DeepSeek: From Controversy Straight to the Stock Exchange
The Chinese AI lab DeepSeek is at the center of attention in July 2026 for several reasons. While the West is thinking about how to defend against it, the company itself is jumping from one milestone to the next. According to Bloomberg and TechCrunch, DeepSeek wants to raise an additional $1.5 billion at a valuation of $71 billion — just one month after closing its very first external funding round of $7 billion.
The planned IPO on China's STAR Market (China's equivalent of Nasdaq) is expected to take place in 2027, with the filing itself expected later this year. Investors include Tencent and the Beijing National AI Industry Investment Fund. DeepSeek's valuation rose from $50 billion to $71 billion in less than six weeks — record pace even by the standards of exuberant Chinese tech companies.
However, parallel to this financial triumph, DeepSeek faces serious accusations from US AI leaders. Both OpenAI and Anthropic have accused DeepSeek (along with Moonshot AI and MiniMax) of so-called "adversarial distillation" — systematically mining knowledge from their models in order to train their own AI systems.
Anthropic specifically documented approximately 24,000 fake accounts and more than 16 million exchanges of specially designed queries aimed at extracting the capabilities of the Claude model. In an open letter to US lawmakers, OpenAI described "new, obfuscated methods" of Chinese data mining. In April 2026, OpenAI, Anthropic, and Google DeepMind agreed to share information about these attacks — prestigious rivals acknowledged that defense against Chinese labs is more important than mutual competition.
For Czech users, it's important to know: DeepSeek is freely available via web interface and API, without geographic restrictions in Czechia. Its models are among the cheapest on the market. However, the question of data security when using Chinese AI services remains open — especially for companies working with sensitive information.
OpenAI Launches GPT-5.6: Three Models for Three Types of Users
On July 9, 2026, OpenAI launched the new GPT-5.6 model family, which comes in three variants with different performance levels and pricing:
- Sol — the flagship model, best performance for coding, science, and cybersecurity. Price: $5 per million input tokens, $30 for output.
- Terra — a balanced model for everyday work deployment. Price: $2.50 / $15 per million tokens.
- Luna — the most economical variant for bulk processing. Price: $1 / $6 per million tokens.
GPT-5.6 Sol will also run on Cerebras infrastructure at speeds of up to 750 tokens per second — significantly faster than previous models. The "Ultra" feature enables coordinating multiple AI agents in parallel, targeting complex enterprise tasks where a single instance isn't enough.
For Czech users: GPT-5.6 is available via ChatGPT and API without geographic restrictions. The ChatGPT interface supports Czech, and the API is of course language-neutral. If you work with AI in a company, Terra offers a good performance-to-price ratio for daily use.
Meta Rewrites Its Own Rules: The First Paid Closed Model
Mark Zuckerberg has long promoted openness as Meta's AI strategy. The Llama model series was released for free and as open-source, giving Meta significant influence in the developer community. But in July 2026, a turning point arrived.
Meta released Muse Spark 1.1 — its very first paid, closed AI model. Price: $1.25 per million input tokens and $4.25 for output. The model is primarily intended as an agentic coder — an AI assistant for automated programming and tool orchestration.
Why is Meta changing course? Analyses suggest that maintaining frontier models as open-source is no longer economically sustainable. AI value is gradually shifting from model weights to the agentic execution layer — meaning how well AI can work autonomously on complex tasks. By closing the model, Meta wants to earn inference fees and fund further research to keep pace with OpenAI and Anthropic.
The Llama series remains open-source, but Muse Spark 1.1 clearly signals that Meta is playing a double game: the community gets cheaper models for free, while enterprise customers pay for top-tier performance.
OpenAI and Disney: A Story with a Bitter Ending
At the end of 2025, Disney and OpenAI entered into a historic partnership — Disney became the first major licensing partner for Sora, OpenAI's generative video platform. Disney pledged a $1 billion investment and access to over 200 characters (Marvel, Pixar, Star Wars). It looked like a breakthrough for AI video content.
However, in March 2026, OpenAI completely shut down Sora — and Disney received a warning just 30 minutes before the public announcement. The reason? OpenAI strategically pivoted toward coding, enterprise AI, and robotics. Sora was generating costs of approximately $15 million per day, while the app's total revenue reached only about $2.14 million from 11.7 million downloads. The math simply didn't work out.
The OpenAI vs. Disney case is instructive: even the most ambitious AI partnerships can be canceled overnight if the model doesn't find economic justification. For Disney, it means the loss of both investment and strategic position; for OpenAI, it was a painful but probably necessary step.
What Does This Mean for the Czech Market?
July 2026 shows the AI industry in full motion: Chinese companies are expanding despite accusations, American giants are rewriting strategies, and partnerships are formed and broken within months. For Czech companies and developers, this brings several practical implications:
- GPT-5.6 is available in Czechia without restrictions and works reliably in Czech — the new pricing for the Terra version is acceptable for enterprise deployment.
- DeepSeek offers extremely low prices, but the question of data security with a Chinese service is legitimate — especially after the published accusations of distillation attacks.
- Meta Muse Spark 1.1 is not yet localized into Czech, but for developers working in English, it's an interesting cheaper alternative.
- The EU AI Act came into force in 2025 and regulations for high-risk systems are still being refined — keep an eye on how individual players are meeting compliance requirements for the European market.
What is adversarial distillation and why do OpenAI and Anthropic consider it a threat?
Adversarial distillation is a technique in which someone systematically sends specially designed queries to a large AI model — not to get answers for themselves, but to train their own model on those answers. The result is that a smaller model "copies" the capabilities of the larger one without having to go through the entire costly training process. OpenAI and Anthropic consider it theft of know-how and intellectual property into which they have invested billions of dollars.
Why did Meta release its first paid model when it has always promoted open-source AI?
Meta found that maintaining top AI models as freely available is no longer economically sustainable. Training frontier models costs hundreds of millions of dollars, and the open-source approach doesn't allow those costs to be covered. Muse Spark 1.1 is therefore aimed at enterprise customers willing to pay for premium performance, while cheaper Llama models remain available to the community for free.
How fast is DeepSeek growing and what does it plan for the future?
DeepSeek is one of the fastest-growing AI companies in the world. In less than six weeks, its valuation rose from $50 billion to $71 billion. The company plans an IPO on China's STAR Market, aiming to file documents in 2026 and debut in 2027. At the same time, it is developing its own AI chips to reduce dependence on Nvidia and mitigate the impact of US export controls.
