Alright, January in the AI world was just crazy. While others were recovering from the holidays, in Silicon Valley, nobody was sleeping. OpenAI launched healthcare for ChatGPT, Anthropic also decided to treat people, and Google rolled out a whole suite of updates. At the same time, everyone started monetizing their products — advertising in ChatGPT at $60 CPM, ChatGPT Go subscription at $8 per month, Google AI Plus for $7.99. It looks like a race for users’ wallets.

The wildest part is the legal drama between Musk and OpenAI. Now he’s demanding $134 billion in wrongful profits, arguing that he invested $38 million in the early stages. Leaked emails reveal that Ilya Sutskever owns roughly $4 billion in the company, according to 2023 estimates. If we extrapolate to the current valuation of (billion, his stake could be more than )billion. The court hearing is scheduled for April.

Now, about the products. Claude Code officially integrated into VS Code and gained access to the terminal — a full development cycle in one window. Plus, a browser extension called Claude in Chrome, which can click, fill out forms, and automate UI tests. Cowork for macOS allows the agent to work directly with files, though it’s a bit scary — the model can delete files if it decides they’re part of the task.

Google released TranslateGemma — open models for translation supporting 55 languages. The small version with 4B parameters works like a 12B base Gemma. They also launched Personal Intelligence — deep integration of Gemini with Gmail and Google Photos, which can find information about you across different services. Sounds powerful, but raises privacy concerns.

Chinese companies aren’t falling behind. Alibaba released Qwen3-Max-Thinking, which benchmarks against GPT-5.2. DeepSeek-OCR 2 reads documents like a human, not like a scanner. Moonshot AI introduced Kimi K2.5 with a swarm mode — up to 100 specialized sub-agents can run in parallel.

On the security front, Nvidia got caught red-handed — employees contacted Anna’s Archive $850 pirate library$100 and downloaded about 500 TB of protected content. Management approved this a week later, citing competitive pressure. Wikimedia Foundation finally started monetizing content — tech giants will pay to use Wikipedia for training models via a commercial API.

Inside Anthropic, they’re tightening the screws. They blocked xAI engineers’ access to Claude Opus 4.5 via Cursor, and banned the use of personal subscriptions in third-party clients like OpenCode. Officially, it’s about UX, but in reality, it’s protecting unit economics.

Interesting cases: Claude helped grow tomatoes via Arduino, controlling lights and watering. JPL NASA used Claude to plan the route for the Perseverance rover — the model charted a path through a rocky field over 400 meters. Axiom Prover solved all 12 problems at the Putnam math competition, generating proofs in Lean language.

Meta bought Manus for (billion. The startup demonstrated real automation use cases and was generating about )million in annual revenue. But the deal could be complicated — the founders started in Beijing, attracting regulatory attention from both sides.

In the startup world, Mirai Murati’s $2 Thinking Machines Lab$100 had a soap opera. CTO Barrett Zoff and his colleague Luke Matz returned to OpenAI after a conflict. Zoff was hiding a relationship with an employee who was later fired, he was demoted himself, then demanded full control over technical decisions. Murati found out about his negotiations with Sam Altman and fired him. The startup lost half of its founders.

Regarding tools: Cursor optimized agent workflows through Dynamic Context Discovery, reducing token consumption by 47%. Amarsia manages infrastructure for AI features without backend coding. Nativeline generates clean Swift code for iOS. There’s even an app for working with one-time codes and two-factor authentication, integrated with agents for secure automation.

Research shows progress in reasoning, scaling computations, and working with memory. DeepSeek stabilized multi-threaded connections via projection onto the manifold of doubly stochastic matrices. Nvidia proposed training during testing — the model updates weights in real-time while reading context. KVzap compresses the KV cache by 3–4 times almost for free.

Overall, January showed that the AI industry isn’t slowing down — it’s accelerating. Monetization, competition, legal dramas, real use cases. Agents are already managing files, coding, planning routes on Mars, and growing plants. The only question is when all this will become the norm, not just news.