Meta Throws Its Hat Into the AI Coding Ring With Muse Code
Meta made a bold move this week, launching Muse Code in beta — a terminal-based coding agent that marks the company's first serious foray into the rapidly growing developer tools market dominated by Anthropic's Claude Code and OpenAI's Codex.
Powered by the new Muse Spark 1.2 model, Muse Code is designed to tackle end-to-end software engineering workflows: planning changes, writing code, debugging, and validating results across large codebases. What sets it apart is its use of persistent background agents that build up context over time rather than starting fresh on each task, plus a crash-safe event log that lets it resume interrupted work seamlessly.
Meta is competing partly on price. Muse Spark 1.2 API access comes in at $1.25 per million input tokens and $4.25 per million output tokens — undercutting many rivals. The tool is available on macOS and Linux, though it currently lacks the dedicated app interfaces that Claude Code and Codex offer.
Early benchmarks show Muse Code outperforming Codex but trailing Claude Opus 5 on coding tasks. The release is the first coding-specific product to ship from Meta Superintelligence Labs (MSL), signaling that Meta is serious about monetizing its AI capabilities beyond the consumer space.
EU AI Act Transparency Rules Are Now Live — and They Have Teeth
As of August 2, the EU AI Act's transparency obligations under Article 50 are fully enforceable across all member states. This is the most significant regulatory milestone yet for AI governance, and it applies to existing systems from day one — no grandfathering.
The rules require AI providers and deployers to be transparent in four key areas:
- Direct interaction: Users must be told when they're communicating with an AI system rather than a human
- AI-generated content: Generative AI outputs must carry machine-readable markers
- Biometric systems: Emotion recognition and biometric categorization systems face additional disclosure requirements
- Deepfakes: AI-generated media and text on matters of public interest must be explicitly disclosed
The penalties are no joke: fines of up to €15 million or 3% of global annual turnover, whichever is higher. Generative AI providers already on the market get a limited grace period until December 2, 2026 to comply with the content-marking requirements, but every other obligation is immediately enforceable.
The European AI Office has published a voluntary Code of Practice on Transparency to help companies demonstrate compliance, with signatories benefiting from a presumption of conformity during enforcement reviews.
The White House Built a Secret Framework for Vetting Frontier AI Models
The White House quietly reviewed a new framework for evaluating frontier AI models at a closed-door meeting on August 4, inviting executives from OpenAI, Anthropic, Microsoft, Meta, and Nvidia — but it won't publicly release the document.
The framework, developed under Executive Order 14409 ("Promoting Artificial Intelligence Innovation and Security"), gives the government up to 30 days of pre-release access to frontier closed-source models. During this window, NSA-led reviewers evaluate whether the models could be used to discover software vulnerabilities or carry out sophisticated cyberattacks.
The program is voluntary, but its structure creates clear incentives for participation. Notably, open-weight models are explicitly excluded from the framework's scope, which critics argue creates a competitive asymmetry favoring open-source approaches while leaving closed-model developers under heavier scrutiny.
The secrecy has drawn sharp criticism from policy researchers. TechPolicy.Press raised five key unanswered questions, including what counts as "state-of-the-art" and how national security risks are defined — criteria that will determine which models fall under the framework.
Google's Brain Drain: Two Top Researchers Defect to Rivals
The reverberations from Google's talent exodus continue to reshape the competitive landscape. In June, two of the company's most prominent AI researchers left within 24 hours of each other: Noam Shazeer, co-lead of Google's Gemini project and co-author of the legendary "Attention Is All You Need" paper that introduced the transformer architecture, departed for OpenAI. The very next day, John Jumper, who won the 2024 Nobel Prize in Chemistry for his work on AlphaFold, announced he was joining Anthropic.
The departures cut deep. Shazeer was instrumental in building the transformer — the architecture that underpins virtually every modern AI system. His return to OpenAI (he briefly worked there years ago) gives Sam Altman's company one of the most respected minds in deep learning. Jumper's move, meanwhile, gives Anthropic a world-class scientist whose protein-folding breakthroughs demonstrated AI's potential to transform scientific research.
The losses have raised questions about Google's ability to maintain its position at the AI frontier. After strong showings in 2024 and 2025, the company's 2026 has felt like a step backward relative to OpenAI and Anthropic — the very labs it's losing talent to.
AI Labs Smash Lobbying Records as Policy Stakes Escalate
AI companies are pouring unprecedented sums into Washington influence campaigns. Both OpenAI and Anthropic broke quarterly lobbying records in Q2 2026, with Anthropic spending $1.97 million — a 26% jump from Q1 — and nearly tripling its year-over-year lobbying bill to $3.53 million for the first half of 2026.
The spending surge wasn't unprompted. Anthropic ramped up lobbying after the Commerce Department forced it to disable Claude Fable 5 and Claude Mythos 5 over national security concerns — a dramatic intervention that demonstrated just how high the regulatory stakes have become for frontier AI developers.
Anthropic's lobbyists have been meeting with lawmakers in both chambers of Congress, as well as officials at the White House, Commerce Department, and Treasury. The company now outspends Nvidia on lobbying and nearly matches Oracle — a remarkable shift for a company that was barely on K Street's radar two years ago.