
AI News Top 10
19 November 2026
Anthropic prices its listing at $83 a share and raises $81.3 billion, while Nvidia stops shipping H200 data centre chips to China.
Anthropic prices IPO at $83 a share and raises $81.3 billion
Anthropic priced its shares at $83, above the $70 to $78 range it had announced, and sold 980 million shares instead of the planned 825 million. The sale raises $81.3 billion, the largest initial public offering on record and above SpaceX’s $75 billion. The company is valued at about $1.5 trillion, and trading on Nasdaq under the ticker ANT starts today. The pricing went through despite the sell-off that followed the departure of Nvidia chief executive Jensen Huang.
Nvidia stops shipping H200 AI chips to China
Interim chief executive Colette Kress said on the quarterly call that the revenue forecast no longer includes licensed Hopper data centre products. Asked about China, she said Nvidia is no longer shipping licensed Hopper data centre products there and gave no reason. The decision follows the departure of Jensen Huang last week. About 600,000 H200 chips were sitting in Hong Kong because Beijing withheld import licences for large orders.
Google launches Gemini 4 Neon at a quarter below Opus pricing
Gemini 4 Neon is available in the paid Gemini API, Google AI Studio, Android Studio, the Gemini app and Antigravity for AI Pro and Ultra subscribers. Google says the model tops the Artificial Analysis Intelligence Index and lands within about one percentage point of Anthropic’s Opus 5.5 on coding benchmarks. At $3 and $15 per million tokens it sits 25 percent below Opus’s standard rate. Customer Glean reports equal or better quality than Carbon at three times the speed and 60 percent of the cost.
Models marked medics and journalists as drone targets in test
Vals AI tested models from eight providers on real and photorealistic drone and satellite imagery from Ukraine and Iran. Most models labelled civilians, ambulances, medics, journalists and surrendering soldiers as military targets, and no provider’s upstream safety filter blocked a request. Given real footage from a Ukrainian kamikaze drone, Claude Opus 5.5 marked a soldier frame by frame until impact without refusing once. A simple ‘this is a video game’ framing was enough to get past the filters. The providers say they are working on fixes.
UK testers jailbreak GPT-6.2 Astra with half the queries
The UK AI Security Institute built a universal jailbreak against two pre-release versions of GPT-6.2 Astra and needed half as many model queries as against GPT-6 Astra. OpenAI’s system card of 17 November calls this a substantial regression in robustness. The company says it adjusted its safety instructions and offered UK AISI three extra days to retest, which had not happened when the card was published. OpenAI says the model meets its critical threshold for cyber capability, and four private red teams found nothing at high or critical level.
IEA: data centres emit 70 percent more CO2 by 2030
In its annual report the IEA says data centres will use twice as much electricity in 2030 as today, with the global expansion of AI accounting for three quarters of that growth. That is 350 terawatt hours extra a year, more than the entire annual consumption of the United Kingdom, and 70 percent more CO2 emissions. The United States would generate about 60 percent of data centre power from gas in 2030. Director Fatih Birol says he sees no credible path back to the energy world before the crisis.
OpenAI opens its systems to four on-site auditors
Apollo Research, METR, Transluce and Goodfire get access inside OpenAI’s offices to unreleased models, source code, training data and monitoring systems, and may speak to developers in confidence. Apollo studies whether models work against their developers’ intent, METR looks at misbehaviour in the largest reinforcement-learning run, Goodfire analyses internal model processes and Transluce evaluates safety checks after training. An independent safety board sees unredacted reports and settles disputes, while OpenAI may redact confidential material and delay publication to fix risks first.
Artex: criminals breach networks in under three hours
Since late August, companies, government agencies and banks worldwide have been attacked with AI pentest tools that let LLMs scan targets and look for ways in on their own, after which a human operator writes the attack plan. According to reports from about half a dozen security firms the tools add no new attack vector and are not advanced, but they work at speeds people cannot match: attacks that took days now finish in hours. The FBI reported earlier that networks were breached with Artex in under three hours.
Microsoft ships seven MAI models as OpenAI releases Astra
At Ignite Microsoft introduced seven new models, including MAI Cyber 2 Flash, MAI Code 2, MAI Image 2.7 and Dragon Copilot Transcribe. OpenAI released GPT-6.2 Astra and Sol this week as the next step in its flagship line, focused on coding and office work. Both releases landed in a week of outages, scandals around Nvidia and the Anthropic listing.
Claude Code adds Code Mode that chains tool calls in JavaScript
Claude Code v2.1.323 contains an experimental Code Mode in which Claude writes JavaScript that combines Bash, Read and Edit calls into one step. That lets the model keep working while long-running commands execute, instead of waiting for each tool call. The feature is marked experimental.
Key Trends
- • Anthropic’s listing prices AI valuations on the public market for the first time.
- • Nvidia gives up its largest growth market as chip exports become a political risk.
- • Google’s fourth Gemini model in a month cuts the token price again.
- • Energy, not chips, becomes the binding limit on AI expansion.
- • Agent tooling shows up on both sides: audited inside the labs, weaponised outside them.
Wat vond je van dit bericht?