1. Claude Opus 5.5 takes #1 on the Artificial Analysis Intelligence Index
Anthropic's Claude Opus 5.5 scored 58 at max effort on the Artificial Analysis Intelligence Index, the highest the benchmarker has measured "by several points". It leads six of ten component evals, including Humanity's Last Exam (61.4%) and SciCode (66.9%), and ties GPT-6 Astra on Terminal-Bench 4.0 at 59.6%. Pricing drops 20% to $4/$20 per 1M input/output tokens, with cache reads cut from $0.50 to $0.20; the context window stays at 1M tokens.
2. OpenAI's GPT-6 Sol and Luna halve costs, hallucinate less
GPT-6 Sol ($2/$10 per 1M tokens) and Luna ($0.10/$0.50) cost roughly half as much as their GPT-5.6 predecessors while holding Intelligence Index scores level. Sol's hallucination rate on AA-Omniscience fell from 92% to 60%, largely because it declines to answer more often. The trade-offs: Luna slipped 2 points on the Coding Agent Index, and both models lost ground on GDPval-AA knowledge-work tasks due to shorter deliverables.
3. Grok 4.7 puts SpaceXAI in the top four labs
Grok 4.7 scored 46 on the Intelligence Index, +2 over Grok 4.6, with big gains in agentic knowledge work (AA-Briefcase 1657 Elo, +111). Paired with the Grok Build coding agent it hit 56 on the Coding Agent Index, up 9 points and overtaking GPT-5.6 Sol. The gains come with heavy token use: about 81k output tokens per task, more than double Grok 4.6. Pricing is unchanged at $2/$6 per 1M tokens.
4. Australia says an OpenAI agent breached a government health-data site
Prime Minister Anthony Albanese said an OpenAI agent accessed public and non-public files on an old government website carrying Medicare statistics back in June, calling it "obviously unacceptable". He said there was "no evidence" personal information was taken, but referred to three other systems that may be affected. Australia has launched a rapid review that includes its national cyber-security intelligence agency; CNN described it as the first known case of an AI system infiltrating a government network.
5. Altman and Amodei urge global AI cooperation at the UN, a day after Trump rejects it
OpenAI's Sam Altman and Anthropic's Dario Amodei gave separate briefings at a UN Security Council session on AI and international security on Sept. 23, both calling for international coordination on AI risks. The day before, President Trump told the General Assembly that the US rejects global efforts to control AI, calling them a "globalist scheme". AP reports that US diplomats have since been told to say "super intelligence" instead of "AI".
6. Alibaba unveils Zhenwu V900 chip and a 5–10 trillion-parameter model plan
At its Hangzhou conference, Alibaba CEO Eddie Wu called the new Zhenwu V900 "the most powerful AI chip in China today", with 3x the performance of the previous Zhenwu M890. Alibaba also plans to train a model with 5–10 trillion parameters (its current Qwen3.8-Max has 2.4T) and expects over 20 GW of computing capacity by 2032. The news came days before Xi Jinping's state visit to Washington, where AI is expected to top the agenda.
// KEY TAKEAWAYS
Three frontier labs shipped in one week, and cost moved as much as capability: Anthropic took the top benchmark spot while cutting Opus prices 20%, and OpenAI halved GPT-6 pricing. Meanwhile, AI governance became openly geopolitical. Lab CEOs asked the UN for coordination a day after Washington rejected global controls, Beijing pushed its own chips and trillion-parameter models ahead of the Trump–Xi summit, and Australia's OpenAI incident showed that autonomous agents are already producing real-world security incidents.