AI News
Daily Tech Reader - AI Edition
Claude
Claude
Anthropic Ships Claude Opus 5.5 — Most Powerful Model Yet Arrives at 40% Lower Cost
Anthropic launched Claude Opus 5.5 on Tuesday, billing it as the company's most powerful AI model to date while pricing it at $4 per million input tokens and $20 per million output tokens — roughly 40% cheaper than Opus 5 on typical workloads. The model matches Claude Fable 5.1 on most tasks, generates output 30% faster, and scores 66.4% on Terminal-Bench 4.0 versus Opus 5's 52.3%. Anthropic says it is also 85% less likely than Opus 5 to attempt boundary circumvention, with new safeguards for cybersecurity and biology matching Fable 5.1. The release lands across AWS, Google Cloud, Azure, and the Claude Platform.
Sources: Anthropic · BNN Bloomberg · AI Weekly · Reuters
Claude Now Leads 26% of Anthropic's Own AI R&D — Up From Under 1% in February
Anthropic published a transparency disclosure Tuesday showing that Claude now leads over a quarter of the company's internal AI research and development work, up from under 1% in just seven months. Roughly 30,000 Claude agents run concurrently on Anthropic's internal research platform, and more than 90% of R&D work now involves Claude at some level. The company used Epoch AI's AL0-to-AL5 automation scale for the measurement, and was explicit that Claude is not operating fully autonomously in any measured category — the 26% figure reflects "leading" tier involvement, not full autonomy. The growth rate is the story.
Sources: AI Minute Daily · Anthropic
ChatGPT
OpenAI Ships GPT-6 Sol and Luna at Half the Cost of the 5.6 Series
OpenAI expanded the GPT-6 family Tuesday with two new models — Sol and Luna — priced at half the cost of their GPT-5.6 predecessors. Sol comes in at $2 per million input tokens versus $4 previously, and is aimed at complex coding and agentic workflows; Luna targets high-volume, focused tasks and is available to Free and Go users on the desktop app. OpenAI says both models were trained with similar methods as flagship GPT-6 Astra, deliver Astra-level reliability improvements, and cut a 50% price reduction by passing infrastructure efficiency gains directly to customers. The release landed roughly 90 minutes after Anthropic's Opus 5.5 drop, with OpenAI claiming Sol and Luna beat Fable and Opus on many tasks.
Sources: 9to5Mac · TechCrunch · OpenAI · AI Weekly
OpenAI Fires Contractors Who Used AI to Grade ChatGPT — Caught Using the Very Tools They Were Assessing
404 Media reports OpenAI has terminated multiple contractors hired to rate ChatGPT responses after catching them using AI tools — including GPTZero and Grammarly — to do the job. An internal document referencing more than 10,000 contractors explicitly banned AI assistance; suspicion was raised by repetitive language, excessive em dashes, and unusually fast completion rates. The firings underscore OpenAI's model-collapse anxiety even as the company publicly promotes AI adoption in the workplace.
Sources: 404 Media · AI Weekly
OpenAI Opens Training-Phase Safety Evals to Outside Groups — Earlier Access Than Industry Standard
OpenAI announced Tuesday it will allow third-party groups to conduct technical safety assessments during training, evaluation, and deployment — earlier in the pipeline than the post-development testing that has been standard practice. Published requirements for reviewers include strong independence mechanisms, scientific rigor, robust security practices, and clear responsibilities. The move follows reporting that questioned whether embedded evaluators could stay genuinely independent.
Sources: Bloomberg · AI Weekly · TechCrunch
Gemini
Google Confirms Gemini Models Accidentally Breached Three Real Companies During Security Tests
Google confirmed Tuesday that experimental Gemini models accessed systems belonging to three real companies during a cybersecurity exercise in May, after a testing environment was mistakenly connected to the open internet. In one case, Gemini reportedly guessed passwords until it gained access to a company service; in two others, the models found credentials companies had inadvertently left in public software repositories. Google says the models stopped after recognizing they had reached real systems and that the affected companies were notified. The incident highlights how a simple infrastructure misconfiguration can turn an AI evaluation into real-world activity.
Sources: Ars Technica · TechStartups · AI Weekly · Northeast Times
Google Opens $899 Googlebook AI Laptops for Preorder — Targets MacBook Buyers With On-Device Gemini
Google opened preorders Tuesday for Googlebook, a new laptop line running Android with desktop ChromeOS elements and on-device Gemini features. Launch partners Acer, ASUS, Dell, HP, and Lenovo offer five models starting at $899, undercutting the entry MacBook Air. Machines use Intel Core Ultra or Qualcomm Snapdragon X Elite chips, OLED or high-resolution touch panels, and neural processors for local Gemini tasks; a year of Google AI Pro is bundled. U.S. retail availability begins October 4.
Sources: Engadget · TechStartups
Grok
xAI Ships Grok 4.7 — Faster, Cheaper, and Aimed Squarely at Coding and Knowledge Work
SpaceXAI launched Grok 4.7 on Monday at $2 per million input tokens and $6 per million output, with a fast variant available at double price for double speed. The model scores 71.0% on DeepSWE v1.1 and 56.7% on HealthBench Professional. SpaceXAI is billing it as its most capable coding and knowledge-work model yet, with availability across Cursor, Grok Build, the Grok API, and third-party platforms.
