Sony Music Publishing, Warner Chappell, and other music publishers sued Anthropic and co-founders Dario Amodei and Benjamin Mann late Friday in the Northern District of California. The suit alleges a "brazen campaign" of illegal torrenting and scraping of copyrighted works, including books containing lyrics and sheet music, to train Claude.
Anthropic was earlier ordered to pay $1.5B in the Bartz case over pirated works.
Andreessen Horowitz launched a $1.1B 'Machine Age' fund dedicated to AI's physical infrastructure: chips, memory, data centers, cooling, and robotics. The firm, long known for software bets, says it will back 'faster, more efficient systems' and 'power efficient edge devices' to fuel AI's real-world buildout.
OpenAI said Thursday it will start showing ads on ChatGPT's Free and Go tiers in India, where the company reported more than 100 million weekly active users back in February. The rollout begins with 50 brands, with agencies WPP and Omnicom as partners.
An ad manager arrives next month, letting marketers run campaigns with a daily minimum budget of ₹725 ($7.60). OpenAI started serving ads in the U.S.
in February and in Europe this month.
Perceptron, founded by ex-Meta FAIR researchers Armen Aghajanyan and Akshat Shrivastava, launched Isaac 0.5, an open-weight vision model built to help robots perceive, reason, and act in factories and warehouses. Trained on a million hours of general and ego-centric video, it targets manufacturing, logistics, security, and mobility.
The startup previously raised $16M in 2024 and is closing a new funding round.
Anthropic is launching a $5 million grant program to fund independent research on how AI affects user wellbeing. Grantees get funding, Claude access, and technical support to build open-source evaluations and benchmarks, covering conversations involving companionship-seeking or mental health crises.
Applications close September 21, 2026.
Hugging Face has been approached about a sale at a valuation of $13B or more, Business Insider reported, and has hired a bank to gauge bidder interest. Talks are early, no deal is signed and no bidder has been named.
The figure is nearly triple the $4.5B post-money valuation from its 2023 Salesforce Ventures-led round. CEO Clem Delangue says the firm is close to profitability; it rejected a $500M Nvidia investment at a $7B valuation earlier this year.
A free "stealth" model called Ox Alpha appeared on OpenRouter on Aug 20 and has set off a guessing game over who built it. The listing shows a 1,048,576-token context window, 131,072 max output tokens, text/image/video input and $0 pricing, with the provider staying anonymous through the preview.
Stripe CEO Patrick Collison called it "very impressive." Speculation has centered on Z.ai's GLM family, though one report later floated an unreleased Microsoft MAI.
Guidelight AI Standards, the nonprofit founded by former OpenAI safety lead Steven Adler, graded five frontier labs on six AI control practices using only public disclosures. Anthropic and OpenAI tied at C+ (2.50/5), followed by Google D+ (1.50), xAI D- (0.83) and Meta F (0.67).
No company scored above 3 on any practice, and Anthropic and Meta both scored 0 on having a containment plan for a misaligned model.
Nvidia published research arguing the harness, not the model, drives long-horizon agent performance. Its custom harness, Agentic Variation Operators (AVO), lifted Claude Opus 5 from 30% to 100% on the interactive reasoning benchmark ARC-AGI-3, clearing all 183 levels across 25 environments.
Opus 5's bare 30% was still the top score of any model tested. AVO's key addition: a "supervisor" agent that redirects the main agent when it stalls.
Pew Research found that 10% of a random sample of 10,000 English-language webpages collected in July 2026 show significant signs of AI authorship. Among pages published after ChatGPT's November 2022 release, the share rises to over one-third.
It skews by domain: about one-in-ten .com pages, 4.6% of .org, and roughly 1% of .edu and .gov. Pew ran ~490,000 Common Crawl pages from the past five years through the Open Pangram detection model.
Google is giving eligible college students a free 12-month Google AI plan: AI Pro (normally $19.99/mo) in the U.S. and AI Plus across 140+ other markets, redeemable until Dec 31, 2026 with a payment method required.
Also rolling out today to all Gemini app users: a student hub at gemini.google.com/students, study notebooks with diagnostic quizzes, interactive 3D visualizations, and Deep Research in Gemini Live, which runs multi-step reports in the background.
