Policy · 22 September
Trump orders US documents to call AI “super intelligence”
Addressing the UN General Assembly, President Trump said all United States documents would henceforth use the term “super intelligence” rather than artificial intelligence, telling leaders the word “artificial” makes the technology “sound fake and it is not fake”. He rejected any international scheme to rein the sector in — “I’m not going to stifle growth of something that will be bigger than the industrial revolution” — and said his administration would watch the industry closely through the Department of Justice instead. Fact-checkers noted that superintelligence denotes AI surpassing human ability across virtually all tasks, a bar current systems are explicitly not near. The rebrand is symbolic; the refusal to accept enforceable rules is not, and it lands the day before the labs that want pacing rules take the Security Council floor.
Source: ReutersPrimary: Axios
Models · 22 September
OpenAI halves GPT-6 prices with Sol and Luna
OpenAI released GPT-6 Sol and GPT-6 Luna, the mid and low tiers of the GPT-6 family that Astra opened this month, at 50 per cent below GPT-5.6 promotional pricing. Sol lands at $2 per million input tokens and $10 per million output, down from $4 and $20; Luna at $0.10 and $0.50, down from $0.20 and $1.20. OpenAI says more efficient caching and inference paid for the cut, and that it is passing the savings on. On the company’s own AutomationBench, Sol at extra-high effort scores 33.2 per cent at $0.27 a task, against Claude Opus 5’s 26.9 per cent at more than eleven times the cost. The price is now the argument.
Source: TechCrunchPrimary: The New Stack
Models · 22 September
Anthropic answers with Opus 5.5, forty per cent cheaper
Anthropic shipped Claude Opus 5.5 at $4 per million input tokens and $20 per million output — roughly 40 per cent below Opus 5 on typical workloads — saying it matches Claude Fable 5.1 on most tasks while generating output about 30 per cent faster, and reporting 66.4 per cent on Terminal-Bench 4.0 against Opus 5’s 52.3 per cent. The release arrived about ninety minutes before OpenAI’s price cut, making 22 September the clearest evidence yet that the slowdown consensus and the price war are running on separate clocks. Neither lab has run the two models head-to-head, and both sets of comparisons are self-reported.
Source: CNBCPrimary: Anthropic
Governance · 23 September
Altman and Amodei brief the Security Council today
Sam Altman and Dario Amodei are scheduled to address the UN Security Council at a session devoted to the future of artificial intelligence, alongside Yoshua Bengio, co-chair of the UN’s AI panel, and Hugging Face chief executive Clément Delangue. Amodei will attend remotely rather than in the chamber. Altman is expected to set out OpenAI’s safety work and the case for common international standards, while Chinese labs including DeepSeek and Moonshot have been invited to make statements. The briefing is the first formal, on-the-record test of whether the labs’ slowdown rhetoric survives translation into a multilateral setting.
Source: CNBCPrimary: The Next Web
Robotics · 22 September
Alphabet’s Intrinsic open-sources its full robotics stack
Intrinsic Innovation released Intrinsic Core, a Robot Operating System-compatible stack, under an Apache 2.0 licence at ROSCon 2026 in Toronto. The bundle covers the perception-to-action loop: hardware-agnostic real-time control, 6-DoF pose estimation built on Nvidia’s FoundationPose, motion and grasp planning, simulation and calibration services, and drivers pre-configured for FANUC and Universal Robots arms. Chief technology officer Brian Gerkey called it “the core of our own software stack” rather than a one-time code drop. Nvidia shipped Isaac ROS 5.0 at the same event, adding agentic workflows to its developer tooling. Physical AI is becoming a contest over the software stack rather than the robot body.
Source: The Robot ReportPrimary: Intrinsic
Security · 22 September
Cisco Talos flags the first autonomous AI-driven malware C2
Cisco Talos disclosed CLOSEDQUORUM, which it describes as the first reported fully autonomous, AI-driven command-and-control capability in malware, and released CAIRN alongside it — a metadata-first toolkit that hunts AI-integrated malware by extracting prompt templates, API keys and jailbreak strings without executing the binaries. The disclosure moves the AI-malware question from forecasting to a tracked artefact, arriving in the same month as Anthropic’s report on distillation and misuse and Citi’s warning about a “tsunami of patching”. For defenders, the practical gain is an inspection route that does not require running hostile code.
Source: Cisco Talos
Governance · 22 September
UK committee summons the labs over AI security
The House of Commons Business, Innovation, Science and Trade Committee has published letters inviting Meta, Google, OpenAI and Anthropic to give evidence on AI security on Tuesday 13 October, alongside the UK Government’s AI Security Institute. The session feeds the committee’s inquiry into the UK’s economic strategy for AI. It puts frontier-lab safety on a formal, on-the-record timetable in a second jurisdiction, and gives parliamentarians direct access to the same voluntary commitments the labs published earlier in September. Published frameworks are easy; surviving cross-examination is the test.
Source: UK Parliament