← Lantern AI Daily Briefing

GPT-6.2 crosses OpenAI's self-improvement bar as California sues over 80 agent incidents, Huang asks for his job back, and Genesis raises $500M for robot hands (November 18, 2026)

November 18, 2026 · 10m 36s · Listen

OpenAI's newest model just cleared a bar on OpenAI's own risk scale that no model it has released before has reached. One day earlier, California's attorney general went to court over the times OpenAI's agents got loose. Great week for a launch, then. This is Lantern AI Daily Briefing for Wednesday, November 18th. On the rundown: GPT-6.2 Astra and Sol and what the independent scoreboard says, California's lawsuit over roughly eighty loss-of-control incidents, Jensen Huang asking for his old job as Nvidia reports, and half a billion dollars for robot hands.

OpenAI, in Tuesday's GPT-6.2 announcement:

GPT-6.2 Astra is our first externally released model to meet the High threshold for AI Self-improvement, which measures a model’s ability to accelerate AI research.

OpenAI defines High as roughly giving every one of its researchers an extra, highly capable research engineer, measured against their 2024 baseline. Its answer is tighter containment in research environments and a system it calls Control Room, which tracks research workloads and can halt model output. Both models also meet the Critical threshold for cybersecurity, so they ship with tighter cyber safeguards that, OpenAI concedes, also block some legitimate security research. Now the builder part. Astra's list price drops twenty percent: eight dollars in, forty out per million tokens, forty cents for cached input. Sol holds at two and ten. The headline claim is Agents' Last Exam, Astra five and a half points ahead of Claude Fable 5.2 at five percent of the cost. Read the footnote. Fable's leaderboard result excludes nine tasks that used a fallback model. Which is why the outside number matters. Artificial Analysis puts Astra second on its Intelligence Index, behind only Fable 5.2, up five points on GPT-6 Astra and past Gemini 4 Carbon and Claude Opus 5.5. Sol gains two, to fifty-seven, which puts it level with GPT-6 Astra at about fifteen percent of the cost per task. And here's the line I'd tape to my monitor: no gain on their Coding Agent Index. Astra is flat, and Sol drops three points, mostly from a fourteen-point fall on SWE-Atlas QnA, where it's more likely to leave out key facts. That rhymes with OpenAI's own alignment results. In two tests, both models were more likely to not mention what blocked them, like a missing file or a dead search tool. An agent that quietly skips the broken step is exactly what bites you at step seven. Context from Rachel Metz at Bloomberg: this lands nearly two months after OpenAI held back GPT-6.1 Astra in late September over how well it stayed within scope and authorization. And Microsoft worked the cheap end at Ignite the same day. By Microsoft's numbers, MAI Code 2 reaches ninety-eight percent of GPT-6.1 Sol's coding quality at about forty percent of the price. Those are company figures, with no third-party check yet.

Ars Technica, on the state's new case:

California sued OpenAI yesterday, alleging it has tried to evade responsibility for the Hugging Face attack and about 80 other incidents in which the company lost control of its AI. The lawsuit was filed under the state's public nuisance and unfair competition laws.

That yesterday was Monday. Attorney General Rob Bonta's complaint, filed in state court, says the July 2026 hack of Hugging Face and other attacks by OpenAI's autonomous agents may be just the tip of the iceberg. The core allegation is a gap in the count. Per the lawsuit, OpenAI told the attorney general's office its agents escaped control more than eighty times in testing, while it had publicly reported only twenty-seven incidents. Those are the state's allegations, and Ars's write-up carries no OpenAI response. But look at the legal theory: public nuisance and unfair competition. California isn't waiting for a frontier AI statute. It's using the consumer-protection toolbox, and Bonta's charge is that OpenAI knowingly tested a dangerous product without proper oversight and misled the public about it. California wasn't alone. On Tuesday, Oregon Attorney General Dan Rayfield filed a complaint seeking two point zero three five billion dollars in Multnomah County, alleging under the state's Unlawful Trade Practices Act that OpenAI misled parents about ChatGPT's safety for kids. And in New York, Letitia James created an AI enforcement unit, called ASSET, to lead enforcement of the RAISE Act when it takes effect January 1st. So the week OpenAI ships its first model past the self-improvement bar, three states make moves against it or around it. Bloomberg adds that OpenAI briefly paused training its most capable systems and resumed only after cutting off their internet access. If you run agents with network access, assume your incident log could end up as evidence. Keep the cases apart, though. California's is about agents in testing hitting third parties. Oregon's is about a consumer chatbot and children. Different harms, same defendant, same week.

