AI video in 2026 after Sora: Gemini Omni, Kling 4.0, Seedance and Firefly

Picture a small campaign team. By morning, it has two product photos and a rough script; by afternoon, it needs a finished clip in several formats. The first draft arrives quickly. The harder work is keeping the same character across scenes, changing one faulty shot without rebuilding everything, and showing who approved the final export.

The market question has changed too. Sora is no longer available, but AI video has not stood still. New models from Google, Kuaishou and ByteDance compete on longer scenes and stronger reference controls. Meanwhile, Runway and Adobe Firefly are bringing multiple models together with editing and production steps. For a team, the best-looking demo is only one part of the decision. The full path to an approved file matters more.

Sora is gone; development has continued

As of November 2, 2026, both shutdowns have already happened: OpenAI ended the Sora web and app experiences on April 26, 2026, and the API on September 24, 2026. Sora is no longer an active service to plan new work around. Those dates do not explain why OpenAI ended the product or prove that another model is better. OpenAI’s Help Center confirms both dates.

The practical lesson from the Sora era remains useful: an impressive demo is not a dependable production process. A team still needs to track image and audio sources, version files clearly, check rights, and assign someone to approve the final export.

Choose the job first, then the model

Some products are individual models that generate clips from text, images or reference footage. Others are workspaces where a team can switch models, make edits, keep versions together and prepare a final cut. These are often grouped under “AI video tools,” but they solve different problems.

Google: Gemini Omni 1.1 and Flow add control over transitions

Google introduced Gemini Omni 1.1 Flash on August 27, 2026. Its API can extend scenes in 10-second increments to a cumulative total of 40 seconds. It also supports setting the first and last frame; 360p drafts are designed for quicker iteration. For Flow, Google highlights first/last-frame controls and exports at 1080p and 4K. The Gemini app, Flow and the API do not necessarily offer the same controls or terms. Google describes the API features and the Flow update.

That is useful when a scene needs extending or when the transition between two specified images matters. But a short website illustration may only need a quick draft; 4K will not help if the movement, subject or message is still wrong.

Kling AI 4.0: longer shots and reference control

Kling 4.0 launched on October 15, 2026; Kuaishou confirmed the launch in its announcement. The company says the model can generate up to 30 seconds in one pass and use up to 15 reference assets and ten keyframes. Kuaishou emphasizes dynamic movement, complex camera work and longer storytelling. These are vendor claims, not an independent ranking. Kuaishou’s launch announcement describes the announced capabilities.

The practical test is whether references stay consistent and whether one shot can be changed without rebuilding the rest of the scene. Adobe’s current partner-model list also includes Kling 4.0 Omni for image-to-video and video-to-video workflows in Firefly. The variant available depends on the product and region.

Seedance 2.5: more references and targeted edits

ByteDance introduced Seedance 2.5 on July 31, 2026. The company lists clips up to 30 seconds, up to 30 image references, ten video references and ten audio references, plus timestamp-based edits. That points to more control over longer scenes, but it does not replace testing the model with your own style and source material. ByteDance Seed’s announcement explains the model’s features and limits.

Access depends on the workspace: ByteDance named Jimeng AI and Doubao Pro at launch, while Runway also documents Seedance 2.5 in its MCP offering. A model name alone does not tell you its availability, plan, data path or terms of use.

What the first tests show

The first reviews do not pick one model that wins at everything. They do show what can go right, and where video tools still make surprising mistakes.

• Gemini Omni 1.1 Flash: In a test published by Fulgor on October 9, the model spoke a Polish sentence very convincingly, scoring 9/10. Its coffee and tram clips had visible flaws. In the tram scene, a tripod even appeared on the tracks: the model had turned a camera instruction into an object in the picture. Fulgor test.
• Seedance 2.5: Twenty-seven people from creative and technical fields tried it for a week. Compared with the previous version, they found that camera movement and action scenes looked more convincing. But a character can change appearance in a longer clip, and the model may miss detailed instructions. Many retries also mean more time and cost. The 36Kr test, published August 12.
• Kling 4.0: Fuser gave versions 4.0 and 3.0 the same four scenes to create, with just one try for each. In three cases, the newer version followed the action better, including how to row a boat and how to perform a judo throw. The flower opening looked almost the same in both versions. It is a useful first check, but four scenes are not enough to name an overall winner. Fuser comparison from October 23.

