Skip to content
Daily BriefNewsDaily Brief

9 AI stories from September 23, 2026: Claude Opus 5.5 at Fable-level performance and 40% lower cost, GPT-6 Sol and Luna price cuts, UN Security Council AI briefing with frontier lab CEOs, and 6 more model and industry updates

· by Pondero Newsdesk · 9 stories

AI news daily brief: 2026-09-23

Nine stories today: three model releases, two regulatory developments, two product and open-source launches, one cybersecurity takedown, and one hardware infrastructure deployment.

Anthropic Releases Claude Opus 5.5 at Fable-Level Performance and 40% Lower Cost

Anthropic released Claude Opus 5.5 on September 22, 2026, the first model in its new 5.5 family. Per Anthropic's announcement, Opus 5.5 performs at the level of Claude Fable 5.1 on most tasks while costing 40% less on typical workloads. Pricing is $4 per million input tokens and $20 per million output tokens, with cache reads dropping to $0.20 per million tokens, a 60% reduction versus Opus 5. On benchmarks, Opus 5.5 surpassed Fable 5.1 on agentic coding, knowledge work, computer use, visual chart recognition, and multidisciplinary reasoning. Anthropic says the model completed a 680,000-line code migration in under a day. It achieves the company's strongest automated behavioral audit scores to date and shows greater resistance to prompt injection than Opus 5. Usage limits are increasing for Pro, Max, Team, and Enterprise plans. Sonnet 5.5 and Haiku 5.5 are expected within weeks.

Full story: Anthropic releases Claude Opus 5.5

OpenAI Launches GPT-6 Sol and Luna With Permanent 50% Price Cuts Across API Tiers

OpenAI introduced GPT-6 Sol and GPT-6 Luna on September 22, 2026, per the company's announcement, expanding the GPT-6 family below the flagship Astra model. GPT-6 Sol, at $2 per million input tokens and $10 per million output tokens, targets complex coding and agentic workflows. GPT-6 Luna, at $0.10 per million input tokens and $0.50 per million output tokens, targets high-volume tasks such as summarization and extraction. An OpenAI spokesperson confirmed to VentureBeat that these prices are permanent, not introductory rates. Sol is 50% cheaper in both directions than GPT-5.6 Sol; Luna is 50% cheaper on input and 58.3% cheaper on output than its predecessor. Both models are live in the OpenAI API now.

Full story: OpenAI launches GPT-6 Sol and Luna

UN Security Council Holds First Open Briefing With AI Lab CEOs and Chinese AI Companies on Global Safety Risks

France convened an open high-level briefing on AI and international security on September 23, 2026 during UNGA high-level week, per Security Council Report. Briefers included Yoshua Bengio, co-chair of the UN Independent International Scientific Panel on AI, OpenAI CEO Sam Altman attending in person, Anthropic CEO Dario Amodei joining remotely, and Hugging Face CEO Clement Delangue. Chinese AI companies DeepSeek and Moonshot were invited to make statements. The session is the second UNSC briefing on AI but the first to include frontier lab leadership and the first to bring Chinese and Western AI companies into the same forum at Security Council level. Whether a joint statement or resolution emerges is the immediate output to watch.

Full story: UN Security Council holds first AI briefing with lab CEOs and Chinese AI companies

Microsoft and Partners Take Down EvilTokens, AI-Powered Phishing-as-a-Service Platform Targeting Enterprise Accounts

Microsoft's Digital Crimes Unit, with authorization from the U.S. District Court for the Eastern District of Virginia, announced the takedown of EvilTokens, per The Hacker News. EvilTokens was an AI-powered device code phishing platform targeting enterprise email accounts. Partners in the operation included OpenAI, Cloudflare, Coinbase, Health-ISAC, SpyCloud, The Shadowserver Foundation, and TRM Labs. The platform used AI to personalize phishing lures, design financial fraud roadmaps, and exploit stolen device authorization codes. Two men were arrested by the Metropolitan Police Service on September 11, 2026 in connection with the platform. Whether Microsoft publishes a full Digital Crimes Unit report on the EvilTokens attack chain is the next signal to track.

