The day's tech, sifted: Jul 10, 2026

What matters today: OpenAI had its biggest launch day yet, splitting GPT-5.6 into three durable tiers (Sol, Terra, Luna) and pairing it with ChatGPT Work, an agent that turns a goal into finished docs, sheets, and sites, even as Artificial Analysis found the best AI agents still clear barely half of real office tasks and independent testers landed softer than OpenAI's own launch framing. The same day, OpenAI absorbed two blows: product chief Fidji Simo is stepping down after a severe illness flare-up, and news publishers led by the New York Times asked a judge to sanction the company over its copyright-case discovery conduct. Meta had an equally rough day: Chinese regulators are forcing it to unwind its $2B acquisition of Manus with Tencent stepping in, and the EU Commission found Instagram and Facebook's "addictive design" violates the DSA, even as a SemiAnalysis look at Meta Superintelligence Labs showed its compute ramp closing in on Anthropic and OpenAI. Money kept pouring into the buildout regardless: US VC funding hit a record $412.7B in H1, SK Hynix raised the largest-ever foreign IPO on a US exchange, and the Fed tapped Marc Andreessen for a new task force on AI's economic impact.

AI / LLMs

• OpenAI split GPT-5.6 into three durable capability tiers instead of one model: Sol for hard problems like coding and security research, Terra for high-volume business tasks, and Luna for fast, cheap everyday work, priced at $5/$30, $2.50/$15, and $1/$6 per million input/output tokens; Sol also became the first verified frontier model to beat a full ARC-AGI-3 game, scoring 7.8% against Opus 4.8's prior-best 1.5%, and GPT-5.6 is now the default model behind Microsoft 365 Copilot and GitHub Copilot.
• OpenAI launched ChatGPT Work, a GPT-5.6 agent that turns a goal into finished sheets, slides, docs, and web apps, staying on multi-step projects for hours and checking in for approval on key actions; it's rolling out today to Pro, Enterprise, and Edu plans, with Plus and Business following within days.
• Independent testing landed softer than OpenAI's launch framing: product consultant Claire Vo's own benchmark had Sol beating Claude Fable 5 on prototyping and PRDs but still preferred Sonnet 5 for agentic voice work, while Factory's Droid platform folded Sol, Terra, and Luna into its automatic task-routing rather than crowning a single winner.
• Artificial Analysis launched an independent leaderboard for ServiceNow's enterprise-agent benchmark and found even the best model clears only 51% of real office tasks, a sober counterpoint to a day full of frontier-model victory laps.
• Anthropic will start usage-based billing for Claude Fable 5 on July 12, pushing subscribers to pay per-token for its best consumer model while working to return it to subscription plans "when sufficient capacity allows," Wired reports, a sign of how tight frontier compute still is.
• Apple has reportedly held meetings with startup PrismML about running much larger AI models directly on iPhones: PrismML has already shrunk Alibaba's 27B-parameter Qwen 3.6 down to run entirely on an iPhone Pro.
• GLM 5.2 prepared a real UK small business's quarterly VAT return, processing 59 transactions in 68 minutes for $2.73 in token costs and landing within 7 pence of the human-prepared figure, one of the more concrete "AI does real knowledge work" demonstrations yet from a Chinese open-weight model.
• Cognition published a trustworthiness evaluation alongside SWE-1.7, its coding model built on Moonshot's Kimi K2.7 Code base, finding the post-trained SWE-1.7 refused problematic requests the base Kimi model complied with, and scored comparably to US frontier labs on the same trust evals.

