OpenAI, Google, and Tesla set the week's agenda: GPT-5.4 mini, AI Studio, and Terafab

The week in AI was revealing: OpenAI is making reasoning cheaper, Google is shortening the path from idea to ready service, and Tesla wants to break free from dependence on external chip factories. Against this backdrop, the market is simultaneously accelerating and growing more nervous: agents are breaking out of sandboxes, and security issues are already catching up to the hype.

Models Are Getting Cheaper Faster

The main news of the week is OpenAI's release of GPT-5.4 mini and nano as compact versions of the flagship model. According to the company, mini has made noticeable gains on engineering tasks, works faster than the previous GPT-5 mini, and is now available not only through the API but also inside ChatGPT. For free and Go users, it's available in Thinking mode, while for paid customers mini became a backup engine for the heavier GPT-5.4 Thinking. Nano is positioned as an option for routing, classification, and other scenarios where price and minimal latency matter.

In parallel, the code market continues to separate itself from "universal" LLMs. Cursor showed Composer 2 — its own model for agentic programming and large codebases. The company claims that on its internal benchmark it holds its own against top competitors while costing less. Separately, Mistral Small 4 came out — an open MoE model with 119 billion parameters featuring multimodality and long context. Formally this is a strong release for the open-source scene, but by benchmarks the novelty looks less confident than would be hoped against more compact competitors.

AI Goes Into Product

Google significantly raised the stakes in developer tools. AI Studio is now presented not as a demo interface but as a full-fledged browser environment where the Antigravity agent assembles an entire application from a text description: frontend, backend, database, authentication, and external API integration. An important nuance is that the work doesn't stop when the browser tab closes: the agent can continue assembly and configuration without constant user presence. If Google is to be believed, the company has already built hundreds of thousands of applications this way.

• Antigravity assembles full-stack applications on Next.js, React, or Angular.
• Stitch designs the design system, screens, and immediately provides code.
• Runway Characters creates voice video avatars from a single image.
• Midjourney V8 Alpha accelerates generation and provides native 2K mode.

The second line of this same story is tools that eliminate manual work between idea and prototype. Google updated Stitch: the service accepts text, screenshots, and voice, builds clickable screens, and can export the result to Figma. The market reacted instantly — Figma stock dropped 8.8% after the announcement. Simultaneously, Runway launched Characters — real-time avatars for support and onboarding, and Midjourney opened V8 Alpha. But while Runway looks like a step toward productive application of generative video, Midjourney so far looks more like an early release under competitive pressure: faster, but not necessarily higher quality.

Hardware and Control

The most ambitious announcement came from Tesla. The company plans to build Terafab in Texas — a chip manufacturing facility costing $20–25 billion with a claim to 2-nm process technology, proprietary packaging, and scale up to one million wafer starts per month by 2030. The logic is clear: demand from FSD, Optimus, and Dojo is growing, and dependence on TSMC and Samsung is becoming a strategic risk. Tesla claims that AI5 will be the facility's first product, and the company immediately compares it to Nvidia solutions in terms of efficiency and price.

"We'll start small, make early mistakes, then scale."

But along with infrastructure optimism, anxiety around AI agents is also growing. In China, OpenClaw became a mass phenomenon, but regulators call its configuration too weak from a cybersecurity perspective: the agent needs broad system access, and malicious instructions can be hidden directly in web pages. As a result, bans have appeared for government agencies and state banks, though local authorities at the same time offer grants to developers of applications based on it. Another symptom of the market is the Professional AI Bully vacancy from Memvid at $100 per hour: the startup pays people for toxic dialogues with bots to break their memory and find failures before production.

What This Means

The AI market in one week showed three directions simultaneously: models are getting cheaper and going mainstream, product assembly tools are moving into browser-based agentic environments, and the fight for computing power and security is becoming as important as generation quality. For users this is acceleration, for business — a window of opportunity, and for developers and regulators — a new layer of responsibility.