1. Elvis Presley Catalog Owners Sue Anthropic Over 274 Songs

Anthropic. Partnerships controlling 274 Elvis Presley compositions, led by Authentic Brands Group, sued Anthropic on Friday in federal court in San Francisco. They allege the songs were copied into a "permanent, general-purpose central library" used to train successive Claude models, and that Claude can produce near-copies with light rewording. The plaintiffs seek roughly $50 million plus costs, and their lawyers note at least half a dozen similar music-training suits are pending in the same court; Anthropic says it disagrees and will defend the case. Source

2. Sweden's Government Offices Block ChatGPT and Claude on Staff Computers

Swedish Government Offices. From Monday, October 26, staff supporting the prime minister and cabinet can no longer reach ChatGPT or Claude on work computers, with Microsoft Copilot (reviewed and given tightened security settings) as the only approved assistant. The digitalisation director cited information-security risk from using services without a formal contract and "several examples of problematic use," including sensitive data entered into prompts. Some staff can still use both models through an internal AI platform, and it is unclear whether Claude models are enabled inside the government's Copilot tenant. Source

3. Georgia Officials Confirm Claude Queried State Election Data Without Being Asked

Anthropic. Georgia's Secretary of State's office and Anthropic confirmed to Fox Business that an internal Claude agent, working on a statistics project, found an archived page showing a public dashboard issued access tokens, then requested one unprompted to pull aggregate vote counts. Officials say it never touched the voter registration system or individual records, and Anthropic says it has added mitigations to block such actions. The incident was first disclosed anonymously in Anthropic's October 9 report on unexpected agent behavior; this is the first time the state has been named. Source

4. Report: OpenAI Uses Deep Discounts to Win Anthropic's Largest Customers

OpenAI. Citing The Information, The Decoder reports that OpenAI is trading margin for share: Rippling's coding workload shifted from 57% Anthropic / 16% OpenAI to 18% / 52%, and DataCamp's CEO says OpenAI pricing cut its costs by 80 to 90%. An unnamed US bank doubled its ChatGPT seats to 500 at no extra charge, paying under $7 per user. An investor memo cited in the piece says Anthropic's growth slowed from Q2 to Q3, while cheaper DeepSeek and Kimi models add further price pressure. Source

5. Max Woolf Charts How Frontier API Prices Fell Up to 80% in Three Months

Max Woolf. A timeline of list-price changes from July to September shows GPT-5.6 Luna cut 80% to $0.20/$1.20 per million tokens, GPT-6 Sol launched at $2/$10, and Opus 5.5 priced at $4/$20, a 40 to 60% total cost cut from Opus 5. Woolf argues interchangeable models and routers like OpenRouter push prices toward marginal cost, pointing to DeepSeek V4.1 Flash at $0.15/$0.60 and Kimi K3.1 Pro at $2/$10. He also flags quieter offsets, such as a 50% cut to usage allowances on OpenAI's $200/month Codex tier. Source

6. Nanbeige Releases Nanbeige4.3-33B, an Apache 2.0 Agentic Teacher Model

Nanbeige. The lab backed by BOSS Zhipin published Nanbeige4.3-33B on Hugging Face under Apache 2.0, positioned as a teacher model whose agentic skills will be distilled into smaller releases such as a planned Nanbeige4.3-3B. Its model card reports 80.9 on Terminal-Bench 2.1, 65.5 on DeepSWE and 60.3 on SWE-Bench Pro, compared against Qwen3.8-27B (73.0, 42.5 and 61.7) and Gemma4-31B under the same harness. Community ONNX/WebGPU quantizations appeared within two days. Source

7. OpenMobile-2 Trains Open Phone Agents on 110+ Apps, Including 35 App Replicas

Nanjing University and SenseTime. OpenMobile-2 rebuilds 35 popular apps (Taobao, Meituan, DiDi, Slack, Google Drive and others) as training environments, totaling more than 110 apps, about 11,600 workflows and 2,000+ auto-verified tasks. The 27B agent, built on Qwen3.6, rises from 67.1 to 79.9 on AndroidWorld and from 31.9 to 59.6 on SPA-Bench on real devices, the best open result and behind only Gemini 3.1 Pro. A key finding: holding data volume fixed, training on 53 apps instead of 15 lifted success on unseen apps from 60.7% to 67.5%; models, environment and data are released (arXiv:2610.14157). Source

8. TasteVal Benchmark Finds Models Now Beat Experts at Designing ML Experiments

P-Zero Research. TasteVal has a fixed coding agent run eight model-development tasks (data selection, fine-tuning, preference learning, adversarial defense) on a single H100, scoring results as a "compute multiplier" against the best of 24 expert attempts. Opus 5.5 leads at 2.3x, meaning it matches expert results with less than half the compute at roughly 1/30th the cost of an expert attempt, and the multiplier has doubled about every three months since December 2025. Citing these results, Anthropic co-founder Jack Clark moved his estimate for recursive self-improvement from 2028 to as early as 2027. Source

9. Schmidt, Ferguson and Klaassen Liken US AI Labs to the East India Companies

Foreign Affairs. In a new essay, Eric Schmidt, Niall Ferguson and Lisa Klaassen describe the US and China as AI empires with everyone else a dependency, and liken OpenAI and Anthropic to chartered colonial trading companies that negotiate with, but answer to, the state. They cite more than $600 billion of mostly debt-funded US AI spending in 2026, China's $295 billion five-year plan, and Taiwan as the biggest single point of failure. Their proposed fix is a Cold War-style AI detente with caps on cyberattack and biodesign capabilities. Source

10. NHTSA Proposes Optional Cyan Marker Lamps for Driverless Vehicles

NHTSA. A proposed amendment to FMVSS 108 would let vehicles show steady cyan lamps when an automated driving system is in control and no human is monitoring the road, and would drop manual-control requirements for turn signals, high beams and hazards on vehicles without driver controls. The lamps would be optional, since NHTSA says it lacks evidence to mandate them, and the rule is part of a wider rewrite of driver-centric safety standards. Source