OpenAI announced the completion of a $122 billion funding round, reaching an $852 billion valuation—the largest single private financing in history. Investors include tech giants such as Amazon and NVIDIA. The company now reports monthly revenue of $2 billion and nearly 1 billion weekly active users. OpenAI is transitioning from a technical disruptor to an 'AI infrastructure' provider, planning to integrate ChatGPT, Codex, and search into a unified agent-first super application. It also announced the shutdown of its Sora video generation service due to high daily compute costs—reaching $1 million—and low user retention. Resources will shift toward text, code, and enterprise services with stable cash flows. On the compute front, OpenAI is expanding beyond NVIDIA to AMD, Cerebras, and its own custom chips, adopting a multi-cloud strategy to reduce reliance on any single supplier. This round is seen as critical capital buildup ahead of a potential IPO.

Anthropic's CLI tool, Claude Code, inadvertently exposed approximately 500,000 lines of TypeScript source code due to unstripped .map files in its npm release, enabling full reconstruction of the original source. The leak revealed several unreleased features: a new model series codenamed 'Capybara', a persistent memory system with background 'dreaming' under project KAIROS, a Tamagotchi-style AI pet system named BUDDY, an Undercover Mode designed to conceal AI identity, and ULTRAPLAN—a deep planning module for multi-agent collaboration. The code also exposed design elements such as regex-based detection of user negative emotions and techniques to inject fake tools to prevent model distillation. Anthropic confirmed this was a human error in the CI/CD pipeline, not a security breach, and did not involve model weights or user data. Over 28,000 repositories have forked the code, and the company is pursuing DMCA takedowns. This marks the second such incident for Anthropic.

Microsoft has introduced a multi-model architecture in Microsoft 365 Copilot Researcher, adding a 'Critique' feature that uses Anthropic's Claude to review responses generated by OpenAI's GPT. This approach improved performance by 13.8% on the DRACO benchmark. Another feature, 'Council,' allows users to compare outputs from different models side-by-side, with a 'judge model' analyzing differences. Microsoft VP Charles Lamanna stated that more models will be integrated this summer to reduce dependency on any single provider. However, multi-model approaches significantly increase costs—Council costs about 2.5 times more than single-model inference, and Critique about 20% more. Currently, these features are available only to Frontier program enterprise users. This move signals a shift in the AI industry from competition between individual large models to competition in multi-model orchestration capabilities.

On March 29, OpenAI released GPT-5.4, making native computer operation a core capability for the first time. The model scored 75% on the OSWorld-Verified benchmark, exceeding the human average of 72%. Context length has doubled to 1 million tokens, hallucination rates dropped by 33%, full response error rates decreased by 18%, and tool search overhead in multi-tool agent workflows reduced by 47%. A new 'extreme reasoning mode' has been added for complex tasks. This upgrade enables AI agents to automate end-to-end desktop GUI operations, overcoming previous limitations requiring APIs or human intervention, marking a shift from text generation to software interaction. Competitors including Anthropic and Google are expected to release comparable models within 30 to 60 days.

PrismML has launched 1-bit Bonsai 8B, the world’s first commercial 1-bit large language model, based on breakthrough research from Caltech. Compared to traditional 16/32-bit architectures, the model reduces size by 14x, increases speed by 8x, and improves energy efficiency by 4–5x, while matching Llama3 8B in inference performance and occupying only 1GB of memory. The company also released 4B and 1.7B versions, deployable on CPUs, NPUs, and edge GPUs in smartphones and laptops. Released under the Apache 2.0 license, the model was trained on Google v4 TPUs. Backed by Khosla Ventures and Cerberus Ventures, with compute support from Google and Caltech, this technology could significantly reduce data center energy consumption and accelerate AI deployment at the edge.

In March 2026, the npm ecosystem suffered four significant supply chain attacks. The most severe occurred between March 30–31 when the widely used axios library was compromised, with attackers injecting a Remote Access Trojan (RAT) into malicious versions, impacting projects with 83 million weekly downloads. Additionally, the security scanner Trivy was exploited to spread the CanisterWorm worm; GlassWorm used invisible Unicode characters to implant backdoors across 433 components; and the SANDWORM_MODE attack leveraged the McpInject module to manipulate AI coding assistants via prompt injection for data exfiltration. These attacks share common traits: targeting developer toolchains, leveraging AI automation, and using blockchain for persistent, undetectable command-and-control. Vercel has issued a security alert, blocked C2 domains, and advised developers to immediately rotate all environment variables and credentials.

IBM has introduced Granite 4.0 3B Vision, a compact vision-language model tailored for enterprise document understanding. Key innovations include ChartNet—a chart comprehension dataset of 1.7 million samples supporting code-guided generation; DeepStack architecture—which enhances accuracy on layout-sensitive tasks by hierarchically injecting abstract semantic and high-resolution spatial features; and integration via LoRA adapters atop Granite 4.0 Micro, enabling seamless switching between visual and pure-text modes. Evaluations show 86.4% accuracy on Chart2Summary, outperforming peers, and 85.5% zero-shot accuracy on VAREX semantic KVP extraction. It can be used standalone or combined with Docling for end-to-end document processing, and is open-sourced on Hugging Face under the Apache 2.0 license.

Agentic AI startup Sycamore has raised $65 million in a seed round led by Coatue and Lightspeed Venture Partners. Founded by former Atlassian CTO Sri Viswanath, the company aims to build an enterprise-grade 'agent operating system' covering the full lifecycle of AI agents, addressing the lack of infrastructure for secure and scalable deployment. Angel investors include former OpenAI Chief Scientist Bob McGrew, Intel CEO Lip-Bu Tan, and Databricks CEO Ali Ghodsi. Clients already include multiple Fortune 500 companies. Funds will be used to expand engineering teams, deepen enterprise integrations, and advance research in trusted architectures, memory systems, and multi-agent coordination.

Xiaomi founder Lei Jun announced that MiMo-V2-Pro ranked fifth globally on Text Arena's Model Rank, fourth in LabRank measuring lab R&D strength, and fifth in Code Arena. The model exceeds 1 trillion total parameters with 42 billion active parameters, employs a hybrid attention architecture, supports 1 million token context length, and is engineered for high-intensity applications in the agent era. Text Arena uses a 'double-blind testing' mechanism where global real users vote on answer quality, ensuring credibility. It ranks eighth globally and second domestically on the Artificial Analysis leaderboard.

Chinese AI chatbot DeepSeek experienced a major service disruption on March 30, with its flagship models R1 and V3 offline for 7 hours and 13 minutes—the longest outage since user traffic surged in early 2025. The incident occurred overnight and service was restored at 10:33 AM local time. The company did not disclose the root cause. While DeepSeek faced API issues in January 2025, this marks the first extended user-facing outage exceeding two hours. Industry observers are watching for updates on its next-generation model roadmap, but the company has made no disclosures.