Leading Video Models: Sora, Veo, Runway, Kling, and Pika

"In 2023 video AI was a curiosity. In 2024 it became a craft. In 2025 it shipped Hollywood. In 2026 we are arguing about union contracts."

Pixel, Sora-Watching AI Agent

20.7.1 The 2026 Frontier: Sora 2, Veo 3, Runway Gen-4

Three models hold the consumer-grade frontier in mid-2026: OpenAI's Sora 2, Google's Veo 3, and Runway Gen-4. They differ less in absolute quality (all three produce broadcast-acceptable output) and more in their strengths.

Sora 2 (OpenAI, October 2025) shipped with native audio generation, character consistency across multi-shot sequences (the "cameos" feature lets a user define a character and reuse it across many shots), and 60-second-plus clips at 1080p. The audio is the most striking new capability: Sora 2 generates synchronized sound effects, ambient audio, and intelligible dialogue alongside the video in a single forward pass, eliminating the need for a separate Foley step. The character-consistency feature works by training the model to attend to a special "subject" token derived from a few reference frames; subsequent generations reuse the same token.

Veo 3 (Google DeepMind, April 2025) leads on visual fidelity and physical-world realism. The model is rumored to have been trained with extensive simulation-grounded data plus reinforcement signals from a perception model, giving it stronger handling of fluid dynamics, cloth, hair, and crowd behavior. Veo 3 also has the best handling of complex camera moves (a topic Section 33.3 covers in depth). It does not produce audio in the same forward pass; Google ships it alongside a separate Lyria-2-based music generator that the user composes in post.

Runway Gen-4 (June 2025) targets the professional filmmaker rather than the consumer prompt-user. Its interface treats video generation as part of a non-linear editor: you upload reference images, character sheets, and storyboard frames, and the model fills in shots that match the reference style and characters. Runway's strength is the integration layer, not the raw model quality (Sora 2 and Veo 3 produce somewhat better individual shots in head-to-head blind tests); for a working production team, the difference between "best model in a notebook" and "best workflow in a film tool" matters more than the difference between FVD scores.

20.7.2 The Second Tier: Kling, Pika, Luma Dream Machine

A second tier of models occupies the $5-15 per minute price band and produces 80-90% of the quality of the frontier at a tenth of the cost. Kling 2.0 (Kuaishou, March 2025) is the dominant model in China; it produces excellent character motion and was the first commercial system to ship physically grounded long-form video. Pika 2.0 (Pika Labs, late 2024) offers a strong consumer creator product with a friendly UI and good handling of stylized animation (anime, cartoons, motion-graphics styles). Luma Dream Machine (Luma AI, mid-2024) introduced the camera-control workflow that Veo 3 later improved on; Luma's product is now more of a specialist tool for camera-driven shots than a general-purpose video generator.

These second-tier models are often the right choice for production teams that need predictable throughput, stable pricing, and acceptable quality for B-roll, social-media content, or stylized animation. The first-tier (Sora 2, Veo 3, Gen-4) is the right choice when shot quality is the hero of the project and the budget can support $50-200 per minute of finished content.

20.7.3 Capability Matrix

20.7.4 Evaluation: FVD, VBench, and Human Preference

Video generation lacks a universally accepted single metric, in the way that image generation has FID and CLIP score. The two automatic metrics most often cited are FVD (Fréchet Video Distance, adapted from FID by computing feature statistics on Inception-3D features over short clips) and VBench (Huang et al., 2024), a battery of 16 quality dimensions including subject consistency, motion smoothness, dynamic range, scene coherence, and aesthetic quality. Both correlate with human preference but neither is a perfect proxy: a model that scores well on FVD can produce shots that humans clearly prefer worse than a model that scores poorly.

In practice, frontier-comparison work in 2025-2026 has converged on arena-style human preference evaluations: paired blind comparisons where humans pick the better of two videos given a shared prompt, aggregated across thousands of prompts. The Chatbot Arena equivalent for video is the Hailuo Arena (Minimax), the LMSys Video Arena, and Artificial Analysis's video leaderboard. These rankings shift month-by-month as new models ship, and the standings as of mid-2026 are: Sora 2, Veo 3, Kling 2.0, Runway Gen-4, Pika 2.0 in roughly that order on overall arena score, with sub-rankings that vary substantially by category (motion-heavy, character-driven, abstract).

