1. OpenAI's Astra Solves Ten Long-Open Problems in Mathematics
OpenAI announced that an internal version of its next-generation model, Astra, solved ten problems in mathematics and theoretical computer science that had been open for a decade or more — for roughly $2,000 in total compute. Headline results include the first explicit construction of a non-sofic group (a question open since Gromov introduced soficity in 1999), improved upper bounds on high-dimensional sphere-packing density down to the Cohn-Elkies threshold (the first improvement to that bound since 1978), and a refutation of Connes' rigidity conjecture. The proofs were formally verified with the Lean proof assistant, with human researchers authoring the papers.
2. Alibaba Releases Qwen3.8-Max, a 2.4-Trillion-Parameter Flagship — Open Weights Promised
Alibaba released Qwen3.8-Max on August 3, a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters — the largest and most capable Qwen to date. Alibaba's own numbers show it beating GPT-5.6 Sol and Claude Fable 5 on several evaluations, including PaperBench (93.0 vs 90.5 and 88.8) and Terminal-Bench 2.1 (86.6), though no independent evaluations exist yet. Notably, Alibaba says open weights are coming within a week — the first time a Max-class Qwen would be released publicly rather than kept behind an API.
3. DeepSeek V4 Flash Exits Preview With Aggressive Pricing
DeepSeek's V4 Flash officially left preview at $0.14 per million input tokens and $0.28 per million output tokens, scoring 82.7% on Terminal-Bench — reportedly beating DeepSeek's own 1.6-trillion-parameter V4 Pro on agentic benchmarks at a fraction of the cost. The release continues the Chinese lab's pattern of undercutting frontier pricing by an order of magnitude and keeps pressure on Western labs in the fast-inference tier.
4. Trump Abruptly Cancels Signing of AI Cybersecurity Executive Order
President Trump canceled a scheduled Oval Office signing ceremony for a new AI executive order, reportedly over concerns the text could undermine America's lead over China. The proposed order had emerged from growing pressure within the banking and financial sectors over AI-related cybersecurity risks. The reversal underscores the administration's tension between industry-friendly acceleration and mounting security concerns from regulated sectors.
5. US State-vs-Federal AI Regulation Battle Intensifies
Fallout from December's Executive Order 14365 — which directs federal agencies to preempt state AI regulation through lawsuits and withheld funding — continues to reshape the US compliance landscape. Colorado repealed and replaced its landmark AI governance law with the narrower SB 26-189, effective January 1, 2027, while new California CCPA regulations now require pre-use notices and opt-outs for automated decision-making. Child safety, AI infrastructure, and state procurement remain exempt from federal preemption.
6. Google Cancels Standalone AI Studio Mobile App Despite 800K Preorders
Google canceled its planned AI Studio mobile app for iOS and Android — despite drawing more than 800,000 preorders since I/O 2026 — and will instead fold its app-creation features into the Gemini app on mobile and desktop. The move consolidates Google's consumer AI surface area around a single Gemini brand rather than maintaining parallel developer-oriented apps.
7. MIT-IBM's PaTH Attention Outperforms RoPE on Long-Context Reasoning
Researchers at the MIT-IBM Watson AI Lab introduced PaTH Attention, a new positional-encoding approach that improves state tracking and sequential reasoning over long texts. On reasoning benchmarks it outperformed RoPE, the rotary positional encoding used in most of today's large language models — a potentially significant architectural shift given how widely RoPE is embedded in the current model stack.
8. Claude Sonnet 5 Intro Pricing Ends September 1
Anthropic's Claude Sonnet 5 introductory pricing window closes in under a month: input pricing rises from $2 to $3 per million tokens on September 1. Teams building on Sonnet 5 have a few weeks to lock in usage patterns and budgets before the roughly 50% input-price increase lands.
// KEY TAKEAWAYS
The clearest signal this week is AI crossing from benchmark-chasing into original research: Astra's ten Lean-verified solutions to decade-old open math problems, for about $2,000 in compute, is a qualitatively new kind of result. Meanwhile the China price-and-openness offensive escalates — Qwen3.8-Max claims frontier parity with open weights promised, and DeepSeek V4 Flash ships near-frontier agent performance at commodity prices. In Washington, AI policy remains in flux: a canceled executive-order signing and the ongoing federal preemption fight leave US compliance rules unsettled even as the EU AI Act enters full enforcement.