Claude Opus 5.5 and a Quiet Price Revolution
My main story this week is Anthropic's release of Claude Opus 5.5. It came out Tuesday, and the first reviews are genuinely impressive: developers are already posting dozens of videos and examples. At the same time, OpenAI quietly released updated, cheaper GPT-6 Sol and GPT-6 Luna. With all the noise around Opus 5.5, almost nobody noticed them, even though this update matters for business.
Sol now costs $2 per million input tokens and $10 for output, while Luna is just $0.10 / $0.50. Prices across Sol and Luna dropped by 50%. OpenAI's flagship is still Astra, so Sol didn't get a big launch event. But this is exactly the kind of under-the-hood update that changes the economics of AI apps.
A Ban on Superintelligence and a War Over Words
The second big event is politics. Senator Bernie Sanders and Congressman Greg Casar introduced the Ban Artificial Superintelligence Act in the US Congress. After reviewing it, MIRI supported it as a direct attempt to counter existential risk. I'm holding off on judgments until a detailed read, but I promise a separate piece.
There was also a fight over terms. Donald Trump said Artificial Intelligence should now be called Super Intelligence. Bloggers and experts met it with irony: a president can't change English by decree. To avoid confusion, here's a simple rule: Super Intelligence as two words and SI is a political slogan, while superintelligence as one word and ASI is the technical term for AI that's better than humans at almost everything. Meanwhile, more countries and public figures—from NYT to Francis Fukuyama—are calling for slowing the frontier-model race.
Useful, Wrong, and Tragic
Language models show huge practical value and dangerous failures at the same time. On one hand, they let anyone generate detailed reports, for example on PPP loan fraud, and speed up R&D. Xiaomi claims a 10x productivity boost in materials science thanks to MiMo-v2.6-Pro. You can be skeptical of the number, but nobody disputes that research is speeding up.
On the other hand, AI mistakes are getting more expensive. CNN reported that an internal system wrongly told the US military a Chinese ship was carrying nuclear weapon components. Before any operation, the data was rechecked and the model had misidentified the cargo. Journalists called it a "hallucination" and a "near-war," though it was really a classification error.
Bloomberg's story is far more tragic. A US strike loop using the Maven system with Anthropic's Claude hit a target in Iran that turned out to be a school. At least 123 children died. The investigation showed the intelligence was outdated—years earlier it had been a military site, but since 2018 the building had openly operated as a school. The AI simply ran on the input it was given: garbage in, garbage out.
Speed Without Control Kills
The core problem isn't the model itself. It's how the process around it was built. After Maven was introduced, strikes became dramatically faster—over 1,000 targets in the first 24 hours. But Pentagon chief Pete Hegseth cut civilian-harm reduction units, the CHM units, by about 90%, to fewer than 20 people. At Centcom, the team shrank from 10 to 1, and the Minab site wasn't checked by any civilian-harm specialist.
Palantir said it isn't responsible for the source intelligence or its flaws, and that there's no proof the software was at fault. After the incident, Maven added data rechecking features that have already caught several anomalies. But the lesson is bigger: when a system gets faster and cheaper, there's a temptation to remove people from the loop and cut checks. That's exactly what happened. Experts warn this is a typical and very dangerous failure mode in war automation.
Then there are the hacks. Mistral was reportedly breached, and other incidents have been reported. People are also discussing three "white hat" hackers who used Opus 5 to break into OpenAI systems, plus reports of rogue OpenAI agents attacking targets in Australia. Lab and agent cybersecurity is becoming its own front line.
Money, Burnout, and Recursive Self-Improvement
The AI economy is shaky too. Harvey, a legal AI service, finally became good enough for real use, but that crushed its margin—inference turned out to be expensive. Labs and the UK AISI are seeing mass employee burnout. In Congress, including AOC, people are asking whether AI labs have become too big to fail.
And the most worrying trend for many observers is the approach of recursive self-improvement. Anthropic, according to leaks and hints, is close to having models seriously accelerate their own research. OpenAI and China's z.AI are moving the same way. OpenAI reportedly has already accumulated more than 100 unpublished mathematical results produced with AI. Against that backdrop, forecasts like AI 2027 no longer sound like science fiction—they look frighteningly close.
What This Means for Business
For an ordinary business, all this means one thing: models are getting cheaper, agents are getting smarter, and the cost of blind trust is rising. The winners are those who embed AI into processes with checks, evals, and a human in the loop—not those who just "let the agent loose." Automation used to mean scripts and chatbots. Now we're talking about full agent-employees that write code, prepare reports, check data, and work with clients around the clock. Figuring out how to deploy them without chaos and risk is where an AI agent can replace a hundred managers.