# Generate a video with Runway Gen-4 via the official Python SDK
# pip install runwayml
import os
import time
from runwayml import RunwayML

client = RunwayML(api_key=os.environ["RUNWAYML_API_KEY"])

# Gen-4 supports image-to-video, text-to-video, and reference-driven generation.
# Here we use a reference image as the first frame plus a motion prompt.
task = client.image_to_video.create(
    model="gen4_turbo",
    prompt_image="https://example.com/reference_frame.jpg",
    prompt_text=(
        "the character turns their head slowly to look at the camera, "
        "warm golden-hour lighting, shallow depth of field, cinematic"
    ),
    duration=10,                  # seconds; Gen-4 caps at 30
    ratio="16:9",
    seed=42,
)
print(f"Task {task.id} queued")

# Poll until completion; Gen-4 turbo typically returns in 2-4 minutes for 10s.
while True:
    status = client.tasks.retrieve(task.id)
    print(f"  state={status.status} progress={status.progress}")
    if status.status in ("SUCCEEDED", "FAILED"):
        break
    time.sleep(15)

if status.status == "SUCCEEDED":
    print(f"Output URL: {status.output[0]}")

20.7.5 Open Alternatives: CogVideoX, HunyuanVideo, Wan 2.1, Open-Sora

The open-weight video ecosystem in 2026 has four serious players. CogVideoX-5B remains the easiest model to fine-tune; the community has produced hundreds of LoRAs and full fine-tunes on Hugging Face. HunyuanVideo-13B (Tencent, December 2024) is the highest-quality open weight, competitive with Kling 2.0 on standard benchmarks; it requires a multi-GPU setup (4x A100 minimum for inference). Wan 2.1 (Alibaba, March 2025) introduced an explicit causal VAE that supports streaming generation and was the first open model to ship credible long-form clips (60-second cap). Open-Sora 2.0 (PKU-YuanGroup, late 2025) is the largest open Sora-style implementation, with 13B parameters and an open dataset of around 80k hours.

These four serve different community roles. CogVideoX is the teaching model and the fine-tune base. HunyuanVideo is the quality reference if you have the hardware. Wan 2.1 is the streaming-and-long-form reference. Open-Sora is the recipe-transparent reference. None matches the closed frontier on out-of-the-box quality, but all four are good enough for product use cases where licensing flexibility, on-premise deployment, or domain-specific fine-tuning matters more than the absolute quality ceiling.

20.7.6 Production Cost, Latency, and Throughput

Video generation is the most expensive multimodal modality in production. A 10-second 1080p Sora 2 generation consumes roughly 100-200 GPU-seconds on an H100; pricing it at $0.30-$1.00 per second of output reflects the hardware cost plus margin. For comparison, a single image from Stable Diffusion 3 is about 0.5 GPU-seconds, and a 30-second piece of audio from MusicGen is about 5 GPU-seconds; video is 50-100x more expensive per second of output.

Throughput is similarly constrained. A frontier model like Sora 2 takes 1-5 minutes of wall-clock time per generation; a typical production team running daily content cannot replace humans with synchronous calls and has to design async batch workflows around the latency. The product implication is that video AI is currently a "send your request, get an email when it is ready" UX rather than the "type a prompt and see it in 3 seconds" UX of image generation. This will change as inference efficiency improves, but the 50-100x compute gap relative to images is structural and will persist.

20.7.7 The Arms Race and Its Limits

The pace of capability improvement in video AI shows no sign of slowing, but several structural limits will shape the next 18 months. Training compute is the most visible: Sora 2's reported training cost is in the high tens of millions of dollars, and a hypothetical Sora 3 at 10x scale would require a billion-dollar training run that only Google, OpenAI, and Microsoft can finance. Training data is the second: high-quality video with clean captions is scarce, and the synthetic-data trick that worked for image and text generation (use a strong model to label data for the next-generation model) is harder for video because video labels are richer (camera path, character identity, motion description) and synthetic labeling errors compound.

The third limit is product. Even the best 2026 model produces shots that need human selection and editing; the prompt-to-finished-video pipeline that some 2024 commentators predicted has not materialized. The 2026 production workflow is human creative direction plus AI shot generation plus human editing, and the editing-and-curation step has not been automated away. The arms race in capability has not yet translated into elimination of the human role; it has translated into different humans doing different work at different scales.