In 2026 generative media split into two very different markets. In images, OpenAI is back on top: GPT Image 2.5 leads the blind-vote leaderboard, ahead of Google, xAI and Microsoft. In video, OpenAI gave up: Sora was shut down, and Chinese labs such as Alibaba, ByteDance, MiniMax and Kuaishou now hold most of the top places, with Germany's Black Forest Labs as the only European name in the top 10. Kling alone earned more than RMB 850 million in a single quarter. And since 2 August 2026, the EU AI Act requires AI-generated content to be marked. Here is the market in five charts, and what it means for companies in Croatia and the DACH region.
The short answer
Image generation has become a commodity with a premium tier. The best models are separated by a few dozen Elo points, while prices differ by a factor of twenty. Video generation is still a frontier race: quality jumps every few months, clips now run up to 20 seconds with native audio, and one minute of 1080p video costs between about $2 and $34 through the model makers' APIs. The four companies we follow most closely play very different roles: OpenAI competes head-on in images, Anthropic deliberately does not generate images at all, Mistral licenses Black Forest Labs' FLUX models for its assistant, and Manus orchestrates other companies' models inside its agent.
Images: OpenAI is back in front
OpenAI released gpt-image-2 on 21 April 2026 together with ChatGPT Images 2.0, the first image model that "thinks" before it draws: it can reason about layout and search the web before rendering. In September it followed with two GPT Image 2.5 variants, Sunburst and Flare. They now hold the top three places of the text-to-image leaderboard. Google's Nano Banana 2.1, released in October, is fourth, and Microsoft's in-house MAI-Image models have become a serious alternative within a year.
Quality comes at a price. At 1024×1024 in high quality, GPT Image 2 costs about $211 per 1,000 images through the API, roughly $0.21 per image. Google's Nano Banana 2.1, only 13 to 40 Elo points behind, costs $33.60, and Meta's Muse Image $10. For most business use cases, such as product visuals, social posts or presentation graphics, the cheaper models are good enough. The premium tier pays off where text inside the image, exact layouts or brand consistency matter.
Video: Sora is gone, Chinese labs lead
The biggest surprise of the year was OpenAI's exit from video. On 24 March 2026, six months after the much-hyped Sora 2 app launch, OpenAI announced it would shut Sora down. The app closed on 26 April and the API was scheduled to end on 24 September. OpenAI cited compute demand and said the Sora team would focus on world-simulation research for robotics; usage had already been falling since the start of the year. A planned $1 billion Disney investment tied to Sora was dropped at the same time.
The video leaderboard today is led by Alibaba's Wan 3.0 (Elo 1158), ahead of Utopai X, ByteDance's Dreamina Seedance 2.5 and MiniMax H3. Black Forest Labs' FLUX 3 from Freiburg is sixth with 1131, the only European model in the top 10. Google's Gemini Omni Flash 1.1 is eighth. Runway, which led the old leaderboard with Gen-4.5 in December 2025, is no longer in the top 15.
Prices vary even more than for images. One minute of 1080p video costs $4.80 with MiniMax H3, $9.12 with Gemini Omni Flash 1.1, $12 with the top-ranked Wan 3.0, $17.40 with FLUX 3 and $34.12 with Seedance 2.5. Black Forest Labs also offers a draft mode at $0.06 per second, so a 20-second preview costs $1.20 before you render the final version. For a 30-second product video with ten iterations, the model choice decides whether you pay $25 or $170 in compute.
The money: Kling shows that video AI can pay
Most generative-media companies do not publish revenue. Kuaishou, which owns Kling AI and is listed in Hong Kong, does. Kling earned about RMB 150 million in Q1 2025, over RMB 300 million in Q3 2025 and RMB 340 million in Q4 2025, a total of about RMB 1.04 billion (≈ $150 million) for 2025. In 2026 growth accelerated: over RMB 650 million in Q1 (+300% year over year) and over RMB 850 million in Q2 (+200%). Around 70% of early revenue came from consumer subscriptions, with professional creators and advertisers a growing share.
In the West, Black Forest Labs raised $300 million at a $3.25 billion valuation in December 2025, with Salesforce Ventures, a16z, NVIDIA, Canva and Figma among the investors. Its FLUX models power image generation in Adobe, Picsart and Mistral's assistant, and the FLUX 3 backbone is also being tested in a robotics model on Audi production lines. Midjourney, which has never raised outside money, is estimated by third parties at around $500 million in annual revenue, though it does not confirm figures.
Anthropic, OpenAI, Mistral, Manus: four different strategies
| Company | Own image model | Own video model | Strategy in 2026 |
|---|---|---|---|
| OpenAI | Yes: GPT Image 2 (Apr), GPT Image 2.5 (Sep) | No longer: Sora shut down (app Apr, API Sep 2026) | Lead in images inside ChatGPT and the API; video moved to research |
| Anthropic | No | No | No pixel generation. Claude creates charts, diagrams and interactive visuals as code, and Claude Design (Anthropic Labs, 17 Apr 2026) builds prototypes, slides and one-pagers. Image models are connected via MCP or Canva |
| Mistral | No (licensed) | No | Image generation in its assistant runs on Black Forest Labs' FLUX models: a European stack from Paris and Freiburg |
| Manus | No (orchestrates) | No (uses Google Veo 3) | The agent plans scenes and calls third-party models, turning one prompt into a finished video or visual story |
| Yes: Nano Banana 2.1 | Yes: Gemini Omni Flash, Veo | Distribution through Gemini, YouTube and Workspace | |
| Black Forest Labs | Yes: FLUX 3 Image | Yes: FLUX 3 (Jul 2026) | Model supplier to Adobe, Picsart, Mistral and others; open weights for FLUX 3 Dev announced |
For a business user, Anthropic's choice matters. If your company standardises on Claude, which we often recommend for text, analysis and coding, you will need a second tool for photorealistic images or video. In practice that is either an image API connected to Claude via the Model Context Protocol, or a design tool such as Canva or Adobe that already embeds FLUX or another model.