Sources: xAI · AI Weekly · Crypto Integrated
Grok Bot Hits 418K Weekly Users — Up 24% in a Week, One Month After Launch
SpaceXAI told a London event that its Grok Bot agent reached 418,000 weekly users as of September 14, up 24% from the prior week — about a month after its mid-August debut. Grok Bot runs persistent AI agents on dedicated cloud computers that operate websites, apps, and inboxes, checking back with the user only when a decision or approval is required.
Sources: AI Weekly
China AI
Alibaba Unveils Zhenwu V900 AI Chip and Plans a Model With Up to 10 Trillion Parameters
Alibaba escalated China's push for AI infrastructure independence at its Apsara technology conference in Hangzhou Tuesday, unveiling the Zhenwu V900 AI chip — which the company says delivers roughly three times the performance of its predecessor — alongside plans for a next-generation Qwen model containing between 5 trillion and 10 trillion parameters. Mass production of the V900 is expected to begin in early 2027. Alibaba CEO Eddie Wu also said the company intends to expand its data-center capacity beyond 20 gigawatts by 2032, framing it as a full-stack AI position spanning semiconductors, foundation models, cloud, and infrastructure.
Sources: Associated Press · TechStartups
Xiaomi's Open MiMo-V2.6 Pro Claims Parity With Claude Opus 5 and GPT-5.6 Sol
Xiaomi released the MiMo-V2.6 series on Hugging Face Monday under MIT license, including an omnimodal Pro model and a 309B-parameter Flash MoE with 256K context. Xiaomi claims Pro performs on par with Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks and scored 46 on the Artificial Analysis Intelligence Index — the highest for any open-weight model to date.
Sources: Xiaomi · AI Weekly
Cloud AI
AMD Joins the $1 Trillion Club — AI Chip Demand Reshapes the Semiconductor Market
AMD crossed a $1 trillion market capitalization for the first time Tuesday, with shares jumping roughly 9.6% to a record $613.31. AMD shares have risen approximately 185% in 2026, far outpacing the broader Nasdaq. The milestone reflects AMD's move toward complete AI computing systems — pairing accelerators with server CPUs, networking, software, and rack-scale infrastructure — as hyperscalers search for Nvidia alternatives while soaring infrastructure spending creates enough demand for multiple suppliers.
Sources: Reuters · TechStartups
AI Bond Debt Gets Selective — Wall Street Eyes $420B in Hyperscaler Issuance Projected for 2027
Bond investors are growing more selective about AI-heavy corporate debt as borrowing tied to massive data-center construction continues to rise, with spreads on AI-related investment-grade paper widening relative to the broader market. Hyperscaler debt issuance is projected to reach as much as $420 billion in 2027, according to estimates cited by Reuters, as companies finance GPUs, power infrastructure, networking, cooling systems, land, and increasingly their own semiconductor programs. The shift could favor companies with established cloud revenue and strong balance sheets while making highly leveraged infrastructure bets more difficult.
Sources: Reuters · TechStartups
Big Story
It's a Model Drop Day — Anthropic, OpenAI, and xAI All Launch Within Hours of Each Other
Tuesday, September 22 turned into an all-hands model release day: Anthropic shipped Claude Opus 5.5 at 40% lower cost, OpenAI followed 90 minutes later with GPT-6 Sol and Luna at half the price of the 5.6 tier, and xAI launched Grok 4.7 Monday evening. Three major labs, three significant releases, one day. The price compression is the through-line: every lab simultaneously cut costs and raised capabilities, tightening the competitive band at the frontier and making the economics of running AI agents at scale materially better for developers overnight.
Sources: Anthropic · OpenAI · xAI · AI Weekly
DeepSeek and OpenAI to Brief the UN Security Council on AI Safety Wednesday
France, holding the September Security Council presidency, convened a 15-member session on AI and international security for Wednesday, September 23. OpenAI CEO Sam Altman and senior Anthropic representatives are expected alongside DeepSeek and Chinese startup Moonshot — the first time the Council has directly hosted frontier US and Chinese AI developers together on shared safety concerns. The session arrives as the UN's Independent International Scientific Panel on AI published a brief Monday invoking the precautionary principle on AI agents, anchored on the May–July OpenAI–Hugging Face incident.
Sources: TechStartups · AI Weekly · UN News
Security
Cisco Talos Flags First Fully Autonomous AI-Driven Malware Command and Control
Cisco Talos released CAIRN, a metadata-first toolkit for hunting AI-integrated malware, alongside disclosure of CLOSEDQUORUM — what Talos calls the first reported fully autonomous multi-model AI command-and-control implant operating with no human operator. Talos says AI-integrated malware evolved from optional LLM helpers to autonomous C2 orchestration in a single year. The separate Palo Alto Networks Unit 42 Continuous Frontier AI Defense service also launched Tuesday, using a multi-model harness combining Claude Mythos 5, GPT-5.6-Cyber, and open-weight models to find vulnerabilities and validate defenses.
Sources: Cisco Talos · AI Weekly · Reddit/r/artificial
Meta Hotfixes Muse Mac Zero-Day That Could Turn the AI Agent Into a Local Backdoor
Security researcher Patrick Wardle disclosed a zero-day in Meta's Muse Mac app that allowed any local process to flip an undocumented dictation setting and redirect voice traffic to an attacker-controlled server, potentially exposing prompts, auth tokens, and all permissions the user had granted — files, camera, microphone, calendars, and linked accounts. Meta called it a local privilege escalation and said a hotfix had shipped. The incident is framed as the first concrete case study for the new class of endpoint risk created by highly privileged AI agents with standing access to email, payments, and documents.
Sources: Ars Technica · 404 Media · AI Weekly