The Commercial Race

It is June 2024. OpenAI’s Sora has been a research report for four months, its clips circulating as demos with no sign-up form behind them, and every serious lab in text-to-video has read the same signal: whoever ships a paying product first inherits a market that does not yet have an incumbent. In a single month, Kuaishou, Luma, and Runway each launch one, and by the time Sora itself finally opens to subscribers in December, four other companies already have.

Where We Were

By early 2024, the video generation field had a capability ceiling and a product vacuum sitting side by side. Sora and Spacetime Patches covers what OpenAI actually disclosed in February 2024: a diffusion transformer that tokenizes video into spacetime patches, trains jointly on images and clips of native duration and resolution, and generates up to a minute of coherent footage. What that chapter also establishes is what Sora was not, in February 2024 — a product. OpenAI’s technical report withheld training data, compute budget, and parameter count, and offered no API, no waitlist conversion, and no price. For four months, the most capable video model anyone had publicly described was also the one nobody outside a small group of testers and filmmakers could use.

The best a paying customer could actually buy in early 2024 was still built on the older, bolt-on-U-Net lineage: Stable Video Diffusion, released openly in November 2023, generated 14 or 25 frames from a single conditioning image, a ceiling set by the fixed-resolution, fixed-duration training a convolutional U-Net requires. Runway’s own commercial product, Gen-2, and early entrants like Pika Labs — which had already shipped Pika 1.0 to a waitlist of over half a million users in late November 2023 — sat in the same generation: short, low-resolution clips, useful for storyboarding and social content, nowhere near Sora’s disclosed minute-long, 1080p demos. The gap between what a research report described and what a subscription actually delivered was the opening every other lab with a video model in training was racing to close.

Price only became comparable once every vendor had shipped something to price. PNG ↓

None of the five labs that shipped a product in the ten months after Sora’s announcement had access to OpenAI’s implementation. Each had to arrive at a competitive architecture, and a competitive price, on its own. And none of them, in early 2024, had a public price for video generation quoted in a unit a buyer could put into a spreadsheet — Runway and Pika billed in monthly subscriptions with no disclosed per-second rate, and nobody had yet shipped a metered API for video at all. Price only became something a buyer could compare once the race itself forced it into the open, over the following ten months.

The Breakthrough

The breakthrough this chapter tracks is not a single method — it is a market converging on the same one within months, and then competing on the axis Sora’s report never mentioned: price per second. Google announced Veo at I/O in May 2024, a text-to-video model Google positioned directly against Sora’s disclosed capabilities. Runway followed on June 17 with Gen-3 Alpha, described in its own announcement as “the first of the next generation of foundation models trained by Runway on a new infrastructure built for large-scale multimodal training” — a rebuild, not an incremental update to Gen-2. Kuaishou’s Kling, unveiled June 6 and opened for testing days later, generated up to two minutes of 1080p video at launch, longer than anything else on the market at the time — but aimed first at Kuaishou’s own domestic short-video audience, with an English-language global beta following on July 25, seven weeks after the domestic launch. That gap mattered commercially: for seven weeks, Kling was a capability announcement for most of the world and a shipped product only inside China, the same shape of gap Sora’s report had created at a larger scale. Luma AI’s Dream Machine, released June 12, undercut both on friction: a free public web tool rather than a waitlist, which reportedly reached a million users within four days. Pika Labs, the one 2023-era incumbent in this list, spent 2024 shipping incremental updates to its existing product rather than a Sora-class rebuild, and lost visible ground in the process — a reminder that the race rewarded whoever rebuilt on a new architecture, not whoever had shipped first.

By October, Meta entered from the research side rather than the product side. Movie Gen (Polyak et al., 2024) is a cast of foundation models whose largest video component, a 30-billion-parameter transformer, generates up to 16 seconds of 1080p video at 16 frames per second from a context of up to 73,000 video tokens — and, distinctively for 2024, pairs that video model with synchronized audio generation in the same system. Meta did not ship Movie Gen as a public product in 2024; the audio-generation mechanism behind it, and its eventual path into Meta’s own apps, is covered in Sound and Speech. What matters for the race is the signal Movie Gen sent in October 2024: synchronized audio, not just longer or sharper video, was about to become the next axis of competition.