Devtools & Infra

• pgrust, a from-scratch Rust reimplementation of Postgres, now passes all 46,000-plus queries in Postgres's own regression suite while staying disk-compatible with Postgres 18.3 installations; it's not production-ready or performance-tuned yet, but the goal is a multithreaded, connection-pooled Postgres without C's memory-safety baggage.
• Zig creator Andrew Kelley published a pointed rebuttal to Bun's widely-shared Rust-rewrite writeup, calling its claim that the old Zig codebase wasn't fuzzed "an outright fabrication" and arguing the binary-size and performance wins the Anthropic-backed team touted were reachable in Zig too, if compile times had gotten the same attention.
• Ghostty and Zig's earlier departures from GitHub are becoming a small trend: Gentoo is now moving parts of its infrastructure to Codeberg too, citing 112 hours of GitHub downtime across 48 outages over the past year, even as GitHub itself keeps growing (36 million new developers in 2025).
• Google shipped LiteRT.js, a high-performance runtime for running AI models client-side in the browser, claiming up to 60x speedups over CPU execution when a GPU or NPU is available via WebGPU or WebNN, and roughly 3x over existing browser AI runtimes on the same hardware.
• GitHub rebuilt Copilot's code review by giving the agent narrower, purpose-built tools instead of broad filesystem access: a general-purpose grep tool actually made review worse by inviting sprawling investigations, while scoped tools built for the review task cut noise and improved precision.

Security & Privacy

• The EU Parliament extended the "Chat Control 1.0" CSAM-scanning derogation to 2028 via a rushed pre-recess procedure: 314 MEPs voted against it, 47 short of the 361-vote absolute majority second-reading rules require to block it. Days earlier, the US House passed the KIDS Act 267-117, a package EFF says would force privacy-eroding age verification onto the whole internet; it now heads to an uncertain Senate vote.
• The European Commission preliminarily found that Instagram and Facebook's auto-play, infinite scroll, and personalized recommendations amount to addictive design that violates the Digital Services Act, saying the features "shift the brain into autopilot mode" and telling Meta to fix them or face fines.
• The European Commission declined to extend the Digital Markets Act's interoperability mandate to social networks in its first DMA review, citing "no clear demand," leaving Meta and TikTok free to keep blocking the cross-platform switching the law was meant to enable.
• Following EFF's reporting on a Texas sheriff's office searching 83,000-plus license-plate-reader cameras to track a woman suspected of self-managing an abortion, the nonprofit Mayday Health put up Houston billboards warning drivers they may be surveilled, expected to reach a million drivers over four weeks, a low-tech answer to a high-tech dragnet.
• Microsoft patched RoguePlanet (CVE-2026-50656), a Windows Defender zero-day that let remote attackers gain admin control even with real-time protection disabled; the same researcher who disclosed it in June has now sent Microsoft scrambling on several zero-days in a row.

Startups & Industry

• Fidji Simo is stepping down from her full-time role leading OpenAI's product and business work after a severe flare-up of Postural Orthostatic Tachycardia Syndrome, a chronic condition she's had since 2019; she'll become a part-time adviser, with her responsibilities split between Greg Brockman, CFO Sarah Friar, and chief strategy officer Jason Kwon.
• News organizations led by the New York Times asked a judge to sanction OpenAI, alleging the company claimed for years it couldn't search its training data or chat logs while internally running a "Bloom filter" system called Project Giraffe and sitting on roughly 78 million de-identified ChatGPT conversations it used to gauge its own infringement exposure.
• Chinese regulators are forcing Meta to unwind its $2B acquisition of AI startup Manus, with Tencent and Manus's original backers ZhenFund and HSG in talks to buy it back at the same valuation; Beijing cited national-security concerns over foreign ownership of Chinese-rooted AI tech, and Tencent would take the largest stake while staying a minority holder, keeping control onshore. The same day, a SemiAnalysis look at Meta Superintelligence Labs found its RL-environment work and compute ramp closing in on Anthropic and OpenAI, a sharper trajectory than the regulatory setback suggests.
• US venture funding hit a record $412.7B in H1 2026, up 30% on all of 2025, with AI startups pulling in $355.9B of that, 86% of every dollar; PitchBook warned the concentration leaves the market exposed if AI returns disappoint, since venture's power law means any gains would concentrate in a handful of winners anyway.
• SK Hynix priced the largest-ever US IPO by a foreign company at $149 a share, raising $26.5B in a book more than 7x oversubscribed, then opened 14% higher at $170 on its Nasdaq debut; the memory chipmaker plans to funnel proceeds into AI-memory capacity like its Yongin cluster.
• Fed Chair Kevin Warsh named Marc Andreessen and Xbox CEO Asha Sharma to a task force studying AI's economic impact, alongside economist Charles Jones, on leave from Stanford to Anthropic's policy institute; all three are AI bulls like Warsh himself, and the task force is due to report by year-end.
• Capital kept finding new corners of the AI stack to fund: 23-year-old Mercor acquired RL-environment builder Deeptune three months after backing its $43M Series A, scraping-infrastructure startup Oxylabs took its first outside money at a $3.6B valuation from Warburg Pincus, and Carlyle sold data-center power platform Copia to EQT for $2.6B, a fivefold return underscoring how much private equity wants exposure to AI infrastructure right now.
• Humanoid robots closed in on human capability from two directions at once: 1X unveiled a tendon-driven hand for its Neo robot with 25 degrees of freedom, 22 in the fingers and palm plus 3 at the wrist, strong enough to lift a 20-pound kettlebell yet gentle enough to pick a grape off its stem; separately, teleoperated humanoid robots removed gallbladders from live pigs in a UC San Diego preclinical trial published in Nature, with human surgeons piloting one and two robots working together on a second procedure, pitched as bringing surgical-robot capability to rural clinics at a fraction of the cost of dedicated systems.
• Microsoft's carbon emissions rose 25% in 2025 to 34 million metric tons, attributed mainly to data center expansion, the same week an analysis found the five biggest data center spenders, Alphabet, Amazon, Meta, Microsoft, and Oracle, added roughly $350B in debt over the past five years to fund the buildout.

