AI News for September 23, 2026

OpenAI released GPT-6 Sol and Luna, two faster, cheaper models built on GPT-6 Astra's advances, with API prices cut 50% versus GPT-5.6 promotional rates.

Anthropic's Opus 5.5 and OpenAI's GPT-6 Sol and Luna launched 90 minutes apart with lower prices, while Meta's Muse agent faced safety and disclosure controversies.

Five competing pharmaceutical companies jointly trained a federated AI model on over 20,000 proprietary drug-discovery structures without sharing data, achieving state-of-the-art protein-ligand co-folding accuracy and launching a follow-on initiative for binding-affinity prediction.

Anthropic and OpenAI released Opus 5.5 and GPT-6 Sol and Luna within days of each other, continuing an accelerating release cadence despite employee and CEO calls to pace frontier AI development.

AI Weekly's September 23 roundup curates top stories including OpenAI's Instagram ad surge, the RoboFollow VLA benchmark exposing instruction-following gaps, a 1,024-agent LLM system, YouTube's AI Custom Feeds, Amazon opening Seller Central to Claude, and China overtaking the US in elite AI talent.

Runway launched DIFFUSE, a hybrid marketplace platform connecting AI-native creatives with brands and studios, citing internal research that 17% of creative job postings now require AI skills.

Business Insider reports that OpenAI has sharply increased its sponsored influencer posts for ChatGPT, partnering with parenting, fitness, and business creators even as some influencers face audience backlash over AI brand deals.

Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Sol/Luna launched hours apart with sharp price cuts, making cost rather than benchmarks the headline story for teams paying monthly AI bills.

Ema raised a $77m Series B led by Creaegis to scale its agentic AI Employees platform, now deployed at Fortune 500 scale with 50X revenue growth over two years.

Alibaba's official Apsara Conference press release confirms Qwen 4 is currently in training with no release date, pricing, or API name announced, while teasing a Qwen 4.5 and Qwen 5 roadmap scaling to 5-10 trillion parameters.

Legal AI startup Chamelio raised a $26 million Series A led by Entrée Capital just five months after its seed round, as its annual recurring revenue quadrupled over that stretch.

A reference page from endoflife.date listing Anthropic Claude model versions with their release, deprecation, and end-of-life dates.

The Indian Express AI section index page aggregates recent AI headlines, including OpenAI's GPT-6 Sol and Luna models, Meta's Muse AI agent, and a California AI safety executive order.

OpenAI's Codex CLI released version 0.156.0, adding a fullscreen TUI, always-on voice conversations, a /usage analytics dashboard, default worktree sessions, new terminal themes, Mermaid and equation rendering, and sandbox security fixes.

A reference calendar listing AI model releases by date across OpenAI, Anthropic, Google DeepMind, DeepSeek, Qwen, and other major labs, last updated September 22, 2026.

A GitHub repository hosts an opinionated, citation-backed handbook for engineers building software with AI coding agents, maintained by a pipeline of AI agents that discover, review, and synthesize updates.

LM Market Cap's LLM Updates page is a live reference tracker listing 462 models from 61 providers, including recent releases like GPT-6 Luna and Claude Opus 5.5.

U.S. TRANSCOM is deploying randomised AI routing, autonomous network healing, IoT sensors, digital twins, and blockchain to protect military logistics from adversary predictive modelling and contested disruption.

Nebius raised pay-as-you-go prices for Nvidia GPU instances, AMD CPU capacity, and memory, with memory jumping about 41%, as AI compute demand continues to outpace supply.

The Gujarat government has launched a state-wide GPU-as-a-Service framework managed by Gujarat Informatics Limited, offering subsidized centralised GPU computing to government departments, educational institutions, and selected AI startups.

McKinsey partners explain that even as per-token AI costs plummet, enterprise AI bills keep rising because agentic workloads consume far more compute, with the same task varying up to 30x in cost, and they recommend tracking spend by task-level economics rather than token price.

Perplexity built its CobbleDB key-value store in two months with two engineers and AI coding agents, cutting batch read latency by 82% and storage costs by over 20%.

NVIDIA's AI Day Singapore showcased fifteen Southeast Asian partner organizations building on its Nemotron and CUDA stack, with Singapore dominating the list and key ASEAN markets Indonesia and the Philippines notably absent.

A daily tech news roundup covers Microsoft's disruption of the AI-driven EvilTokens phishing network, Chrome 154's patching of 108 flaws, an actively exploited Arista VeloCloud zero-day, and a federal investigation into Comma.ai's hands-off driving technology.

In a podcast interview, Gaurav Gupta of Decimal Point Analytics explains how agentic AI differs from generative AI and what foundations—enterprise knowledge bases, data pipelines, and defined workflows—organizations need before deploying it in critical processes.

A GitHub issue on the retrievr-mcp repository documents that Perplexity's Sonar /chat/completions endpoint is scheduled to shut down on 2026-09-27, breaking the repository's perplexity plugin and Atlas's retrievr configuration.

R-Capital Research argues hyperscaler capex will exceed Morgan Stanley's $1.4 trillion 2027 estimate and examines ARM CPUs, AMD's Helios, and Nvidia's Rubin in the AI data center buildout.

The Guardian's Technology + AI topic page aggregates recent articles on AI, spanning UK regulation, datacentre investment, AI doomsday debates, labor impacts, and big tech's promises.

A GitHub repository documents an automated GitHub Actions workflow that aggregates AI ecosystem signals from many data sources into bilingual daily digest reports.