OpenAI launched ChatGPT for Teens, automatically enabled for eligible accounts identified as under 18 on Free and paid personal plans, rolling out globally from Aug 18. Age-appropriate protections are on by default, guided by OpenAI's Under-18 Principles in the Model Spec, and homework reminders trigger when possible cheating is detected.
Study mode adds guiding questions and quizzes, while parents get study hours, quiet hours and safety notifications.
404 Media planted a tracker in a rare book and followed it to VGT3, Amazon's Las Vegas facility where bulk-bought books are stripped of their spines and scanned for AI training, destroying the originals. Out-of-print titles are prized because nothing published before 2022 was written by an LLM; training on AI text risks "model collapse." Amazon says it buys books "through commercial channels to improve the products and services customers use."
Anthropic CEO Dario Amodei pushed back on investor Gavin Baker, who argued on the All-In podcast and X that Amodei's AI risk warnings fueled the US backlash — especially against data centers — and that he had 'lost the argument' on regulation. Amodei called it "fundamentally a crisis of trust": ordinary people don't trust companies, governments or the tech industry.
He conceded the fair critique is that "we haven't yet delivered on our big promises."
Anthropic detailed how Claude's new text watermark works. It uses Google DeepMind's SynthID-Text approach, nudging the randomness Claude uses to pick among equally valid words; Anthropic says quality is unchanged.
Future Claude models ship watermarked at launch worldwide, and models released before Aug 2, 2026 get it over the coming months. A detection API is planned but not live.
It survives light edits, a full rewrite removes it, and code is barely marked.
Z.ai released GLM-5.3, built on the same base as GLM-5.2 with every gain from scaled post-training. Z.ai calls it the top open-weights coding model: +50% on its in-house Code Bench, and Terminal Bench 3.0 jumps from 4.6 to 28.3.
Cyber skill grew faster than expected — 84.5 on CyberGym, ahead of GPT-5.6 Sol's 83.6 — and it found 2,436 vulnerabilities across 269 projects. Live now on the GLM Coding Plan and ZCode; weights land in two weeks after safety hardening.
DeepSeek has shipped DeepSeek-V4-Pro-0813, the GA build of its flagship model, superseding the preview. Open weights are on Hugging Face under MIT, with a 1M-token context, 384K max output and a DSpark speculative decoding module.
On DeepSeek's own numbers it scores 87.9 on Terminal Bench 2.1 and 62.7 on DeepSWE, above Opus-4.8 at 85.0 and 58.0, but trails on HLE (42.7 vs 49.8). API prices rise Aug 16, off-peak V4-Pro at $0.66/M in, $1.98/M out.
SpaceXAI has released Grok 4.6, built on the Grok 4.5 foundation with a focus on long-running agents and interactive work. It scores 61 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Sol Max and up from 56 for Grok 4.5 High, and 1753 on GDPVal-AA v2 vs 1728 for Sol Max.
Pricing starts at $2/M input and $6/M output tokens, with a fast variant at double. Live now in Cursor, Grok Build, the API, OpenRouter, Vercel and Cloudflare, with 500K context.
Google's Gemini app has passed 1 billion monthly active users, Sundar Pichai said on X, calling it Google's fastest growing product ever and its 14th to reach that mark, alongside Search, Gmail, Android, Maps, Chrome, Play and YouTube. The app was at 400M in May 2025, 900M in July 2026 and 950M in late July.
Google says 63% of users reach Gemini by voice, 1 in 5 Gemini Live interactions use camera or screen sharing, and it generates 150M+ images a day.
Meta Superintelligence Labs open-sourced Muse Glimmer, a 30B multimodal model distilled from its closed Muse Spark, now on Hugging Face under Apache 2.0. It targets always-on local agents — tool calls, coding, file and screenshot handling — and runs offline on one consumer GPU, under 20GB quantized within a 24GB envelope.
Benchmarks: 75.5 MCP Atlas, 74.6 DeepSearch QA, 76.0 SWE-Bench Verified, 51.2 SWE-Bench Pro. 131K+ context, 100+ languages.
