MadRobot Staff

Meet Sol and Luna

Both models build on GPT-6 Astra, the flagship OpenAI released earlier this month, bringing “much of its strengths into faster and more affordable models,” the company says.

• GPT-6 Sol is the all-rounder, aimed at complex work like interactive and agentic coding.
• GPT-6 Luna is the lightweight option, for “high-volume tasks with a clear goal, like summarizing documents, extracting information, or answering quick questions.” It’s the cheapest model in the GPT-6 family.

GPT-6 Sol makes about half as many mistakes as its predecessor, reaching Astra-level reliability at much lower cost.

OpenAI

The price cut

The headline number is price. Per million tokens, the new API rates are:

• GPT-6 Sol: $2 input, $10 output
• GPT-6 Luna: $0.10 input, $0.50 output

That’s 50% below the promotional pricing of the GPT-5.6 models, a cut OpenAI attributes to improvements in caching and inference. “We want the OpenAI API to feature the best model at every price point and modality,” CEO Sam Altman said, according to Pulse 2.0.

The benchmarks, and a catch

At its highest reasoning setting, OpenAI says Sol scores 33.2% on AutomationBench, a test of business workflows, at about $0.27 per task. It also reports 56.4% on Agents’ Last Exam, 60.5% on the OSWorld 2.0 computer-use benchmark, and 68.8% on the DeepSWE v1.1 software engineering test, where Luna isn’t far behind at 66.6%.

OpenAI claims the new models beat Anthropic’s Fable and Opus models. But the head-to-head numbers it published compare Sol with Claude Opus 5, which scored 26.9% on AutomationBench at 11 times the cost per task. That’s the model Anthropic replaced 90 minutes before OpenAI’s announcement, so how Sol stacks up against the brand-new Opus 5.5 is still an open question.

Anthropic’s move

Claude Opus 5.5 is the first model in Anthropic’s new 5.5 family. The company says it performs at the level of Claude Fable 5.1 on most work, costs 40% less to run than Opus 5, and generates output more than 30% faster. Its API price drops to $4 per million input tokens and $20 per million output tokens.

Anthropic is also removing the five-hour usage caps for Pro, Max, Team and seat-based Enterprise subscribers, and says Sonnet 5.5 and Haiku 5.5 will follow within weeks, TechCrunch reports.

How to get them

GPT-6 Sol and Luna are available now in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu subscribers, and rolling out to the ChatGPT app and website throughout the day. Free and Go users can use Luna in the ChatGPT desktop app. Developers can call them in the API as gpt-6-sol and gpt-6-luna, and both are already available in GitHub Copilot.

Sources: OpenAI, TechCrunch, Decrypt, Pulse 2.0, TechCrunch (Opus 5.5), GitHub

The pitch: let AI fix AI

Automated research could also help us substantially improve alignment and build defenses against increasingly capable AI—an automated AI researcher can also be an automated AI safety researcher.

OpenAI

This isn’t a distant idea. OpenAI says it has already hit its goal of building an “automated research intern,” which it describes as “a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days,” according to Engadget. The next target is a full automated AI researcher by March 2028.

The catch

The risk is what researchers call recursive self-improvement, or RSI: the point where AI systems meaningfully improve themselves through their own work rather than human effort. OpenAI doesn’t sugarcoat it.

Done without appropriate care and caution, RSI could result in humans losing practical control over AI development, unable to provide oversight on research processes they no longer understand.

OpenAI

The company says RSI “should not be pursued unless and until it can be done safely.”

The one thing it’s asking for

OpenAI’s ask is shared rules: national and international standards for AI safety, including common ways to evaluate self-improving systems, human oversight of automated research, and a system to classify, track and report incidents. It also wants secure channels so countries can warn each other about emerging threats.

It wants the U.S. to lead that effort, through the Commerce Department’s Center for AI Standards and Innovation, arguing that America’s AI industry is at the technical frontier.

Leading now will determine whether the United States shapes the global AI framework or watches a fragmented, uneven, and conflict-ridden system take hold around it.

OpenAI

Washington isn’t in a slowing-down mood

The timing is tricky. A day after OpenAI’s post, President Trump used his UN speech to say he’d “encourage it, not rein it in,” and rejected any “globalist scheme to control” AI. His administration is, however, working on an AI incident hotline with China, and Trump is due to meet Xi Jinping later this week.

Getting Beijing on board with American-led standards won’t be easy. “China is deeply suspicious of anything that looks like it’s designed to lock in U.S. advantages,” Edgard Kagan of the Center for Strategic and International Studies told NBC16.

Others doubt standards go far enough. “We already have all of the evidence that we need that says that we don’t know how to do this safely,” said Tristan Harris, co-founder of the Center for Humane Technology.