Sora's report opened the race; every other arrow is a lab answering it on its own schedule, with two separate paths both landing on native audio. PNG ↓

Google shipped that axis to paying customers seven months after Movie Gen’s report. Veo 2, released in December 2024 alongside Sora’s own public launch, focused on motion realism and prompt adherence rather than audio. Veo 3, announced in May 2025, closed the gap Movie Gen had opened: its technical report describes Veo 3 as using latent diffusion in which “the diffusion process is applied jointly to the temporal audio latents, and the spatio-temporal video latents,” producing a single clip with dialogue, sound effects, and ambient audio generated alongside the picture rather than added afterward. Where Movie Gen had demonstrated the capability in a research report, Veo 3 shipped it behind a metered, per-second API price — the same axis every other 2024 entrant had already started competing on, now extended to sound.

The Mathematics

None of the five products this chapter covers publish a new training objective — each is, in its disclosed form, a diffusion transformer over spacetime patch tokens, the architecture and loss Sora and Spacetime Patches already derives in full. What changed in 2024 was that the cost of running that architecture became the thing customers priced, shopped, and compared. A simple accounting identity makes the trade every lab was actually making explicit:

\text{Cost}_{\text{sec}}(m) = \frac{n_{\text{step}} \cdot \mathrm{FLOPs}_{\text{fwd}}(N)}{\pi \cdot s}

$N$ is the number of spacetime patch tokens a clip decomposes into, defined the same way Sora and Spacetime Patches defines it — more resolution, more duration, or coarser compression all push $N$ up. $\mathrm{FLOPs}_{\text{fwd}}(N)$ is the compute one forward pass of the diffusion transformer costs over that many tokens; $n_{\text{step}}$ is how many denoising steps the sampler runs before it hands back a clean clip; $\pi$ is the provider’s realized compute price, in floating-point operations purchasable per dollar, on whatever accelerator generation it runs the model on; and $s$ is the number of seconds of finished video one generation call produces. $\text{Cost}_{\text{sec}}(m)$ , the dollar cost per second of footage for model $m$ , is what a buyer actually compares across vendors — and every lever a lab pulled during 2024 maps onto one term here: fewer sampling steps is step distillation (Making Video Cheap), a larger $\pi$ is a newer accelerator generation (The Co-evolution Thesis), and a smaller $N$ at the same perceived quality is a better compression network.

Symbol	Meaning
$\text{Cost}_{\text{sec}}(m)$	dollar cost per second of finished video, for model $m$
$n_{\text{step}}$	number of denoising steps the sampler runs
$\mathrm{FLOPs}_{\text{fwd}}(N)$	compute cost of one forward pass over $N$ spacetime patch tokens
$N$	number of spacetime patch tokens in the clip, as defined in Sora and Spacetime Patches
$\pi$	the provider’s realized compute price, FLOPs purchasable per dollar
$s$	seconds of finished video produced per generation call

A quoted retail price is not this number — it is this number plus whatever margin, subsidy, or loss-leader strategy a company layers on top during a land-grab phase. A second identity makes that comparison, capability against price, the way a buyer actually should:

\text{Value}(m) = \frac{\mathrm{VBench}(m)}{\text{Cost}_{\text{sec}}(m)}

What It Enabled

By the end of 2024, “can it generate video” had stopped being the question a buyer asked; “at what duration, resolution, and price” had replaced it. The five entrants that shipped in 2024 already looked different on every axis that mattered to a paying customer, not just on sample quality:

