AI News of the Week (3rd April)
TLDR: Key AI Developments This Week
Anthropic's unreleased Claude Mythos dominated headlines after leaked materials described it as a step change in capability, while Google's TurboQuant breakthrough targets memory efficiency with extreme compression for LLMs. OpenAI closed a historic $122 billion funding round at an $852 billion valuation, and Microsoft launched three first-party MAI models while introducing multi-model intelligence in Researcher. Meanwhile, Anthropic suffered a Claude Code source leak, and OpenAI made Codex more accessible with pay-as-you-go pricing.
Claude Mythos Leak Turns Anthropic's Next Model Into the Week's Biggest AI Story
Anthropic's unreleased Claude Mythos dominated the week after leaked draft materials described it as a step change in capability and the most powerful model the company has built so far. Reports also said Anthropic is warning officials that Mythos could make large-scale cyberattacks materially more likely as frontier models become more autonomous, more precise, and far more useful for offensive security work. Read more
Google's TurboQuant Pushes AI Infrastructure Toward Smaller Memory Footprints and Lower Costs
Google Research's TurboQuant remained one of the most discussed technical breakthroughs of the week, because it directly targets one of the most expensive parts of running modern AI systems: memory overhead. Google says the method enables massive compression for large language models and vector search engines, reinforcing the idea that the next big gains in AI may come not just from bigger models, but from making existing systems far more efficient. Read more
OpenAI Closes a $122 Billion Funding Round and Doubles Down on Platform Scale
On March 31, OpenAI said it closed a $122 billion funding round at an $852 billion post-money valuation, giving it one of the largest private financings ever. The company said the new capital will help accelerate the next phase of AI, and the timing makes the broader strategy clearer: OpenAI is trying to strengthen its position not just as a model lab, but as core infrastructure for consumer, developer, and enterprise AI. Read more
Microsoft Launches MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 as First-Party Models
Microsoft used the week to show it wants to rely less on outside labs by launching three in-house models inside Microsoft Foundry. The company says MAI-Transcribe-1 delivers enterprise-grade speech recognition across 25 languages, MAI-Voice-1 can generate 60 seconds of expressive audio in under one second on a single GPU, and MAI-Image-2 is now its highest-capability text-to-image model. Read more
Microsoft Also Starts Combining Frontier Models Inside Researcher Instead of Picking Just One
Microsoft's other notable move this week was architectural rather than purely product-level. Its Researcher agent for Microsoft 365 Copilot now adds multi-model capabilities called Critique and Council, with Microsoft saying the system can separate generation from evaluation and use combinations of models from frontier labs including Anthropic and OpenAI to improve accuracy, depth, and confidence. Read more
Anthropic Leaks Part of Claude Code and Reveals More of Its Agent Roadmap Than It Meant To
Anthropic also suffered a self-inflicted security embarrassment this week after a release packaging error exposed part of Claude Code's internal source base. Reporting said the leak involved nearly two thousand files and roughly half a million lines of code, with observers spotting hints of unreleased ideas including a Tamagotchi-style coding pet and an always-on agent, even though Anthropic said no customer data or credentials were exposed. Read more
OpenAI Makes Codex Easier to Buy With Pay-As-You-Go Pricing for Business Teams
OpenAI also made a quieter but important commercial move by changing how teams can adopt Codex. Starting this week, ChatGPT Business and Enterprise workspaces can add Codex-only seats with pay-as-you-go pricing and no fixed seat fee, while OpenAI also lowered the annual price of standard ChatGPT Business seats to $20 per user per month, making pilots and narrower rollouts much easier to approve. Read more
