AI News Roundup: OpenAI Drops GPT-5.4-Cyber, Stanford’s 2026 Index Lands, NVIDIA Hits 10-Day Streak

OpenAI Unveils GPT-5.4-Cyber to Counter Anthropic’s Mythos

OpenAI on Tuesday introduced GPT-5.4-Cyber, a variant of its flagship model fine-tuned for defensive cybersecurity work, exactly one week after Anthropic stunned the industry with its own security-focused Claude Mythos Preview. The model will be rolled out on a limited basis to vetted security vendors, government agencies, and academic researchers because of what OpenAI called its “more permissive” design around vulnerability discovery and exploit reasoning.

The move underscores how quickly the frontier labs are racing to claim the cybersecurity beachhead. OpenAI says GPT-5.4-Cyber can audit large codebases, generate proof-of-concept exploits inside sandboxed environments, and chain together findings into full attack paths — the same capabilities Anthropic warned were too dangerous to ship publicly. Both companies are now effectively gatekeeping who gets access to the most capable security models, and the question of whether that privatized defense advantage holds against open-weight Chinese rivals is the dominant policy debate this week.

Anthropic’s Project Glasswing Finds Thousands of Zero-Days

Anthropic’s Project Glasswing — the controlled deployment of Claude Mythos Preview that triggered OpenAI’s response — has now identified thousands of zero-day vulnerabilities across every major operating system and web browser, the company disclosed this week. Partners with early access include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, and NVIDIA, and Anthropic has committed up to $100 million in model credits to fund the program.

The Bank of Canada has now convened with major banks and financial institutions to assess Mythos’s implications for systemic financial cybersecurity, and the U.S. is reportedly considering treating the underlying weights as export-controlled. Bruce Schneier, writing this week, called Glasswing “the first credible attempt at giving defenders a durable advantage” in the AI era — while warning that the same model in adversarial hands would be catastrophic.

Stanford AI Index 2026: China Nearly Closes the Gap

Stanford HAI released its 2026 AI Index this week, and the headline finding is that the U.S. performance lead over the top Chinese model has collapsed from 9.26% in January 2024 to just 1.70% by early 2026. Despite the U.S. producing more models and attracting far higher private investment, Chinese labs like DeepSeek, Moonshot, and Alibaba have effectively closed the quality gap on every major benchmark.

The capability numbers are otherwise historic: SWE-bench Verified scores climbed from 60% to nearly 100% in a single year, and the best models now exceed 50% on Humanity’s Last Exam, up from just 8.8% in 2025. Organizational AI adoption hit 88%, and Stanford pegs the value of generative AI to U.S. consumers at $172 billion annually. The trust numbers, though, are grim — the Foundation Model Transparency Index dropped from 58 to 40, and only 33% of Americans expect AI to make their jobs better.

NVIDIA Logs Longest Winning Streak Since 2023

NVIDIA stock is now on a 10-day winning streak, up roughly 18% over that stretch — the chipmaker’s longest run of consecutive gains since a similar streak in 2023. Investors are reacting to a wave of compute-deal announcements and Tuesday’s release of NVIDIA’s Ising family of open-source models aimed at accelerating quantum-computing workloads.

The rally also reflects the staggering scale of new infrastructure commitments. Anthropic last week announced it will receive roughly 3.5 gigawatts of next-generation TPU capacity from Google and Broadcom starting in 2027, and Meta added $21 billion to its existing $14 billion CoreWeave deal. Globally, AI data centers can now draw 29.6 gigawatts — enough to power the entire state of New York at peak demand.

Anthropic Surpasses OpenAI in Run-Rate Revenue

Anthropic disclosed this week that its run-rate revenue has crossed $30 billion, up from roughly $9 billion at the end of 2025, and the number of business customers spending more than $1 million annually has surpassed 1,000 — doubling in less than two months. Multiple analysts now place Anthropic ahead of OpenAI on enterprise run-rate, the first time the smaller lab has held that lead.

The growth is concentrated in coding (Claude Code and the Claude Sonnet 4.6 API), agentic tooling, and the wave of Mythos-driven security deals. Q1 venture funding overall hit a record $300 billion, with AI capturing $242 billion — about 80% of all global VC.

OpenAI Acquires Hiro Finance in Sixth Deal of the Year

OpenAI quietly closed its acquisition of personal-finance startup Hiro Finance on Monday, marking its sixth acquisition of 2026 — nearly matching its full-year 2025 total. Terms weren’t disclosed, but the deal slots Hiro’s consumer-finance reasoning team into OpenAI’s push toward verticalized agents that handle everything from budgeting to tax filing on a user’s behalf.

Meta Ships Muse Spark Multimodal Model

Meta launched Muse Spark this week, a new multimodal model designed to power features across Facebook, Instagram, WhatsApp, and Meta’s smart-glasses lineup. The model coordinates multiple specialized sub-agents and ships with both fast-response and deeper reasoning modes — Meta’s clearest answer yet to the GPT-5.4 / Claude Sonnet 4.6 / Gemini 3.1 Pro tier.