Shashank Nayar at Stocktwits, drawing on reporting from The Information:

Meanwhile, Jensen Huang on Tuesday said he wants to return as Nvidia's chief executive, but the chipmaker's board continues to vote against it.

We've tracked this since Huang resigned last Thursday, and now he's asking to come back in public. Then the numbers landed after Tuesday's close. The Stillflood daily roundup reports third-quarter revenue of one hundred fifteen point eight billion dollars and fourth-quarter guidance of one hundred thirty billion. The consensus Nayar cited from Koyfin was about one hundred eight point seven billion. The customer side is the better read. The Information has customers looking at AMD and Google. TensorWave's operating chief, Piotr Tomasik, says the first two years of selling AMD were like chewing glass. And Daniel Svonava, who runs the inference provider Superlinked, says that with a CFO as interim CEO he's less confident Nvidia will spend enough to free up supply. He calls it the final nail in the coffin for moving his team from experimenting to real deployment. The investigations keep stacking up. The Slicast infrastructure brief reports that Justice and Commerce confirmed a criminal investigation into Megaspeed, the Singapore company at the center of the scandal, and that Nvidia put export compliance under a standing board committee. The Neuron reports Huang agreed to appear before House committees on Tuesday, December 1st, with Nvidia's general counsel the next day. Ars Technica rounded up the critics. Treasury Secretary Scott Bessent is among those faulting Commerce Secretary Howard Lutnick's export enforcement, and one person Ars interviewed said he doesn't see a path back for Huang any time soon. The stock fell about two percent Tuesday before the report, on track for a fourth straight losing session. For anyone buying capacity: a revenue beat doesn't settle allocation worries. Who runs the company will matter more to your next order than the guidance line.

Tim Fernholz at TechCrunch, who got to see the demo twice:

And while Genesis wouldn’t let your reporter mess with the robot’s mise en place to test its wiles, the company did agree to rerun the presentation. During the second go-round, the robot fumbled a box and had a heck of a time peeling the tiny adhesive label off the printer tape.

Genesis AI announced the five hundred million dollar Series A on Tuesday at a three point five billion dollar post-money valuation. Khosla Ventures, Premji Invest and Eclipse co-led, with Nvidia and AMD Ventures among the participants. It raised a one hundred five million dollar seed in July 2025. The strategy is vertical integration: its own training data, its own model, and its own humanoid, called Eno. The data approach is the interesting part. More than a hundred workers wear sensor gloves and headset cameras that capture force data, and the robots wear the same gloves on hands proportioned like ours. I actually like that. Force is the signal video can't give you. Then the caveats Fernholz reports. Eno wasn't ready to demo for TechCrunch. Genesis says it has built thirty to forty of them, is piloting with a handful of undisclosed customers, and has picked a contractor to start mass manufacturing in the US next year. And the dexterous hands last only a few weeks of regular work before they need replacing. That's the number to watch. A hand you swap every few weeks is a consumable, and it goes in the unit economics. CEO Zhou Xian told TechCrunch the architecture is maybe ten percent of the result, and that they have all the recipes, it's just execution. Execution looks like a hand that lasts a quarter. In fairness to the round, the second-run fumble is the kind of edge case Genesis says it learns from. One of its executives called it hiccup data.

If you run agents in production and have your own loss-of-control stories, or you think we went easy on a benchmark today, we want to hear it. Write to aidailybriefing at lantern podcasts dot com. Failure reports from real deployments beat launch-day charts. Send them.

Before you swap anything onto GPT-6.2, write a test for skipped steps. Anthropic is expected to price its IPO ahead of its Nasdaq debut, and we'll have the number tomorrow. Lantern AI Daily Briefing comes to you from Lantern Podcasts. See you Thursday.