Product	Launch	Max duration disclosed	Max resolution disclosed	Distinguishing feature
Runway Gen-3 Alpha	June 17, 2024	Seconds-scale clips, extendable	Up to 1080p	New training infrastructure, tighter motion control
Kling	June 6, 2024	Up to 2 minutes	1080p	Longest disclosed duration at launch
Luma Dream Machine	June 12, 2024	Seconds-scale clips	Up to 1080p	Free public web access, no waitlist
Google Veo (1)	Announced May 2024	Minutes-scale, per Google’s demos	1080p	Positioned directly against Sora’s disclosed capabilities
Sora	Announced Feb 2024, public Dec 2024	Up to one minute	1080p	Ten-month gap between announcement and any paying customer
Meta Movie Gen	October 2024	16 seconds at 16 fps	1080p	Synchronized audio generated with the video, research report only
Google Veo 2	December 2024	Minutes-scale	1080p	Improved motion realism and prompt adherence
Google Veo 3	May 2025	Minutes-scale	1080p	Native, jointly-generated audio at product scale

Price, once labs started disclosing it, turned out to be quoted in units that do not compare cleanly across vendors — a fact more consequential to a buyer than any single benchmark score:

Provider	Pricing unit	Disclosed rate (mid-2026)
Google, Veo 3 via Gemini API	Per second, standard quality with audio	$0.40 per second
Google, Veo 2 via Gemini API	Per second, with audio	$0.35 per second
OpenAI, Sora 2 API	Per second, tiered by resolution	On the order of $0.10 per second at 720p, reported in Sora and Spacetime Patches
Runway	Monthly subscription, credit-metered	$12–$76 per month across three consumer tiers, credits drawn down per generation at a rate the price sheet does not state in seconds
Kling	Monthly subscription, credit-metered	Credits per clip vary by resolution, duration, and whether audio is included; no published per-second rate

Only one of the five vendors above publishes a price a buyer can convert to dollars per second without first estimating how many credits a given clip will consume. That asymmetry, not any single model’s benchmark score, is what a buyer scoping a real integration has to solve for first.

The market that resulted also split by access model, not just by price. Google, and later OpenAI, exposed Veo and Sora through metered developer APIs suited to being embedded inside another company’s product. Runway, Luma, and Pika stayed consumer-subscription products first, with API access arriving later and priced separately from the web app. Kling ran both at once from a single vendor, a consumer web app in China and internationally plus a developer API, priced in the same opaque credit units either way. None of these are architecture differences — every one of these products still runs a diffusion transformer over patch tokens — they are go-to-market differences that determined which kind of buyer each vendor was actually built to serve.

The race did not settle the market, it made the open questions visible, and each one gets its own chapter next. PNG ↓

The race this chapter covers did not settle the market — it opened the next three questions this book takes up in turn. Open-weight labs answered the same architectural convergence with public checkpoints instead of metered APIs (Open Video Models). The audio thread Movie Gen opened and Veo 3 shipped gets its own mechanism-level treatment (Sound and Speech). And the price-per-second pressure this chapter documents as a market outcome becomes, two chapters later, an engineering problem with its own techniques for solving it directly (Making Video Cheap).

Key Papers

Paper	Authors	Year	Link	Why it matters
Movie Gen: A Cast of Media Foundation Models	Polyak et al.	2024	arXiv:2410.13720	Meta’s 30B-parameter video model paired synchronized audio generation with video for the first time in this chapter’s race, as a research report rather than a shipped product
Veo 3 Technical Report	Google DeepMind	2025	Veo 3 Technical Report (PDF)	Shipped Movie Gen’s audio-video convergence as a metered product, applying diffusion jointly to spatio-temporal video latents and temporal audio latents

PM Lens

The number I check first when a vendor pitches a video model is not the benchmark score, it is the unit their price sheet is quoted in. Google’s Gemini API prices Veo 3 at $0.40 per second and Veo 2 at $0.35 per second, standard quality, audio included — I can put that directly into a cost model. Runway and Kling price in monthly credit bundles instead, and neither publishes how many credits a specific resolution-and-duration combination actually consumes, so the “cheaper” plan on the pricing page is not necessarily the cheaper video once I run my actual use case through it. I ask every video-gen vendor for a per-second cost at my target resolution and duration before I ask about quality, because the 2024 race was won on capability demos, and demos do not carry a line-item cost. A roadmap built on a research report’s samples, the way several teams built on Sora’s or Movie Gen’s, is a roadmap with no number in the budget column until someone ships a price.

Search