is a question its own report cannot answer, because the report was written by the people selling it.