The EU AI Act: labelling is now mandatory
Since 2 August 2026, the transparency duties of Article 50 of the EU AI Act apply:
- Providers of generative AI must mark image, video, audio and text outputs in a machine-readable way so they can be detected as AI-generated. The Commission has not endorsed a single technique; in practice it means metadata (such as C2PA Content Credentials) combined with invisible watermarks.
- Deployers, meaning companies that publish content, must disclose deepfakes (realistic images, audio or video of real people, places or events) and AI-generated text on matters of public interest. Clearly artistic, satirical or fictional works need only a light disclosure.
- The Commission published a voluntary Code of Practice on marking and labelling AI-generated content on 10 June 2026. Signing it is one way for providers to show compliance.
- Fines for breaches can reach €15 million or 3% of worldwide annual turnover.
For most companies in Croatia, Germany and Austria this means: if you use AI images in advertising that show realistic people or situations, label them, keep the provenance metadata, and write the rule into your AI Act compliance and AI usage policy. Copyright is a separate question; see our overview of AI copyright and training-data cases.
What this means for companies in Croatia and the DACH region
- Pick by use case, not by leaderboard rank. For product photos, social media and presentations, a model at $10 to $35 per 1,000 images is enough. Pay for GPT Image 2.5 only where text in the image and exact layouts matter.
- Video is ready for marketing, not yet for everything. Short product clips, social ads and training videos work well and cost a few euros per minute. Longer stories with consistent characters still need human editing.
- Mind data residency and supplier risk. Most top video models come from Chinese labs. For customer data, unreleased products or regulated industries, check where prompts and reference images are processed. FLUX 3 from Germany and Google's EU regions are options; see our guide to sovereign AI and data residency.
- Do not build on a single model. Sora's shutdown showed that even OpenAI products can disappear within six months. Use an abstraction layer or a platform that lets you switch models.
- Label and log from day one. Keep C2PA metadata, mark deepfakes, and document which tool produced which asset. It costs little now and avoids AI Act problems later.
The bottom line
In 2026 AI image generation became a mature, price-competitive market led by OpenAI and Google, while AI video remains a fast-moving race led by Chinese labs, with Black Forest Labs as Europe's contender. OpenAI's Sora exit shows how expensive video still is; Kling's RMB 850 million quarter shows that it can pay. For companies in Croatia and the DACH region, the practical agenda is clear: choose models by use case and cost, keep the option to switch, and label AI-generated content as the EU AI Act now requires.
Frequently asked questions
What is the best AI image generator in 2026?
In blind human preference votes on the Artificial Analysis text-to-image leaderboard (October 2026), OpenAI leads with GPT Image 2.5 Sunburst (Elo 1199), GPT Image 2.5 Flare (1192) and GPT Image 2 (1172), followed by Google's Nano Banana 2.1 (1159), xAI's Grok Imagine Image 2.0 (1155) and Microsoft's MAI-Image-2.6 (1152). Black Forest Labs' FLUX 3 Image and Meta's Muse Image share 1115.
What is the best AI video generator in 2026?
On the Artificial Analysis text-to-video leaderboard (v2.0, October 2026) Alibaba's Wan 3.0 leads with Elo 1158, ahead of Utopai X (1149), ByteDance's Dreamina Seedance 2.5 (1144) and MiniMax H3 (1141). Black Forest Labs' FLUX 3 is sixth (1131) and Google's Gemini Omni Flash 1.1 eighth (1115). Gaps of a few points are within the confidence interval.
Is OpenAI Sora still available?
No. OpenAI announced on 24 March 2026 that it would shut down Sora. The Sora app closed on 26 April 2026 and the API was scheduled to end on 24 September 2026. OpenAI said its Sora team would focus on world-simulation research for robotics. OpenAI still offers image generation through GPT Image 2 and 2.5.
How much does AI image and video generation cost?
At 1024×1024, API prices range from about $10 per 1,000 images (Meta Muse Image) to about $211 (OpenAI GPT Image 2 high quality). For video, one minute of 1080p costs between $2.40 and $34 on the creators' APIs; Wan 3.0, the top-ranked model, costs $12 per minute. Black Forest Labs' FLUX 3 draft mode costs $0.06 per second.
Do AI-generated images have to be labelled in the EU?
Yes. Article 50 of the EU AI Act applies from 2 August 2026. Providers of generative AI must mark images, video, audio and text as AI-generated in a machine-readable way, and companies that publish deepfakes must disclose them. The Commission published a voluntary Code of Practice on marking and labelling on 10 June 2026. Fines can reach €15 million or 3% of global turnover.
Want to use AI images and video safely in your marketing?
We help companies in Croatia and the DACH region choose the right image and video models, connect them to Claude and existing tools, and set up labelling and usage rules that meet the EU AI Act. As a Claude Certified Architect based in Zagreb, we start with your use cases, not the hype.
Talk to an AI consultant