AI Trends: 23 September 2026

Anthropic and OpenAI ship cheaper frontier models 90 minutes apart

Tuesday brought a head-to-head release day. Anthropic launched Claude Opus 5.5 at $4 per million input tokens and $20 per million output, roughly 40% cheaper than Opus 5 on typical workloads. The company says it matches Claude Fable 5.1 on most tasks while generating output 30% faster, scoring 66.4% on Terminal-Bench 4.0 (up from 52.3%) and 81.8% on OSWorld 2.0. Anthropic also says the model is 85% less likely to attempt boundary circumvention than Opus 5. It is the first release since Dario Amodei's "pace the frontier" essay.

About 90 minutes later, OpenAI released GPT-6 Sol and GPT-6 Luna at half the price of the GPT-5.6 tier ($2/$10 per million tokens for Sol). Sol targets complex coding, while Luna is built for high-volume clerical work. OpenAI says Sol makes about half as many factual mistakes as its predecessor. The launches cap a month that CNBC has described as producing "model fatigue" among enterprise buyers. Separately, xAI shipped Grok 4.7 on 21 September at $2/$6 per million tokens.

UN Security Council hosts US and Chinese AI labs together

France, which holds the Security Council presidency this month, convened a session on AI and international security for today, 23 September. OpenAI's Sam Altman and senior Anthropic representatives are expected alongside DeepSeek and Moonshot. It marks the first time the Council has directly hosted frontier US and Chinese developers on shared safety concerns. The session follows the UN's new scientific panel on AI publishing its first brief on 21 September, urging governments to put safeguards around AI agents before their risks are fully understood.

OpenAI opens model training to outside safety testers

OpenAI said it will let third-party groups run technical safety assessments during training, evaluation and deployment of new models, earlier than the post-development testing that has been standard. The move echoes Anthropic's decision last week to embed Accenture as an independent evaluator.

Agent security is becoming a real-world problem

Cisco Talos disclosed CLOSEDQUORUM, which it calls the first reported fully autonomous, multi-model AI command-and-control malware implant, and released a toolkit for hunting AI-integrated malware. Meanwhile, six major banks including Bank of America and Capital One published a joint report warning that AI shopping agents could raise scams, fraud and disputes, proposing five principles for "agentic commerce." A researcher also disclosed a flaw in Meta's new Muse for Mac agent that let any local app steal the user's account token.

Money and infrastructure

AMD closed above a $1 trillion market cap for the first time, riding data-centre revenue up 107% year over year. Alibaba announced new cloud regions in Turkey, Finland and the Netherlands as part of a $53 billion AI and cloud plan. AI training-data suppliers are booming: Snorkel AI raised $350 million at a $3.5 billion valuation and Micro1 raised over $100 million at $4 billion. On the cautionary side, Reserve Bank of Australia Governor Michele Bullock told lawmakers AI investment "might be" a bubble.

Also this week

Anthropic and OpenEvidence are bringing a free clinical AI search tool to physicians in about 100 low- and middle-income countries. Xiaomi released its open-weight MiMo-V2.6 models under an MIT license, claiming parity with leading closed models on agent benchmarks. Alphabet's Intrinsic open-sourced its full robotics software stack. And AI helped mathematicians work through roughly 25,000 cases of an inverse Galois problem.

Compiled from public reporting for the 48 hours ending 23 September 2026.