Full story: Microsoft and partners take down EvilTokens AI phishing platform

Alphabet's Intrinsic Open-Sources Its Full Industrial Robotics Stack Under Apache 2.0 at ROSCon 2026

Intrinsic, Alphabet's industrial robotics subsidiary, released Intrinsic Core on September 22, 2026 at ROSCon 2026 in Toronto, per the company's announcement. The release puts the same production robotics code Intrinsic uses in real manufacturing deployments into the public domain under the Apache 2.0 license. Intrinsic Core is ROS 2-compatible and includes a hardware-agnostic real-time control framework, collision-free path generation, NVIDIA FoundationPose integration, an on-edge ML inference engine, and a developer SDK. Prerequisites are Ubuntu 24.04 or 26.04 LTS with ROS 2 Lyrical Luth. Production-grade control and perception code under a permissive license gives robotics developers a foundation without proprietary lock-in.

Full story: Intrinsic open-sources its industrial robotics stack at ROSCon 2026

Meta Connect 2026 Keynote: Prescription Ray-Ban AI Glasses and Muse Spark Language Expansion Announced

Meta Connect 2026 opened September 23 with Mark Zuckerberg hosting the evening keynote. Meta announced two new Ray-Ban Meta variants, per Yahoo Tech: a standard sunglasses model and a prescription version called Phoenix, priced in the $1,000 to $2,000 range. Phoenix is the first prescription AI glasses model in the Ray-Ban Meta line. Both run on Meta's Muse Spark multimodal AI, with live translation across 14 additional languages and voice-guided pedestrian navigation on screenless models. Muse Spark is also receiving an update improving voice conversation quality and adding a shopping assistant in the US and Canada. The prescription tier opens AI glasses to a substantially larger addressable user base; volume adoption at the current price point is the open question.

Microsoft Becomes First Named Customer for AMD Helios Rack-Scale AI, Deploying 72-GPU Systems on Azure

AMD announced September 22, 2026 that Microsoft is the first publicly named customer for its Helios rack-scale AI platform, per the AMD Newsroom. Microsoft is deploying the 72-GPU MI455X system on Azure for frontier-model inference. The MI455X carries 31.1 terabytes of aggregate HBM4 memory alongside 6th-Gen EPYC Venice CPUs, Pensando Elba networking, and ROCm software. AMD says the MI455X delivers 1.5x the HBM4 capacity of competing offerings at a comparable price tier. AMD will begin shipping Helios to other customers in Q4 2026. Whether Google and Amazon follow with Helios commitments, and how rack-level performance compares to Nvidia H200/B200 systems, are the next data points to watch.

OpenAI, Anthropic, and Google in Talks to Create Industry-Led AI Standards Body Modeled on FINRA

Representatives from Google, Anthropic, and OpenAI have been meeting regularly since July 2026 about forming a self-regulatory standards body for frontier AI, per France 24. The proposal originated with Google DeepMind CEO Demis Hassabis, who publicly proposed a "US-led standards body" modeled on FINRA on July 14, 2026. The talks cover third-party evaluations before release, formal safety checks, and unified risk protocols. No formal governance structure or public announcement has emerged. The talks carry direct weight today given the UN Security Council session, where all three companies are represented.

xAI Launches Grok 4.7 With 2.1 Trillion Parameters, Tops CursorBench and DeepSWE Coding Benchmarks

xAI released Grok 4.7 on September 21, 2026. The model carries 2.1 trillion parameters, 40% more than Grok 4.6, using a new base model, a longer RL run, and training weighted toward difficult multi-hour coding tasks. Grok 4.7 scored 46.3% on CursorBench 4.0 and 71.0% on DeepSWE v1.1 at high effort, per xAI's blog. Pricing is unchanged at $2 input and $6 output per million tokens, per VentureBeat. The model is live in Cursor, Grok Build, and the xAI API, with a GitHub Copilot rollout underway. Whether Claude Opus 5.5 displaces Grok 4.7 at the top of CursorBench as it is added to the benchmark is the key performance comparison ahead.

Sources