According to the Financial Times, OpenAI is preparing the most significant overhaul of ChatGPT since its 2022 launch, transforming it into a 「super app」 that integrates programming tool Codex, AI agents, and third-party services such as Canva and Booking.com. The new version will roll out in the coming weeks, steering free users toward paid products like Codex. Codex's desktop application has surpassed 5 million weekly active users, a sixfold increase. Approximately 2 million enterprise customers contribute about 40% of OpenAI's revenue, expected to rise to 50% by year-end. Internal executives have reportedly declared 「chat is dead,」 with several senior leaders departing, while standalone projects like Sora video generation have been shelved. This strategic shift aims to counter Anthropic's Claude Code and lay groundwork for an IPO.

Per TechTimes, OpenAI is collaborating with Goldman Sachs and Morgan Stanley to prepare a confidential IPO filing as early as September 2026, targeting a valuation between $730 billion and $850 billion. This move follows closely after Anthropic filed its S-1 on June 1 with a $965 billion valuation, and currently achieves an annualized revenue of $47 billion. The public listing processes of these two AI giants mark the first time private AI valuations face scrutiny from public markets. Current AI industry valuations far exceed disclosed revenue levels, and post-IPO financial disclosures will test market acceptance of such high valuations.

Shortly after Anthropic's late-night release of Claude Opus 4.6, OpenAI responded with the launch of GPT-5.3-Codex, combining the coding capabilities of GPT-5.2-Codex with the reasoning strengths of GPT-5.2, achieving a 25% speed improvement. The model sets new benchmarks on SWE-Bench Pro and Terminal-Bench 2.0, capable of handling complex software engineering, web development, and long-horizon tasks, supporting real-time interaction and human intervention. It is already used internally at OpenAI for self-improving research and development. The model also excels in data analysis, document writing, and slide creation, demonstrating strong performance on the OSWorld computer operation benchmark, and has been integrated into ChatGPT's paid service.

Google has signed a $920 million monthly AI computing infrastructure agreement with SpaceX to meet surging demand for its Gemini Enterprise platform. The contract runs from October 2026 to June 2029, during which SpaceX will supply approximately 110,000 Nvidia GPUs and other high-performance hardware as 「bridge capacity.」 According to SEC filings, Google retains full ownership of its AI models, intellectual property, and data. Google may terminate the contract if SpaceX fails to deliver the committed capacity by September 30, 2026. This marks SpaceX's expansion into the AI infrastructure domain, with both companies also exploring long-term collaboration possibilities such as orbital data centers.

NVIDIA and SK Hynix have announced a strategic alliance to tackle the ongoing high-bandwidth memory (HBM) chip shortage projected to last until 2030. SK Hynix has pledged to double its wafer production capacity over the next five years to meet NVIDIA's demands for Blackwell and future architectures. Jensen Huang met with SK Group Chairman Chey Tae-won in Seoul, emphasizing that memory shortages will persist for 「several years,」 with the entire supply chain—including wafers, packaging, and silicon photonics—facing constraints. This strengthens NVIDIA's dominance in the AI data center market and enhances SK Hynix's competitive edge against Samsung and Micron, though it raises concerns about reliance on a single supplier.

Perplexity has introduced a new 「Search as Code」 (SaC) architecture that enables AI models to autonomously build search workflows by generating Python code instead of relying on fixed search APIs. The architecture consists of three layers: a model layer that formulates search strategies, a sandbox to execute code, and a底层 「Agentic Search SDK」 offering composable search modules. Compared to traditional methods, SaC allows parallel queries, programmatic filtering, and deduplication, significantly reducing contextual redundancy. In a test tracking 200 CVE vulnerabilities, SaC achieved an 85% reduction in token usage, with accuracy far surpassing competitors. It led OpenAI and Anthropic systems in four out of five benchmark tests and has been integrated into Perplexity Computer and Agent API.

AI leaders from Google, OpenAI, Anthropic, and Microsoft have publicly endorsed mandatory screening for synthetic DNA and RNA orders, requiring gene synthesis providers to verify customer identities and review sequence content before shipment. The initiative positions gene synthesis companies as critical biosecurity checkpoints to prevent the physical realization of harmful biological sequences designed by AI. The proposal has gained broad support from tech firms, biotech institutions, national security experts, and policy organizations. The U.S. Cotton-Klobuchar Act has prompted the Department of Commerce to develop verification standards, while NIST is building screening benchmark tools. Experts caution that screening alone may not be foolproof and advocate for multi-layered control mechanisms.

NVIDIA has acquired Kumo AI for $400 million, integrating its foundational model technology built specifically for relational data into NVIDIA's AI platform. Kumo AI's flagship product, KumoRFM, is the world's first foundation model designed for enterprise relational databases, enabling tasks such as churn prediction, fraud detection, and recommendations without training—20 times faster than traditional methods—and achieving 30%-50% higher accuracy after fine-tuning. The KumoRFM-2 version, released in April 2026, is based on a 「Relational Graph Transformer」 architecture, processing 5GB of data and 20 million lookups per second, scalable to 500 billion rows. It achieved 89% accuracy on the SAP SALT benchmark, significantly outperforming PhD-level data scientists using XGBoost at 75%.

At the 2026 Microsoft Build developer conference, Microsoft launched seven in-house MAI series AI models, including MAI-Thinking-1 for reasoning and MAI-Code-1-Flash, an accelerated coding model already deployed in GitHub Copilot and VS Code. MAI-Thinking-1 uses a MoE architecture with 35 billion active parameters, trained entirely on commercially licensed data with no involvement from third-party models, offering clearer IP assurance for enterprise clients in finance and law. After restructuring their agreement in May 2026, Microsoft no longer shares Azure AI revenue with OpenAI; instead, OpenAI now pays Microsoft, accelerating Microsoft's push for independent model development. MAI-Code-1-Flash is now available as an option and may become the default in the future.

Security research team Pluto disclosed a remote code execution (RCE) vulnerability (CVE-2026-4372) in the Hugging Face Transformers library. The flaw allows attackers to bypass the trust_remote_code=False safety check by crafting a malicious config.json file, triggering arbitrary code execution when users call from_pretrained() to load a model. The root cause lies in the library's use of setattr() to directly write configuration parameters into internal objects, enabling attackers to exploit the _attn_implementation_internal field to load malicious kernel code. The affected versions were downloaded approximately 232 million times before patching, potentially exposing sensitive information such as cloud credentials, API keys, and SSH keys. Users are advised to upgrade immediately and implement sandbox evaluations.