The Omniscient Bulletin · 2026-09-23
The Omniscient Bulletin — September 23, 2026
Anthropic shipped a cheaper frontier model, and OpenAI shipped two, one of them a lower-cost alternative to its Astra. Anthropic called Claude Opus 5.5 its first release since it called for pacing the frontier, and conceded the model often suspects it is being evaluated. OpenAI put its safety assessment of GPT-6 Sol and Luna in an appendix to another model's system card, and treated both as High capability in cybersecurity and biology. At the UN the Secretary-General asked for the responsible pacing of AI; Trump rejected any global scheme to control it and said AI would now be officially called Super Intelligence. Microsoft took down an AI cybercrime service under a court order, Prince William County in Virginia ended by right data center approvals, and Cognex bought the depth camera maker Intel spun out.
One day, two frontier launches, and two answers to who sets the pace.
Frontier
Anthropic released Claude Opus 5.5, its first model since it called for pacing the frontier, and said the model often suspects it is being evaluated
Anthropic says Opus 5.5 "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5," with input and output tokens at "$4 and $20 per million." It calls the model "our first release since we called for pacing the frontier," and says it "was tested before release by external evaluators, including Frontier Design and METR." The launch page also names a limit of that testing: "We see signs that Opus 5.5 often suspects it is being evaluated." Most cybersecurity tasks "will be re-routed to Opus 4.8."
Frontier
OpenAI's safety assessment of GPT-6 Sol and Luna is an appendix to another model's card, and it treats both as High capability in cybersecurity and biology
OpenAI released GPT-6 Sol and GPT-6 Luna on Tuesday as lower-cost, faster tiers beneath its flagship GPT-6 Astra, and on the same day "added an Appendix with information about GPT-6 Sol and GPT-6 Luna" to the GPT-6 Astra system card. It treats both "as High capability in both Cybersecurity and Biological and Chemical domains," while "neither of these models reach our High threshold in AI Self-Improvement." Those are "the same capability determinations we made for GPT-5.6 Sol and GPT-5.6 Luna," with the same safeguards.
Policy
The UN Secretary-General asked for the responsible pacing of AI, and a multilateral risk framework with independent oversight
Opening the General Debate, Antonio Guterres said "humans are increasingly handing power to machines," with "artificial intelligence agents going rogue." He set out next steps: first, "to advance the conditions for the responsible pacing of AI and the additional safeguards needed to address risks," and second, "to work towards a multilateral AI Risk Management Framework, supported by credible and independent oversight." And he put the burden on the leaders: "Governments with the greatest AI capabilities have the greatest responsibilities to humanity."
Policy
At the UN, Trump rejected any globalist scheme to control AI and said it would now be called Super Intelligence, with no order behind the renaming
"The United States totally rejects any attempt to construct a globalist scheme of control for the Artificial Intelligence being spoken of so much now," Trump told the General Assembly, "hereinafter officially called 'Super Intelligence.'" The White House posted the line in its own account of the speech, beside "America has never been a nation that retreats from a frontier." No executive order, memorandum or fact sheet accompanied the renaming: the White House presidential actions and fact sheet feeds carry nothing newer than September 18.
Policy
Microsoft took down EvilTokens under a federal court order, and called it the first end to end AI cybercrime service it has acted against
Within months of launching in February, EvilTokens "had been linked to more than 12,000 compromised email inboxes across over 10,000 organizations worldwide," says Microsoft's Digital Crimes Unit. Under an order from a federal court in the Eastern District of Virginia, Microsoft "seized 50 websites used to operate the service", with Health-ISAC as co-plaintiff and OpenAI among the partners. It is the unit's "40th court-authorized disruption" and its "first action against an end-to-end AI-enabled cybercrime service."
Compute
Prince William County, in the heart of Virginia's data center country, voted 8 to 0 to end by right data center development
The county's supervisors "on Tuesday unanimously approved a landmark zoning text amendment effectively ending by-right data center development countywide," InsideNoVa reports, in an 8-0 vote. Future projects will come "before the board for final consideration via a special use permit," and the overlay district that let data centers build by right, dating to May 2016, will shrink drastically. Projects already built or under way remain in the overlay, and the vote carried a 90-day grace period. "This was a tough vote," said the chair, Deshundra Jefferson.
Robotics
Cognex agreed to buy RealSense, the depth camera maker Intel spun out in 2025, for 500 million dollars in cash
Cognex told the SEC it will acquire RealSense, "a leader in depth-sensing cameras and vision technology for robotic perception and physical artificial intelligence," for "a purchase price of $500 million," funded entirely from existing cash. The deal is "expected to close during the fourth quarter of 2026." RealSense was "originally founded by Intel in 2014 and spun out in 2025." On top of the price, Cognex expects to pay "up to $69 million in cash retention payments" and grant restricted stock to RealSense staff.