AI News Flash · Daily Brief

Anthropic's Claude Mythos Preview sets a new GPQA Diamond record at 94.6%.

By and · Editorial team, AI News Flash

Capabilities

Anthropic's Claude Mythos Preview sets a new GPQA Diamond record at 94.6%.

Anthropic's Claude Mythos Preview has taken the top spot on the LLM Stats leaderboard for GPQA Diamond, posting a 94.6% score that surpasses the previous best of 94.3% set by Gemini 3.1 Pro as of February 2026. GPQA Diamond is widely regarded as one of the most discriminating frontier reasoning benchmarks, testing PhD-level knowledge across biology, chemistry, and physics. The score is continuously refreshed on the leaderboard, and no independent third-party audit has been published. The result extends Anthropic's lead over all currently tracked models on this category of scientific reasoning.

Why it matters: AI developers and enterprises evaluating frontier models for scientific reasoning now have a new performance ceiling to benchmark against.

Google's Android Bench July update ranks eight models on real mobile dev tasks.

Google's Android Bench leaderboard added eight models in its July release, evaluating them on real-world Android development challenges such as Jetpack Compose migrations and wearable networking using the updated Harbor evaluation framework. Claude Fable 5 leads the overall standings at 84.5, with GPT-5.5 close behind at 80.2. Among open-weight models, GLM 5.2 tops the category at 72.2 and Kimi K2.7 Code follows at 70.4, marking the first publicly scored head-to-head comparison of these Chinese open-weight models on a task-grounded mobile development benchmark. The results give Android developers concrete, task-specific data for choosing a coding assistant.

Why it matters: Mobile developers and enterprises building Android tooling now have the first task-grounded leaderboard data comparing leading proprietary and open-weight models directly.

Technology & Research

Dual-Path Architecture paper proposes joint scaling of compute and model capacity

A new paper on arXiv (2605.30202) proposes a dual-path Transformer architecture that decouples model capacity from inference compute by maintaining a dense capacity path alongside a separate compute-efficient speed path. Standard dense and mixture-of-experts designs tie parameter count and FLOPs together, creating a persistent trade-off between model quality and inference cost. The dual-path approach allows each dimension to be scaled independently, and early benchmark results indicate competitive performance at lower active-parameter budgets. Code and checkpoints are referenced in the paper, though independent reproduction has not yet been confirmed by outside researchers.

Why it matters: AI researchers and infrastructure teams pursuing cheaper inference without sacrificing model quality have a new architectural approach to evaluate and reproduce.

Regulation & Policy

Bipartisan Senate bill would require visible disclosures on all AI-generated media.

Senators Brian Schatz (D-HI), John Curtis (R-UT), and Mark Warner (D-VA) introduced the AI Labeling Act of 2026, a bipartisan bill that would require generative AI providers to attach both visible and machine-readable disclosures to AI-generated audio, video, and image content. Large online platforms reaching at least 10 million monthly U.S. users or generating $1.5 billion or more in annual revenue would be prohibited from removing those disclosures. The Federal Trade Commission would enforce the requirements, and NIST would convene a technical working group to develop detection and labeling standards. The bill creates direct compliance obligations for major AI developers and the platforms that distribute their content.

Why it matters: Generative AI developers and large platforms will face new federal disclosure obligations that reshape how AI-produced content is labeled and distributed.

States passed 109 AI laws in early 2026, even as federal preemption looms.

An NYU Center on Technology Policy analysis published July 7 found that U.S. states enacted 109 AI laws and 28 data center laws in the first half of 2026, a rate slightly below 2025 but still representing substantial legislative activity. The findings arrive as Congress continues to debate the Great American AI Act, whose three-year preemption clause would freeze state laws that specifically regulate AI model development. The enacted state measures reflect cross-partisan priorities: at least six states restricted AI use by health insurers, and several limited AI-enabled dynamic pricing. If the federal preemption clause passes, this wave of state-level regulation could be curtailed significantly before many laws take full effect.

Why it matters: Enterprises deploying AI and policymakers must prepare for a potential collision between an active state regulatory landscape and a sweeping federal preemption proposal.

AI Stocks

(MSFT) Microsoft launches $2.5B 'Frontier' AI-services unit, stock up ~3%

Microsoft unveiled a dedicated $2.5B business unit called Frontier this week, focused on AI services, sending its shares up roughly 3% and contributing to an approximately 7% gain in the iShares software ETF over eight sessions, even as the semiconductor SOXX index fell roughly 8.5% in the same period. The announcement comes as Microsoft approaches its late-July FY2026 close earnings report. Investors will focus on Azure growth and the trajectory of Microsoft's AI business, which is already running at a $37 billion annual revenue rate, a 123% year-over-year increase. The Frontier unit signals a structural commitment to packaging and selling AI capabilities as a distinct enterprise service line.

Why it matters: Enterprise technology buyers and investors will scrutinize Microsoft's Frontier unit as a signal of how large cloud providers plan to productize and monetize AI services at scale.

(NOW) Guggenheim upgrades ServiceNow to Buy at $125, citing 'extinction' valuation

Guggenheim analyst John DiFucci upgraded ServiceNow from Neutral to Buy on July 1, setting a $125 price target that sent shares up roughly 4% on the day. DiFucci's thesis is grounded purely in valuation compression following a roughly 49% decline in the stock over the prior year, explicitly not predicated on success in AI monetization. The upgrade places ServiceNow's July 22 Q2 earnings report in sharp focus as the first full quarter under the company's all-in AI licensing model. Investors will track cRPO growth against management's approximately 19.5% constant-currency guidance and the trajectory of the Now Assist ACV toward its $1.5 billion target to determine whether the valuation-only upgrade thesis holds.

Why it matters: Investors and enterprise software buyers will use ServiceNow's July 22 earnings to judge whether AI-driven licensing restructurings can restore growth after deep valuation declines.