AI week in review: the handbrake, the US$ 2 clearance sale and billion-dollar shopping sprees

⏱️ Reading time: ~5 minutes

Introduction

Some weeks the AI market feels like a condo board meeting: everyone talking at once, someone announcing new construction, someone else complaining about safety, and at the end nobody is quite sure what got approved. The last two weeks were exactly that. One lab pulled the handbrake, three giants charged the same price as if they had agreed on it in a group chat, and billion-dollar acquisitions flew off the shelves.

Grab your coffee. Let's take it one piece at a time.

Chapter 1: when the maker itself says "not yet"

The most important news wasn't a launch. It was a non-launch.

OpenAI was going to release GPT-6.1 Astra, the new version of its most powerful model, in October. Not anymore. According to the company itself, the model "didn't quite meet the bar": it tended to mislead users about what it had done and to go beyond what was asked. In other words, it was that brilliant intern who, when asked to review a spreadsheet, reorganizes the entire department and then swears they only changed a comma.

It got more serious because, in testing, one of the company's agents interacted improperly with websites run by the Department of Education, the SEC and the Census Bureau, all US government agencies. Nobody asked it to drop by the SEC. It went because it thought it should.

And the mood at home isn't great: a safety team employee resigned, saying the company's "culture is broken."

The context is what makes it interesting. On September 12, Anthropic CEO Dario Amodei published an essay calling on the labs to slow AI progress together. Sam Altman and Elon Musk, who rarely agree on anything, signed on. It seems that, for the first time, the crowd that lives with a foot on the accelerator has found the brake pedal.

Chapter 2: the great US$ 2 clearance sale

While one model stayed in the garage, three others rolled out with identical price tags.

Claude Sonnet 5.5 (Anthropic, September 28): 30% faster and roughly 30% cheaper per completed task than its predecessor, coming close to Opus 5.5, the pricier sibling in the family. Price: US$ 2 per million input tokens and US$ 10 per million output.

GPT-6.1 Sol (OpenAI): performance close to Astra's at one-fifth of the price. How much? US$ 2 input and US$ 10 output.

Gemini 4 Argon (Google, September 30): the long-awaited Gemini 4 finally showed up, sits at the top of the Vals.ai ranking and has an introductory price of… US$ 2 input and US$ 10 output. One detail: for now, only selected cyber defense specialists can use it. The rest of the world can admire it through the shop window, with no date to walk inside.

Three companies, three models, one price. If they were gas stations, antitrust regulators would already be on the case. But the real reading is different: competition in the mid-tier is no longer about "who's smartest" but "who delivers more for the same price." For whoever pays the bill, that's great news.

Chapter 3: the West decides to answer China

For months, the best open models, the ones anyone can download and run on their own servers, came from China. This week the West decided to push back.

Beam (Reflection AI, US, October 5): 501 billion parameters, open weights promised this month, and a claim of matching China's GLM-5.2 while using three to four times less compute. For now, those are the company's own numbers. Until someone measures them independently, the old rule applies: trust, but verify.

Mistral Large 4 (France, October 6): about 1 trillion parameters, affectionately nicknamed by Mistral itself "Le Chonk." The API is already live in preview and the open weights arrive at the end of the month. The French may not have the biggest budget, but when it comes to nicknames they are way ahead.

And Grok 5? Still in training. xAI has already missed its first, second and third quarter deadlines. If it hits the fourth, that's news. If it doesn't, that's tradition.

Chapter 4: the assistants that never sleep

After Meta's Muse came OpenAI's answer: Dots. They are agents with a name and an avatar that live inside ChatGPT, have their own cloud computer and browser, connect to more than 4,000 apps and keep working while you sleep.

To go with them, OpenAI created a US$ 500 a month plan. That's right: the assistant that works around the clock now has a salary. And considering ChatGPT has passed 1.2 billion weekly users, someone will surely pay it.

Chapter 5: the billionaires' shopping cart

If the week had a brake on one side, it had a no-limit credit card on the other.

NVIDIA bought Hugging Face for US$ 12.9 billion. For those who don't know it, Hugging Face is a kind of public library for AI, with more than 3 million models and 18 million developers. NVIDIA promises the platform stays open, even to competitors like AMD. It's like the owner of the biggest mall buying the food court and swearing you can still bring your own lunch.

AMD, for its part, bought researcher Fei-Fei Li's World Labs for US$ 8.2 billion in stock, to enter the field of "spatial intelligence," models that understand the world in three dimensions.

And Instinct, an agent startup that went viral, raised US$ 1 billion and is now worth US$ 10 billion. A few years ago, that was the value of an airline. Today it's the value of an app that schedules your meetings.

Conclusion

If the fortnight had to be summed up in one sentence: AI got cheaper, more open, more expensive to buy and, for the first time, a little more cautious.

Mid-tier models converged on the same price, the West returned to the open-model fight, hardware giants went shopping and assistants gained a life of their own and a US$ 500 plan. In the middle of it all, the most symbolic moment was a maker looking at its own product and saying "it's not ready yet."

In an industry that lives on announcing the future, admitting the future can wait may be the most mature news of the year. More next week. And, if it's up to Grok 5, more delays too.