What 40 Seconds and 4K actually mean in AI video

AI video spec sheets lead with two numbers — clip length and resolution — and both are routinely quoted as if they measured one thing. They don’t. “40 seconds” can mean one continuous generation pass or four chained extension rounds. “4K” can mean the pixels a model actually renders, a separate upscale pass, or a third mechanism that is neither. Where a vendor names which mechanism applies, it does so on its own page; most coverage flattens the result into a single headline figure.

The distinction is not pedantry. A 40-second cumulative ceiling reached through extensions behaves differently from a 30-second single pass: every extension introduces a seam, and whether a character’s face survives that seam depends on a number most vendors don’t publish at all — how much prior video the model reads before generating the next segment. Likewise, footage upscaled to 4K and footage rendered at a native resolution carry different detail, different artifacts, and in at least one documented case different surface availability.

This reference normalizes both spec numbers into one taxonomy — native single pass, chained extension with a stated context window, and upscale or regeneration — then scores five 2026 video models against it using each vendor’s own wording: Gemini Omni 1.1 Flash, FLUX 3 Video, Seedance 2.5, MiniMax H3, and Veo 3.1.

Read the vendor pages behind the 2026 video models and a pattern emerges: the headline spec numbers are honest, but they are answers to different questions. On the duration side there are two mechanisms — a native single pass, where the model generates the whole clip in one continuous run, and chained extension, where the model repeatedly appends new segments to existing output. On the resolution side there are three — native generation, post-hoc upscaling, and a third path MiniMax introduced that fits neither bucket. Two plus three: five distinct mechanisms compressed into two spec-sheet numbers.

This post is deliberately about one axis: the unit behind the number — what mechanism a duration or resolution claim actually describes. It is a different axis from availability — whether you can call the model at all — which we scored separately in the buildability audit of what actually ships. A model can be fully buildable with a misread spec, or correctly specced and unavailable; the two failure modes are independent, and the vendor status words quoted in the tables below are there to keep them separate.

Start with the biggest number in the set. Google’s launch post for Gemini Omni 1.1 Flash, published August 27, 2026, states that you can extend videos in 10-second increments up to a total cumulative length of 40 seconds. That is chained extension: the 40-second figure is a ceiling across multiple rounds, not the length of any single generation. The same post adds first/last-frame control — you supply the starting and ending frames of a shot and the model generates continuous motion between them — and a 360p draft mode that Google frames as up to 60% faster and at a third of the cost of the standard 720p resolution. Draft-tier economics are their own story; here the relevant fact is that the headline duration is cumulative.

Now the single-pass claims. ByteDance’s own launch post for Seedance 2.5 states the model “can generate high-quality, 30-second audio-video clips in a single pass” — the vendor’s term is “single pass,” and “one-take” is media shorthand, not ByteDance’s verbatim wording. That doubles the 15 seconds ByteDance references for its predecessor, Seedance 2.0. Beyond the single pass, Seedance 2.5 also supports multi-round extension: users can “smoothly append subsequent shots to existing video outputs,” referencing a prior clip by name, up to “several minutes of content from one workflow.”

Black Forest Labs’ FLUX 3 Video, per its GA rollout, generates clips up to 20 seconds long from a single prompt in text-to-video and image-to-video modes — a single-pass generation with native audio in the same pass. Its separate video continuation mode is capped lower: 5 to 15 seconds of new output per step. The headline 20-second figure applies only to the single-pass modes, not uniformly across the product. And MiniMax H3 generates video with native stereo sound up to 15 seconds — with no extension mode stated in the announcement, and no statement of whether that length is one pass.

Lay the four claims side by side and the trap is visible: the largest headline number in the set — 40 seconds — is the only one that is not a single pass. Ranked by continuous generation, Seedance 2.5’s 30 seconds leads, FLUX 3’s 20 follows, and Gemini Omni 1.1 Flash’s per-round unit is 10. Neither ranking is more “true” — they answer different production questions. A continuous take matters for unbroken camera moves; a high extension ceiling matters for total shot length. A spec comparison that mixes the two units is comparing a marathon time to a relay time.

If a duration claim means chained extension, the next question is the one aggregators almost never ask: how much of the existing video does the model read before generating the next segment? That prior-context window is the mechanism variable behind seam quality. A model that conditions on one final frame can match momentary appearance but has no memory of how a character moved, what left the frame, or where the light was two shots ago. A model that reads ten seconds carries motion, identity, and scene state across the cut.

Google publishes its figure, and frames it as the generational change: Gemini Omni 1.1 Flash’s extension step analyzes up to 10 seconds of prior context — in Google’s own words, a leap from previous models that only referenced the final second. BFL publishes one too: FLUX 3’s continuation mode conditions on up to four seconds of existing video — a materially smaller window than its 20-second single-pass headline might suggest. ByteDance, by contrast, describes Seedance 2.5’s appending behavior qualitatively — carrying character, environment, and pacing forward — but states no numeric window at all.

