July 12th

OpenAI brought GPT-5.6 to general availability as a three-tier family—Sol (flagship), Terra, and Luna—with API pricing per 1M tokens of $5/$30, $2.50/$15, and $1/$6. Sol set a new high of 53.6 on Agents' Last Exam (13.1 points over Anthropic's Fable 5) and topped the Artificial Analysis Coding Agent Index at 80, while adding Programmatic Tool Calling and an ultra four-agent parallel mode—completing the Commerce Department clearance that had gated last month's staggered rollout. (OpenAI)

On launch day, GPT-5.6 became the preferred model in Microsoft 365 Copilot, deepening the OpenAI–Microsoft distribution tie-up and putting the new flagship in front of enterprise users immediately. (OpenAI)

SpaceXAI released Grok 4.5, its first model since the IPO, pitched as a cheaper "Opus-class" workhorse at $2/M input and $6/M output (versus Opus 4.7 at $5/$25) with claimed 2x token efficiency; Musk called it "roughly comparable to Opus 4.7, but much faster." (TechCrunch)

OpenAI also introduced GPT-Live, a full-duplex voice model family (GPT-Live plus a mini variant) that listens and speaks simultaneously and hands deeper reasoning off to GPT-5.5—pushing real-time voice past the turn-taking model. (OpenAI)

Apple sued OpenAI, alleging theft of hardware trade secrets—an escalation between two companies increasingly circling the same on-device AI ambitions. (The Verge)

Anthropic brought Claude Cowork to mobile and web for the first time (previously desktop-only), rolling out to Max subscribers first with other plans to follow. (The Verge)

Fidji Simo is stepping down from leading OpenAI's AGI work due to illness, moving to an advisory role—a notable leadership change at the top of the company's most ambitious effort. (The Verge)

Chinese open models are capturing serious U.S. demand as frontier-lab costs climb: the U.S. token share running on Chinese models (DeepSeek, Z.ai) has topped 30% every week since February 8, peaking near 46%, and Z.ai's GLM 5.2 saw first-week tokens jump roughly 27x and customers about 80x—a direct challenge to OpenAI and Anthropic pricing. (CNBC)

The Bank of England warned that an AI-stock crash could cut UK GDP by 2.2%, noting AI firms now make up roughly half of the S&P 500 and that expected 2028 hyperscaler capital spending has climbed from under $600 billion to over $1 trillion since December—concentration risk that now spans the real economy. (Politico EU)

AI giants are handing out large volumes of free compute to startups in a race to lock in market share early, per the Wall Street Journal—an aggressive land-grab that echoes the subsidized-growth playbooks of past platform wars. (WSJ)

Microsoft laid off roughly 4,800 employees across Xbox and commercial sales in its latest AI-era restructuring, part of a broader 2026 wave in which employers have explicitly cited AI—including Snap (~1,000, or 16%) and Intuit (~3,000, or 17%). (The Verge, TechCrunch)

New research complicates the doom narrative: workers may actually be safer at firms that embrace AI, which are increasing entry-level hiring, even as Stanford data show a roughly 20% drop in young-developer employment since late 2022 and high-AI-exposure graduate unemployment claims run 16,000–22,000 a month. (LA Times)

West Shore Home's CEO aims to double revenue to $2 billion with about 1,000 fewer hires (6,000 versus 7,000) using proprietary AI, with roughly 70% of 2026 sales appointments now triggering an AI scan—a concrete blueprint for growing output while shrinking headcount. (Fortune)

Employers are rewriting what they want from new hires: an analysis of 2.85 million job listings shows a shift toward versatile skills and human judgment AI can't replicate, and one bank's departmental generative-AI use went from 0% to 100% daily in three years. (Business Insider, HBR)

Illinois enacted a landmark AI Safety Measures Act, mandating annual third-party audits, 72-hour incident reporting, and penalties up to $1 million/$3 million, effective January 1, 2028—the most substantial state-level AI regulation yet and a likely template for others. (WTTW)

Unsealed emails revealed a tense standoff between Anthropic's Dario Amodei and DoD Under Secretary Emil Michael over the company's no-autonomous-weapons and no-domestic-surveillance redlines; Michael called the guardrails "just not workable" and pushed for "all lawful uses," and the DoD knew the limits before its $200 million July 2025 contract. (Gizmodo)

A lawsuit alleges a man used xAI's Grok to generate thousands of illegal CSAM images and that xAI reported only one prompt to authorities—spotlighting major gaps in the company's detection and reporting. (Ars Technica)

Anthropic faced backlash after being outed for a secret "Claude tracker" that monitored Chinese users—closing loopholes (Singapore-entity accounts, VPN reimbursements, "transfer station" relays) that gave Chinese firms Claude access, but sitting awkwardly against its own anti-surveillance stance. (Ars Technica)

In the New York Times copyright case, the paper alleges OpenAI faked an inability to search its training data and hid billions of logs—raising the evidentiary stakes in the highest-profile AI copyright fight. (Ars Technica)

The UN's ITU convened the first meeting of a new "AI for Good" Global Commission in Geneva, co-chaired by Salesforce's Marc Benioff and Rwanda's Paul Kagame, with Amazon's Andy Jassy, Anthropic's Jack Clark, Cohere's Aidan Gomez, Microsoft's Brad Smith, and NVIDIA's Jensen Huang—an attempt at global coordination as national rules diverge. (Axios)

The AI buildout's power bill is landing on U.S. manufacturers: data-center demand is straining the country's top grid operator and driving Rust Belt factory electricity bills up faster than other customers', squeezing steelmakers and brick plants and cutting against the "Made in America" push. (Ars Technica)

The rush to build custom AI silicon is jamming one narrow pipeline: Microsoft, Google, Amazon, Meta, OpenAI, Anthropic, and DeepSeek are all designing chips to cut NVIDIA reliance, but they compete for the same scarce TSMC/Samsung foundry capacity, packaging, HBM, and ASML lithography—one analyst warns some "won't see silicon for three years." (Axios)

Facing U.S. export controls, China's DeepSeek is about a year into an effort to design its own AI chips, hiring engineers to reduce its dependence on NVIDIA and Huawei. (Ars Technica)

Nokia is repositioning to supply networking and optical infrastructure for the AI data-center boom, a bid to turn the buildout into a second act. (WSJ)

AI is pushing toward general-purpose autonomy in robotics, with Agility's Digit humanoids already working in warehouses—and in a striking first, surgeon-controlled humanoid robots performed a world-first operation on live pigs at UC San Diego. (Ars Technica, Ars Technica)

Education is becoming AI's most contested proving ground: wealthy U.S. families and elite private schools (Alpha School, Forge Prep) are handing roughly two hours a day of core instruction to adaptive AI tutors, even as a Brown professor who switched to an in-person final saw scores fall about 50% and Nature reported AI-detection tools are so unreliable they flag human-written text as "almost 100% AI." (The Verge, Ars Technica, Nature)

Meta's Superintelligence Labs launched Muse Image—its first model, now powering image tools across the Meta AI app, Instagram, and WhatsApp—then quickly disabled the feature that let users generate AI deepfakes of public accounts after a privacy backlash, and followed up with Muse Spark 1.1, a multimodal reasoning model, plus a public preview of the Meta Model API. (The Verge, The Verge)

Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter open mixture-of-experts model with native 1M-token context and a new sparse-attention scheme—among the largest openly released models and further evidence of China's frontier-scale open-weights momentum. (MarkTechPost)

In healthcare, England's NHS App will use AI to route patients to appropriate services—reaching roughly 200,000 patients over the next year as part of a £10 billion overhaul. (The Guardian)