1. Anthropic Ships Claude Opus 5.5 at ~40% Lower Cost
Anthropic released Claude Opus 5.5 on September 22, priced at $4/M input and $20/M output tokens, roughly 40% cheaper than Opus 5. It scores 66.4% on Terminal-Bench 4.0 (vs 52.3% for Opus 5) and 81.8% on OSWorld 2.0 computer use, with output generation about 30% faster.
2. OpenAI Launches GPT-6 Sol and Luna at Half the Price
OpenAI launched GPT-6 Sol and Luna, with Sol priced at $2/$10 per million tokens versus $4/$20 for the GPT-5.6 tier. Sol targets complex coding with what OpenAI calls Astra-level reliability, while Luna is aimed at high-volume clerical work. OpenAI DevDay follows on September 29.
3. xAI Releases Grok 4.7
xAI launched Grok 4.7 on September 21 at $2 per million input tokens and $6 per million output. xAI reports 71.0% on DeepSWE v1.1, 46.3% on CursorBench 4.0, 38.0% on Terminal-Bench 4.0 and 56.7% on HealthBench Professional.
4. Xiaomi Open-Sources MiMo-V2.6 Under MIT License
Xiaomi released the MiMo-V2.6 series on Hugging Face: an omnimodal Pro model plus a 309B-parameter Flash MoE with 15B active parameters and 256K context. Xiaomi says the Pro model scores 46 on the Artificial Analysis Intelligence Index, the highest for any open-weight model.
5. UN Security Council Hosts US and Chinese AI Labs
France, holding the Security Council presidency, convened a session on AI and international security today, September 23, with OpenAI's Sam Altman, Anthropic, DeepSeek and Moonshot expected. It is the first time the Council has brought frontier US and Chinese developers together. Separately, the UN's 40-expert scientific panel on AI, co-chaired by Yoshua Bengio, published its first brief urging governments to rein in AI agents.
6. Cisco Talos Reports First Fully Autonomous AI Malware C2
Cisco Talos disclosed CLOSEDQUORUM, which it describes as the first reported fully autonomous multi-model AI command-and-control implant running with no human operator. Talos also released CAIRN, a toolkit for tracking AI-integrated malware, noting such malware went from optional LLM helpers to autonomous C2 in a single year.
7. AI Agents Help Solve the Last Open M23 Inverse Galois Case
A six-mathematician team including MIT's Bjorn Poonen and Colorado State's Rachel Pries used multiple AI agents to search M23's symmetry combinations. The work produced an explicit degree-23 polynomial whose Galois group is M23, closing the final open sporadic case of a decades-old problem.
8. Snorkel AI Raises $350M at $3.5B Valuation
Snorkel AI raised $350 million led by Insight Partners and S32, valuing it at $3.5 billion, nearly triple its May 2025 mark. CEO Alex Ratner said annualized revenue passed $350M, up from about $20M a year earlier, on demand for complex AI training data.
9. British Columbia Sues OpenAI Over Tumbler Ridge Shooting
BC Attorney General Niki Sharma filed suit in San Francisco federal court on September 21 against OpenAI and Sam Altman over the February 10 Tumbler Ridge school shooting that killed eight. The province alleges OpenAI's safety team flagged the shooter's ChatGPT sessions in June 2025 and recommended contacting police, but leadership rejected it.
10. Pew: US Views of Data Centers Turn Sharply Negative
A Pew survey of 10,548 US adults found 54% now say data centers harm the environment, up 15 points since January, and 50% say they raise home energy costs, up 12 points. Awareness of data centers rose from 75% to 88%.
// KEY TAKEAWAYS
A price war broke out at the frontier this week: Anthropic, OpenAI and xAI all shipped new models at sharply lower prices, while Xiaomi pushed open weights closer to the top tier. Agents are now the center of the safety debate, from the UN's first brief and Security Council session to the first autonomous AI malware C2. Meanwhile, money keeps flowing into data and applied AI, even as public sentiment on data centers sours.