1. OpenAI's Astra Solves Ten Open Math Problems With Machine-Checkable Proofs
OpenAI announced that an internal version of Astra, its next model family, solved ten problems in mathematics and theoretical computer science that had each been open for at least a decade — including the first explicit construction of a non-sofic group, resolving a question standing since Gromov introduced soficity in 1999. Every solution ships with a Lean 4 certificate published on GitHub under Apache 2.0, so anyone can verify the proofs by compiling them — no peer review required. Total compute cost for all ten: roughly $2,000. Mathematician Thomas Bloom of erdosproblems.com called the results "big news," and the milestone marks AI crossing from doing tasks to doing original research.
2. White House Finishes Frontier-Model Framework — and Keeps It Secret
The White House met its August 1 deadline to complete a voluntary evaluation framework for frontier AI models, mandated by a June executive order. The framework gives the federal government up to a 30-day preview window on new frontier models before broader release, with government and developers jointly selecting "trusted partners" for early access. On August 4, Meta, Nvidia, Microsoft, OpenAI, Anthropic and smaller firms reviewed the draft in Washington — but the benchmarks and thresholds are classified, and the administration won't publicly release the framework itself.
3. Anthropic Releases Claude Opus 5: Near-Flagship Performance at Half the Cost
Claude Opus 5 launched July 24 across all Anthropic platforms, delivering performance close to the company's top model on many tasks at half the price — $5 per million input tokens and $25 per million output, unchanged from Opus 4.8. It posts leading results on software engineering and knowledge-work benchmarks, and Anthropic calls it "the most aligned Opus model" yet. A new effort dial lets users trade capability against token spend, preserving much of the performance at lower cost settings.
4. OpenAI Cuts GPT-5.6 Luna Pricing 80% as ChatGPT Nears 1 Billion Weekly Users
OpenAI slashed the price of GPT-5.6 Luna, its lightweight tier, by 80% to $0.20 per million input tokens, continuing the industry-wide race to commoditize inference. ChatGPT meanwhile reached roughly 1 billion weekly active users. The GPT-5.6 family — Sol, Terra, and Luna — bundles capability tiers with safety features and government access provisions, reflecting how model launches now package pricing and policy together.
5. AI Money Floods the 2026 Midterms
Millions of dollars tied to the AI industry are pouring into the 2026 US midterm elections, with interest groups funded by AI leaders split over how the government should oversee the technology. Innovation Council Action, a political group tied to two Trump advisors, announced it will spend at least $100 million. The spending surge comes as federal preemption of state AI laws remains contested and binding regulation looms on both sides of the aisle.
6. MIT-IBM Researchers Debut PaTH Attention, Outperforming RoPE on 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 outperforms RoPE, the rotary positional encoding used in most of today's large language models — a reminder that core architectural components of the transformer stack are still being rethought even as frontier labs race on scale.
7. EU AI Act Enters Full, Strict Enforcement
The European Union's AI Act is now in full enforcement, setting the de facto global standard for AI compliance. High-risk system obligations, transparency requirements, and general-purpose AI rules are all now binding, and companies shipping into the EU market are treating the Act's requirements as their worldwide baseline. Combined with the US voluntary framework, August 2026 marks the moment AI became cheaper to run but harder to ship without oversight.
8. States Stay in the Driver's Seat: Colorado Replaces Its Landmark AI Law
With federal AI legislation stalled, US states remain the primary source of binding AI regulation. Colorado — which passed the country's most comprehensive state AI governance law with SB 24-205 — repealed and replaced it in May 2026 with SB 26-189, a narrower statute regulating automated decision-making technology that takes effect January 1, 2027. The move illustrates the churn companies face as a December 2025 executive order pushes to preempt state AI laws while carving out child safety, data-center infrastructure, and state procurement.
// KEY TAKEAWAYS
AI crossed a visible research threshold this week: Astra's ten machine-verified proofs of long-open math problems, for about $2,000 in compute, move frontier models from doing tasks to doing original science. At the same time inference keeps getting radically cheaper — an 80% cut on GPT-5.6 Luna, Opus 5 at half effective cost — while oversight tightens on every front: a secret US 30-day preview framework, the EU AI Act in full enforcement, and $100M+ of AI money flowing into the midterms. Cheaper to run, easier to use, harder to ship without someone watching.