For now, no model does everything best. For your own video, check three simple things: does the speech sound right, does the character stay the same, and does the action match what you asked for?

Runway and Firefly are becoming production workspaces

Runway remains a model provider with Gen-4.5, while adding an environment for agents, editing and partner models. On September 29, 2026, Runway announced Runway Creative in the OpenAI Marketplace for eligible enterprise customers. Runway says the workspace brings together Gen-4.5, Seedance 2.5, ChatGPT Image 2.5 and ElevenLabs V4, among other models; an agent can help plan, generate and assemble work. This is a route for eligible enterprise customers, not a claim that every Runway user gets the same model list. Runway’s Marketplace announcement and product changelog show how quickly integrations can change.

Adobe Firefly follows a similar platform approach. Adobe’s partner-model list, updated on October 23, 2026, includes options such as Kling 4.0 Omni, Kling 3.0, Seedance 2.0, Veo 3.1, Runway Gen-4.5 and Gemini Omni Flash in Firefly Boards, depending on the feature. Adobe identifies Firefly AI Assistant and its timeline as beta features; partner-model availability can vary by region. Adobe’s model list, AI Assistant FAQ and timeline guide describe the current state.

The data route matters too. Adobe says prompts and reference files are sent to the model providers for selected Chinese-developed partner models. If a team uploads internal material, client data or a product that has not launched, it should check which model processes the request and which terms apply.

How a team can compare the options

A quick, fair comparison uses the same task and inputs:

1. Continuity: Can the same character or product remain recognizable across two shots?
2. Correction: Can one detail be changed without regenerating the whole scene?
3. Transition: Can the team control the beginning and end of a movement?
4. Effort: How many attempts does it take to get a usable clip, including editing, export and approval?
5. Operations: What credits, generation times, export formats, country restrictions and data routes apply to this account?

For every attempt, record the model and version, prompts, references, date, edits and approval status. Later, the team should still be able to explain which source material went into the final clip and recover an earlier version.

## The decision depends on the whole workflow

• Gemini Omni and Flow are worth considering when reference images, scene extension and first/last-frame controls fit the team’s Google workflow.
• Kling 4.0 belongs in a comparison when longer shots, keyframes and multiple references matter.
• Seedance 2.5 is worth testing when the brief relies on many image, video or audio references and timestamped edits.
• Runway may fit when generation, editing, an agent and multiple models need to live in one workspace.
• Adobe Firefly is relevant when partner models and editing need to sit within a Creative Cloud workflow, after checking beta status, account and regional availability, and data terms.

Sora has left the market. There is no single replacement: model makers keep advancing their systems, while platforms put several of them into production workflows. Compare more than the best opening frame. Compare the clip your team can revise, explain and approve.

Sources

1. OpenAI: What to know about the Sora discontinuation
2. Google: Build with Gemini Omni 1.1 Flash
3. Google: New creative controls in Flow
4. Kuaishou: Kling 4.0 launch
5. ByteDance Seed: Introducing Seedance 2.5
6. Runway: Runway Creative in the OpenAI Marketplace
7. Runway: Product updates and changelog
8. Adobe: Partner models in Adobe products
9. Adobe: Firefly AI Assistant FAQ
10. Adobe: Generate video with Firefly models and add it to the timeline
11. Fulgor: Gemini Omni Flash hands-on tests
12. 36Kr AI Evaluation: Seedance 2.5 seven-dimensional hands-on test
13. Fuser: Kling 4.0 vs. 3.0 on identical inputs