Two of the three current models charted carry a vendor-stated figure; the Seedance row is an honest zero-length bar, because the correct entry is “vendor does not state,” not a guess. The window sizes matter at the seam, a problem we’ve examined in the single-shot-vs-multi-shot coherence problem around the earlier Seedance generation.

The table below isolates the column that existing comparisons skip: the prior-context window per extension step.

Recounting the table: of the four rows, two carry a single-pass headline — ByteDance says “in a single pass” for Seedance 2.5 in so many words, while BFL describes FLUX 3 as producing clips of up to twenty seconds from a single prompt — one is a chained ceiling (Gemini Omni 1.1 Flash), and one states a length without naming a mechanism (MiniMax H3). Exactly two rows carry a vendor-stated prior-context figure — Gemini’s up-to-10 seconds and FLUX 3’s up-to-4 seconds.

Resolution claims split three ways, and vendors are more precise about which path applies than the coverage that quotes them. Google’s Omni 1.1 Flash workflow is draft-then-upscale: the model generates at 360p or 720p, and 1080p or 4K arrive through a separate resolution pass — which is why “Gemini Omni 1.1 Flash generates 4K video” is a misreading, and “offers 4K via upscale” is the accurate sentence. Google’s Veo 3.1 materials use the same verb repeatedly: “Generate videos 1080p and 4K with state-of-the-art upscaling.” BFL is just as explicit for FLUX 3 Video: output “is released at HD, meaning 720p natively, with Full HD available through upscaling.”

The third path deserves the emphasis. Upscalers are separate models bolted after generation; native generation is the diffusion pass itself. MiniMax’s in-context regeneration is neither: the same base model takes its own low-resolution output as context and regenerates it. Whether that produces upscaler-like artifacts or generation-like detail is an empirical question the announcement doesn’t settle — but collapsing it into “upscaling,” as most aggregator writeups do, erases the one mechanistic novelty in the set.

Same discipline as the duration table: the vendor’s verb for the higher-resolution path where one exists, with absences named rather than filled. The Seedance 2.5 row is the instructive one — a major launch whose own post contains no resolution specification at all.

Recounting this table: five rows, of which three carry a vendor-stated native figure (Gemini Omni 1.1 Flash, FLUX 3, MiniMax H3) and two are absences of the first kind — Seedance 2.5 and Veo 3.1, where the pages cited simply do not state a native number. No row claims native 4K. And note the Veo footnote: Google’s January 2026 post restricts the 1080p/4K upscale to Flow, the Gemini API, and Vertex AI — explicitly excluding the Gemini app and YouTube surfaces.

The definitional collapse is usually framed as hype deflation — vendors overstate, careful readers correct downward. The record around these five models shows something more interesting: third-party pages overstate vendors’ own numbers upward, in both of the directions this post tracks.

Case one: MiniMax H3. MiniMax’s own announcement states 2K. Yet multiple secondary pages advertise the model as native 4K at 60FPS. Case two: Veo 3.1. Google’s own posts say “upscaling,” repeatedly and in both the consumer and cloud framings — yet review-style aggregator pages claim the model achieves native 4K rather than upscaling, directly contradicting the vendor. In both cases the primary source is more modest than the coverage, and in both cases the correct citation is the vendor page, not the page with the bigger number.

This is the same failure mode we documented in the availability-language audit — coverage upgrading “preview” to “GA” — now operating on capability numbers. The mechanism is identical: speculation published before a launch outranks the launch in search for a while, and nobody goes back. The practical consequence for anyone budgeting a video pipeline is that resolution claims sourced from anywhere other than the vendor’s page carry a documented upward bias in this category.

The taxonomy compresses into four questions. They take minutes to answer against a vendor page and will usually change which model a production plan actually needs.

Two implications follow from the pattern. First, on interpretation: the vendors are converging on a generate-low, deliver-high architecture — cheap native passes (360p drafts on Omni 1.1 Flash, native 720p on FLUX 3) with resolution added by a separate, separately-controlled step. That split exists because it matches how production actually works: iterate cheap, finish expensive. The spec-sheet confusion is a side effect of an architecture decision that is, on its own terms, sensible. Duration is following the same shape — short native passes, length added by extension — which is why the unpublished prior-context window is becoming the real differentiator while headline seconds converge.

Second, projecting forward: as extension ceilings stretch, expect vendors that publish their context window to make it a marketing number — Google already frames 10-versus-1 seconds as the generational leap — and expect the gap between ceiling and pass to widen faster than single-pass length improves, because chaining is the cheaper axis. When a spec number determines real budget — and per-second pricing differs by mechanism, as we break down in translating these units into cost per finished second — the mechanism check above is the difference between a plan and a surprise. Mapping claims like these to production workflows is exactly the kind of evaluation our AI transformation engagements run before a stack decision, and the same discipline drives how our content engine treats vendor claims in published work: the vendor’s page, the vendor’s verb, the vendor’s number.