OpenAI co-founder and president Greg Brockman has officially taken over the company's product strategy, planning to integrate ChatGPT, Codex, and developer APIs into a unified 'agent-based platform'—a desktop super app combining chat, programming, and browsing. This move comes during Fidji Simo's medical leave as head of AGI deployment and marks a key strategic shift ahead of Google I/O. In an internal memo, Brockman emphasized focusing resources on an 'agentic future,' targeting both consumer and enterprise markets. CEO Altman previously announced the suspension of non-core projects like Sora and OpenAI for Science. The company expects a $14 billion loss in 2026 and faces challenges including executive turnover and litigation from Musk, but is still advancing toward a potential IPO by year-end.

On May 17, xAI launched Grok 4.3 Beta, its next-generation reasoning large model, currently in early access and available exclusively to SuperGrok subscribers. The model has approximately 500 billion parameters, uses a Mixture-of-Experts (MoE) architecture, and supports a context length of up to 2 million tokens. It accepts and generates text, with core improvements focused on enhanced long-context processing and potential native video understanding. It retains the multi-agent collaboration framework from Grok 4.2, where multiple specialized agents work in parallel to improve accuracy. API pricing is set at $3 per million input tokens and $15 for output. The model is closed-source, and no official benchmark results have been released. Hermes Agent v0.14.0 now supports integration with Grok 4.3 via SuperGrok OAuth, supporting up to 1 million tokens of context.

NVIDIA has introduced SANA-WM, a 2.6-billion-parameter open-source world model based on the Diffusion Transformer architecture, capable of generating 60-second 720p videos on a single GPU. Key innovations include frame-level Gated DeltaNet and hybrid attention mechanisms to address memory bottlenecks in long sequences; dual-branch camera control enabling precise 6-DOF trajectory following; and a two-stage generation process that reduces visual drift ΔIQ from 3.09 to 0.31. On 8 H100 GPUs, throughput reaches 22.0 videos/hour—36x faster than LingBot-World. A distilled version can complete denoising in 34 seconds on a single RTX 5090. Trained on 64 H100s over ~18.5 days, the code is now open-sourced under Apache 2.0 at NVlabs/Sana.

Google DeepMind has released Veo 3, its latest text-to-video model, capable of generating videos up to 60 seconds long with native audio synthesis, including dialogue, sound effects, and ambient sounds. This sets it apart from prior models like Sora and Runway Gen-3 Alpha, which require post-hoc audio addition, potentially saving creators significant editing time. Veo 3 is accessible via three channels: Google One AI Premium ($19.99/month), VideoFX on Google Labs (waitlist required), and the Gemini API for developers—currently in limited access. Model architecture, parameter count, and evaluation metrics remain undisclosed. Future scrutiny will focus on API rate limits, content moderation policies, and lip-sync accuracy in extended conversations.

OpenAI has launched a personal finance tool for ChatGPT, allowing users to securely connect bank accounts from over 12,000 financial institutions—including Chase, Fidelity, and Robinhood—via Plaid using read-only access. The tool enables spending analysis, investment tracking, subscription management, and financial planning. Currently in preview for ChatGPT Pro users in the U.S., it runs on GPT-5.5 and was co-optimized with financial experts. For privacy, data is deleted within 30 days after disconnecting an account, and users can manually clear financial memories. However, financial conversations are included in model training by default unless users opt out. This launch follows OpenAI’s acquisition of the personal finance startup Hiro team. Over 200 million users now ask ChatGPT financial questions monthly. On the same day, OpenAI faced a class-action lawsuit alleging it embedded Meta Pixel and Google Analytics tracking tools in ChatGPT.

On May 16, OpenAI announced a partnership with the Maltese government to offer all residents of Malta free one-year access to ChatGPT Plus, contingent upon completing a free course on AI usage. The program, launched in May and gradually expanding, is also open to Maltese citizens living abroad. Malta becomes the first country globally to implement such an initiative. The curriculum, designed by the University of Malta, includes three mandatory modules: AI Fundamentals, AI in Daily Life, and AI Learning. This falls under OpenAI’s 'OpenAI for Countries' initiative, which previously collaborated with Estonia and Greece’s education systems. OpenAI has not disclosed financial terms of the agreement.

On May 17, 2026, Anthropic’s Claude surpassed OpenAI in paid commercial AI adoption market share, reaching 34.4% compared to ChatGPT’s 32.3%. This reversal is primarily driven by strong developer support for Claude Code and a precise go-to-market strategy targeting professional roles. Meanwhile, the open-source model ecosystem continues to thrive, with DeepSeek V4-Flash and lightweight frameworks like llama.cpp drawing attention to emerging 'activation steering' techniques for direct manipulation of LLM internals. According to CAISI evaluations, open-source models overall lag behind U.S. closed-source frontier models by approximately 3 to 7 months, though assessment limitations may underestimate actual capability gaps.

Alibaba Health has launched 'Hydrogen Ion,' a medical AI product designed to provide China’s 5 million doctors with an evidence-based AI assistant. The system employs a four-layer evidence-based AI architecture to ensure each response is supported by authoritative literature, enabling traceability to original text passages. A key breakthrough is its exclusive partnership with BMJ Group, granting access to the full content of 70 journals over the past decade—making it the only medical AI in China with direct access to top-tier BMJ publications. It has also integrated resources from domestic authorities including the Chinese Medical Association and People's Medical Publishing House. Supporting natural language, voice, and image inputs, the tool reduces physicians’ time to digest an SCI paper from 1–2 hours to just 3–5 minutes.