Table of Contents
Every week in the AI world now is worth more than an entire year was five years ago. March 2026 is no exception.
This roundup of artificial intelligence news covers everything important: new models from OpenAI and Anthropic, breakthrough developer tools, unexpected alliances, and a couple of high-profile failures. Read on so you don't miss anything essential.
New language models
GPT-5.4 mini and nano: OpenAI bets on subagents
OpenAI released two compact models — GPT-5.4 mini and GPT-5.4 nano. These are not just cheaper versions of the flagship model: the architecture is optimized for agent pipelines. The idea is simple and elegant — the large model plans and decomposes the task, while the smaller ones execute individual steps in parallel.
At the same time, OpenAI added subagent support in Codex. This is a significant shift: previously Codex worked as a monolithic agent; now tasks can be distributed among specialized executors. The approach is similar to microservice architecture, only for AI tasks.
What this means in practice: the cost per call will drop, speed will increase, and complex multi-step tasks will become economically viable.
Claude: 1 million tokens — now for everyone
Anthropic made the context window of 1 million tokens generally available (GA) at standard prices, with no extra multipliers. Previously, extended context was either in beta or required a separate plan.
1M tokens is roughly 750,000 words, or the entire novel "War and Peace" four times over. Practical uses: analyzing a large codebase in full, working with lengthy legal documents, deep data analysis without chunking.
Mistral Small 4 and the Forge platform
Mistral released Mistral Small 4 — another update to its compact model with improved reasoning. At the same time, the company announced Forge — a platform for training custom models on enterprise data: pre-training, post-training, and RLHF are supported.
Among the first partners are ASML and the European Space Agency. These are serious clients with real requirements for quality and data security.
Chinese models are keeping pace
The Chinese AI market continues to impress:
- MiniMax M2.7 — described as a "self-evolving" model, reached SOTA on the SWE-Pro benchmark
- Xiaomi MiMo-V2-Pro — a model with 1 trillion parameters, aimed at mobile devices and edge inference
Developer tools
Cursor Composer 2: its own model turned out to be a Kimi K2.5 fine-tune
Cursor released Composer 2 — a new version of its AI coding agent with its own model. Claimed price: 86% cheaper than the previous version. The community quickly discovered that under the hood there is a fine-tune of Kimi K2.5 from MiniMax. Cursor did not dispute it.
An interesting precedent: a major Western product openly uses a Chinese model as its foundation. This is either an acknowledgment of the quality of Chinese development, or pragmatic cost saving — most likely both.
OpenAI acquires Astral
Astral — the team behind uv (the fastest Python package manager), ruff (linter), and ty (type checker). OpenAI is acquiring the company, and the entire team is moving to Codex.
This is a strategic move: OpenAI wants to own the Python development infrastructure, not just assist with it. For the community, this is a warning sign — open-source tools are moving under corporate control.
Stripe Machine Payments Protocol (MPP)
Stripe launched the open MPP protocol for machine payments — when an AI agent initiates and completes a transaction on its own, without human involvement. Visa has already joined.
This is an infrastructure shift: agents gain the ability not just to recommend, but to act — buy, pay, book.
1Password Unified Access
1Password launched a solution for secure access by AI agents to enterprise credentials. The agent receives credentials only for a specific task, within strictly defined permissions.
Video and multimodality
Runway: real-time HD video generation
Runway demonstrated real-time HD video generation on Nvidia's latest Vera Rubin hardware. Latency is instantaneous, and the quality is full HD.
Real time changes the possible use cases: interactive video worlds, live visual effects, personalized video content on demand.
Midjourney v8: 5x faster
Midjourney has started closed testing of version 8. Key improvements:
- Generation speed increased by 5 times
- Significantly improved text rendering in images
ElevenLabs: AI music marketplace
ElevenLabs launched an AI-generated music marketplace as part of the ElevenCreative platform. Creators can publish and monetize tracks created with AI.
xAI Text-to-Speech with emotion control
xAI launched a TTS API based on Grok with emotion and intonation control directly through text instructions.
Corporate news
OpenAI in “red code” mode
OpenAI is combining ChatGPT, Codex, and the Atlas browser into a single desktop superapp. At the same time, the company is winding down side projects and focusing on coding and enterprise users — amid preparations for an IPO.
Google Stitch: Figma shares fell 12%
Google significantly upgraded the tool Stitch — an AI-native design canvas with an agent, voice control, and instant prototypes. The tool is free. Figma shares fell 12% on the day of the announcement.
Bezos raises $100 billion for AI-powered manufacturing automation
Jeff Bezos is raising $100 billion for a fund that will buy up manufacturing companies (chipmaking, defense, aerospace) and implement AI automation.
Meta shuts down Horizon Worlds
After ~$80 billion in losses Meta is officially shutting down Horizon Worlds. The VR version will be turned off on June 15, 2026. The metaverse era is officially over.
Security and research
NVIDIA NemoClaw: zero-permissions by default
At the GTC 2026 conference, Jensen Huang introduced NemoClaw — an open-source security framework for AI agents. Principles: zero-permissions by default, isolated sandboxes, and auditing of all agent actions.
Also announced was Nemotron Coalition — an alliance of eight AI labs to jointly create an open-source frontier model.
DarkSword: vulnerability affects 25% of all iPhones
Google, iVerify, and Lookout discovered DarkSword — a zero-click exploit for iOS 18. The attack occurs through infected websites without any user interaction. Around 25% of all iPhones worldwide are at risk.
⚠️ The exploit code is already publicly available. Update iOS immediately.
Anthropic study: 81,000 interviews about fears of AI
Anthropic surveyed 81,000 people from 159 countries about their attitudes toward AI. The main practical fear is not job loss (22%), but unreliability and hallucinations (27%). People are not afraid that AI will replace them, but that AI will give wrong answers in important situations.
Frequently asked questions
How is GPT-5.4 mini different from regular GPT-5.4?
GPT-5.4 mini is a specialized compact version optimized for use in agent pipelines. It is faster and cheaper to invoke, but is designed to handle specific narrow tasks under the control of a more powerful planning model. It is not a replacement for the full version for complex autonomous tasks.
What does 1 million context tokens mean in practice?
One million tokens is roughly 750,000 words. In practical terms: an entire codebase of a medium-sized project, a full year of business correspondence, a large legal document, or a medical history — all of this can be processed in a single request without splitting it into parts. Anthropic's Claude was the first to make such context publicly available at standard prices.
Should you be worried about the DarkSword vulnerability?
Yes. DarkSword is a zero-click exploit, meaning that simply visiting a malicious website is enough for infection. Updating to the latest version of iOS is the only reliable protection. The exploit code is already publicly available, which sharply increases the risk of mass attacks.
What does the shutdown of Horizon Worlds mean for the VR market?
It is a symbolic end to the “metaverse” era as a consumer VR concept. VR/AR technologies are not going anywhere — they are used in education, medicine, and design. But the idea that people would spend their social lives in Meta's virtual worlds has suffered a complete defeat.
Conclusions
This week's artificial intelligence news paints a clear picture: the industry is moving from experiments to infrastructure.
Key trends:
- Agentic systems are becoming the norm — OpenAI, Anthropic, and everyone else are building an ecosystem for multi-agent pipelines
- Security is moving to the forefront — NemoClaw, 1Password Unified Access, MPP with rights control
- Speed and cost are falling faster than expected — Cursor Composer 2, GPT mini/nano, Runway realtime
- Trust matters more than power — Anthropic's study says this plainly
Follow updates on AI Dawn — we publish only what truly matters.
Sources: OpenAI Blog, Anthropic Blog, Mistral AI, Runway, CNBC, Wired