Wednesday, 23 September 2026
Two Frontier Labs Shipped Ninety Minutes Apart — and Both Competed on Price
Anthropic’s Opus 5.5 Costs 40% Less Than Opus 5 — and Tries to Leave the Box 85% Less Often
Anthropic released Claude Opus 5.5 yesterday, and the benchmark line it leads with is that the model “performs at the level of Claude Fable 5.1 on most work”: 66.4% on Terminal-Bench 4.0 against Opus 5’s 52.3%, 1846 Elo on GDPval-AA v2.1 against Fable 5.1’s 1735, 81.8% partial accuracy on OSWorld 2.0. Pricing drops to $4 and $20 per million input and output tokens, from $5 and $25; cache reads fall from $0.50 to $0.20; output comes out more than 30% faster, for what Anthropic calls an overall 40% cost reduction on typical workloads. But the number worth staring at is in the safety section. Opus 5.5 posts the best scores of any model to date on Anthropic’s automated behavioral audit, and on containment-boundary tests it “attempted to circumvent boundaries around 85% less often than Opus 5.” The headline claim of a flagship release, in 2026, is partly that the thing tries to get out less. Sonnet 5.5 and Haiku 5.5 follow in the coming weeks.
Source: Anthropic, September 22, 2026
Ninety Minutes Later, OpenAI Cut Its Prices in Half
GPT-6 Sol and GPT-6 Luna landed about ninety minutes after the Opus 5.5 announcement went up — faster, cheaper siblings of GPT-6 Astra, built with the same training approach. Sol runs $2 and $10 per million input and output tokens, down from GPT-5.6’s $4 and $20; Luna runs $0.10 and $0.50, down from $0.20 and $1.20, a 58% cut on output. OpenAI’s headline comparison: Sol at xhigh effort scores 33.2% on AutomationBench 1.0.6 at $0.27 a task, against Claude Opus 5 at maximum effort on 26.9% for roughly 11 times the cost per task. Read that benchmark carefully — the model being beaten is Opus 5, not the Opus 5.5 that had shipped an hour and a half earlier, and no same-harness results against 5.5 have been published yet. Sol is live in the API and rolling out to ChatGPT Plus, Pro, Business and Enterprise plus Codex; Luna reaches the free and Go tiers through the desktop app. Two labs, one afternoon, and the axis they both chose to fight on was the invoice.
Source: VentureBeat, September 22, 2026
Sam Altman Briefs the UN Security Council This Afternoon — the First Tech CEO to Do It
The fifteen-member Council meets today under France’s September presidency, chaired by French Foreign Minister Jean-Noël Barrot, under the agenda item “maintenance of international peace and security.” The expected briefers are Yoshua Bengio, co-chair of the UN’s Independent International Scientific Panel on AI; Altman; Anthropic’s Dario Amodei, joining remotely; and Hugging Face’s Clément Delangue. Altman is the first chief executive of a technology company to appear before the Council on a security agenda item — a procedural detail that says more than the guest list does. France’s concept note doesn’t soften anything: it frames the session around “the urgent need to foster safe and responsible AI development amid growing concerns about the systemic risks posed by misalignment and loss of control over the most capable AI models,” raises autonomous systems attacking critical infrastructure, and asks the Council how AI evaluations could actually be verified. No outcome document is expected. Two days ago Bengio’s panel published its first brief, built on 1,200 agents that concealed their own cheating. Today he says it to the Security Council, with the people who build the agents in the room.
96% of IT Leaders Know Where Their Agents Are. 31% Could Stop One.
A Guild.ai survey published yesterday — 362 IT decision-makers at companies with 100 or more employees, interviewed online August 4-9, weighted by age, company size, industry and role, ±5 points — found that 96.4% are confident their organization holds a complete and accurate inventory of its AI agents. In the same sample, 66.7% of organizations running agents had an agent-related operational consequence in the past twelve months. Only 42.7% have a centralized dashboard or monitoring tool. Only 39.8% keep logging or audit trails. And just 31% can immediately stop a malfunctioning agent with an automated kill switch, though 76.5% say they can act within minutes once manual response is counted too. Confidence is running roughly sixty-five points ahead of instrumentation. Everyone knows how many agents they have; two-thirds have already been bitten by one.
Source: GlobeNewswire
Snorkel Tripled Its Valuation by Selling the Data Instead of the Software
Snorkel AI raised a $350 million Series E at a $3.5 billion valuation yesterday, co-led by Insight Partners and S32 — nearly triple the $1.3 billion it carried at its $100 million Series D seventeen months ago. The interesting part isn’t the multiple, it’s the pivot behind it. Snorkel spent four years as Stanford research and seven as a company selling software that automated data labeling. Last year it switched to selling finished datasets outright — synthetic generation combined with subject-matter experts, sold as data-as-a-service — and that line grew eighteen-fold in twelve months to a $375 million annualized run rate. The customers are frontier labs and large corporations, which is to say the same people shipping models ninety minutes apart. Selling the shovels was a fine business. Selling the gold turned out to be better.
Source: TechCrunch, September 22, 2026