The agents-radar repository documents a GitHub Actions workflow that aggregates AI ecosystem signals from ten sources and publishes bilingual daily digests as Issues, Markdown, Pages, and RSS, along with an MCP server and web UI.

The extracted page is the LlamaIndex GitHub repository README, describing the open-source framework's document parsing focus, integrations, and LlamaParse platform rather than covering the cited bug issue.

A proposed class-action lawsuit accuses Anthropic, OpenAI, SpaceXAI, and Google of violating antitrust law by coordinating to slow AI development, allegedly harming paid subscribers.

Sen. Bernie Sanders and Rep. Greg Casar unveiled legislation to ban artificial superintelligence and create a federal Department of Artificial Intelligence, a proposal with long odds in the Republican-controlled Congress but a marker of progressive AI regulation priorities.

California Governor Gavin Newsom announced a panel of leading AI experts to advise on his recent executive order, which seeks to accelerate independent oversight of AI companies and advance the creation of an emergency "kill switch" for rogue frontier models.

A publisher coalition suing OpenAI and Microsoft reportedly grew past 550 outlets as consolidated copyright cases advance toward fair-use summary judgment, though no liability has been decided.

A US-China summit produced a bilateral framework establishing 'digital red lines' for AI safety, including a human-in-the-loop requirement for nuclear command-and-control, alongside talks on a new strategic stability model for nuclear arms control.

President Trump used his UN General Assembly keynote to call for renaming AI 'Super Intelligence' and rejecting safety-driven slowdowns, as OpenAI, Anthropic, and Hugging Face CEOs prepared to address the UN Security Council on AI regulation ahead of his state visit with China's Xi Jinping.

This legal analysis explains how Indian courts and existing statutes—privacy, passing off, trademark, and copyright—protect celebrities from unauthorized AI-generated use of their face, voice, and likeness, while balancing free-speech and parody exceptions.

Unions, particularly in manufacturing and the UAW, are negotiating job security provisions, retraining programs, and transparency rules to protect workers from AI-driven automation.

Rep. Riley Moore, a former welder, is telling labor groups like the Teamsters that AI could make blue-collar jobs safer and more productive rather than replace workers, while opposing fully automated trucking.

A practical guide for teachers on using AI tools to save time on lesson planning, grading, differentiation, and classroom tasks, with privacy and safety guidance for the USA, India, and Europe.

A vendor guide outlining a full Google Workspace to Microsoft 365 migration and cutover plan, covering assessment, inventory, tenant and identity setup, and data mapping.

This GitHub issue asks the maintainer of the northbound-job-radar personal job-search tool to decide the fate of an orphaned export module.

AI-generated political ads placing candidates in fabricated situations are proliferating ahead of the November U.S. midterms, a largely unregulated practice adopted despite public distrust of the technology.

AI-generated political ads depicting candidates in fabricated scenarios have proliferated ahead of the US midterms, raising concerns about eroding voter trust amid a lack of federal disclosure rules.

AI-generated political ads are proliferating ahead of the U.S. midterms, raising concerns about voter trust and the authenticity of campaign content.

A Cornell faculty committee released a 238-page report with nearly 70 recommendations for how the university should adapt to technological change, federal-government tensions, and eroding public trust in higher education.

The Guardian's AI topic page collects a wide range of recent stories and commentary, including political debates over AI and disinformation, economic bubble concerns, datacentre controversies, and ethical questions about the technology.

YouTuber MrBeast spent nearly $10 million building a self-sustaining village in Ghana to combat child labor, drawing both praise and familiar skepticism about his philanthropic motives.

Multibagg AI is an AI-powered stock research and analysis platform offering data and analytics for Indian companies listed on the NSE and BSE.

A feature profile recounts how Tubi, dismissed by Hollywood and investors, built a billion-dollar streaming business by licensing cheap B-movies and canceled TV shows for a free, ad-supported catalog.

An investor-oriented analysis argues that networking, optics, and switching stocks like Arista, Cisco, and Ciena are underappreciated beneficiaries of AI data center buildouts, since data movement rather than raw GPU compute is the real bottleneck.

Simply Wall St analyzes how securities class action lawsuits alleging Wix.com misrepresented the competitiveness, costs, and benefits of its AI offerings affect the company's investment narrative and shareholders.

A FinancialContent aggregation page whose headline promises legal AI billing analysis actually delivers only a list of unrelated stock and earnings items, with no substantive article on legal AI costs.

The DICOMPUTE/Open repository README documents its standardized eight-section issue filing format, but the selected text does not actually contain the referenced bug report about fabricated model timestamps.

A morning news digest summarizes top stories including ACA coverage cuts, Texas data-center permit halts, Hurricane Polo, a $1 billion University of Oregon donation, and Trump's UN address.

This is a TechCrunch category landing page describing its general coverage of artificial intelligence and machine learning, not a specific news article.

A roundup of recent cases shows how AI agents and deepfakes are scaling cyberattacks, from an Anthropic-reported Chinese espionage campaign against 30 organizations to Gemini-abusing Android malware and a $25 million deepfake video fraud.

A Trellix report describes how AI-driven attackers are industrializing deception through autonomous espionage agents, mobile malware, deepfake payment fraud, and checkout skimming that exploit misplaced trust in normal-looking interactions.

A GitHub repository maintains a sourced, deduplicated timeline of 2026 agentic AI safety events, updated by a daily Pydantic AI agent cron job using Claude Opus and web search.

A narrative methodological review proposes a framework combining fMRI, AI, and statistical modeling for phenotyping PPPD and tracking residual dizziness after BPPV, while stressing that current findings remain hypothesis-generating.