Hacker News

OpenAI's GPT-5.6 launch (816 comments) and same-day ChatGPT Work push, both covered above, anchored the front page, alongside Tencent's own Hy3 model release, which landed with far less discussion. The EU's Chat Control extension and the Commission's Instagram addictive-design finding, both detailed above, drew heavy front-page traffic too. On the DIY end, one hacker wrangled GLM 5.2 onto underpowered hardware via colibri, a fun companion to GLM's bookkeeping benchmark covered elsewhere; Meta's Muse Spark 1.1 Model API, launched yesterday, kept drawing HN traffic today. The sharpest fight was Zig creator Andrew Kelley's rebuttal to Bun's Rust-rewrite writeup, calling its fuzzing claims fabricated, a pointed hit on the Anthropic-backed Bun team; a companion interview with Mitchell Hashimoto on Zig and Ghostty rode the same wave. Malisper's pgrust, covered above, drew admiration but no one's calling it production-ready yet.

Defense-tech readers compared notes on the Army's brittle logistics (413 comments) and the scramble for cheap hunter-killer drones after Iran destroyed $1B of Reapers (311 comments), both fretting over expensive gear versus attritable hardware. Pangram's look at AI content flooding social feeds, especially LinkedIn, struck a nerve given the day's Instagram design ruling. Show HN's minimalist 18 Words racked up outsized traction (297 comments) with little pushback.

Threads

• OpenAI ran a contradictory day: launching GPT-5.6 and ChatGPT Work with maximum bravado while, hours apart, losing product chief Fidji Simo to illness and getting hit with a sanctions push in the NYT copyright case, and Artificial Analysis quietly undercut the whole industry's agent-benchmark hype.
• Meta had its own split day, a regulatory pincer (Beijing's forced Manus unwind, the EU's addictive-design finding) landing the same day SemiAnalysis reported its AI compute ramp gaining real ground on Anthropic and OpenAI.
• A pattern keeps recurring: a lab ships a flashy self-reported benchmark win (Sol's ARC-AGI-3 score), and independent testers (Claire Vo, Factory, Artificial Analysis) land on a far more mixed verdict, the same dynamic that drove OpenAI to retract its own SWE-Bench Pro recommendation last week.
• Regulators leaned on both ends of the funnel at once: the EU's Chat Control extension and Instagram addictive-design finding, and the US House's KIDS Act, all advancing over real floor or industry opposition the same week.
• Money chased AI infrastructure from every direction: SK Hynix's record IPO debut, a record VC half-year, the Carlyle-EQT Copia sale, Oxylabs' first outside raise, and the Fed's new Andreessen-led task force, the same week hyperscalers' data center debt load and Microsoft's own emissions both came into sharper view.