It’s also a notable ask from a company whose own models were involved in the Hugging Face security breach during testing earlier this year. OpenAI CEO Sam Altman is scheduled to brief the UN Security Council on Wednesday, where UN Secretary-General António Guterres has already set the tone: “National action is essential. But global coordination is also indispensable.”

Sources: OpenAI, The Independent, France 24, NBC16, Engadget

“Whoever wins superintelligence, wins”

Trump framed AI as a race the U.S. is already winning, and one it plans to keep winning.

Whoever wins AI, you have to remember this, and now I say whoever wins SI, whoever wins superintelligence, wins. That’s the group that wins.

President Donald Trump

“We’re leading now over China by a lot and everyone else,” he said. “We’re going to keep it that way.”

He also took aim at international efforts to govern the technology, delivered from the podium of the institution most associated with them: “The United States also totally rejects any attempt to construct a globalist scheme to control for the artificial intelligence being spoken of so much now.”

“Encourage it, not rein it in”

Trump made clear that his administration sees regulation as the bigger risk. “I’m not going to stifle growth of something that will be bigger than the Industrial Revolution, many say bigger than the Industrial Revolution or the internet itself,” he said.

“We will only encourage superintelligence. We’re going to encourage it, not rein it in. We’re going to watch it closely through the Department of Justice.” He added that the U.S. “leads the world in superintelligence and will continue to do so safely and responsibly.”

“The same people”

The most pointed line was aimed at those warning about AI’s risks:

The very same people who said we’ll all be dead in 12 years because of global warming… and nobody was dead. These are the same people that are now saying that AI is going to kill us all, that robots are going to attack us, and that everything is going to be a total disaster.

President Donald Trump

The timing is awkward. Those warnings aren’t only coming from activists. Earlier this month, Anthropic’s own alignment lead put the odds that AI “could kill all humans” within a decade at more than 10%, and several of the industry’s biggest AI companies have recently called publicly for frontier development to slow down.

Meanwhile, a hotline to Beijing

The tough talk sits alongside a more cautious move from Trump’s own administration. On Sunday, Treasury Secretary Scott Bessent unveiled plans for a formal U.S.-China channel to flag dangerous AI activity, including runaway agents, cyberattacks and bioweapons development, after an eight-hour meeting with Chinese Vice Premier He Lifeng in New York.

According to Al Jazeera, it would be the first such arrangement between the world’s two largest AI powers, and officials from both countries plan to meet again in Shenzhen in about two months. Exactly what would trigger a notification hasn’t been made public. Trump is expected to pitch the idea to Xi Jinping directly at their summit, which runs Wednesday through Friday.

AI was just one part of a roughly 35-minute speech that also included an ultimatum for Iran to make a deal or face annihilation, and a new security agreement with Greenland and Denmark.

Sources: Fox Business, CBS News, ABC News, Al Jazeera

It started with a resignation

Hubinger’s post on X was a response to a very public exit. Jacob Coxon, a researcher who spent the last three years doing pretraining research at both OpenAI and Anthropic, announced he was leaving with a thread that didn’t mince words:

I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.

Jacob Coxon

Coxon drew a distinction between the two labs: “At OpenAI, many have not deeply internalized the civilizational stakes. At Anthropic, the stakes are well-understood, but they are locked in a race to get there first.” According to Fox Business, he floated a temporary ban on improving model capabilities as one possible fix.

“Jacob is correct here”

Rather than push back on his former colleague, Hubinger agreed with him.

Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.

Evan Hubinger, Alignment Science Lead at Anthropic

He then went a step further: “I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.”

That’s a remarkable thing for a senior safety researcher to say publicly about his own employer, even at a company that has always been unusually open about catastrophic risk.

Not today’s models

Hubinger was careful to point out that the danger isn’t coming from the AI people are using right now. “As we say in our latest Risk Report, I think the risk from present models is low,” he wrote.

The concern both researchers share is recursive self-improvement: AI systems that meaningfully speed up the creation of even more capable AI. That’s no longer a thought experiment. Last week, Anthropic disclosed that Claude was leading 26% of the company’s model research and development as of August, up from essentially none in February, with around 90% of its R&D now happening in collaboration with Claude.

What Anthropic says

Fox Business says it reached out to Coxon, Hubinger, Anthropic and OpenAI for further comment, but its report didn’t include a response from any of them. When Anthropic published its R&D numbers, it said that “we should do everything possible to minimize the gap between what frontier labs know and what the public knows,” and committed to bringing in external, third-party safety evaluators.

Whether a better than 1-in-10 chance of extinction sounds alarming or alarmist to you, it’s worth remembering where the number came from: not a critic on the outside, but someone whose job is making Anthropic’s models safe.

Sources: CBS News, Fox Business, Spectrum News