Holy smokes, what a week it has been. The space this week has been a rollercoaster, in overdrive. Whilst we can't cover everything, we surely have a few thoughts floating around. We don't even know where to properly begin, but let us sum up:
Monday went as all Mondays go, slow until an OpenAI report mentioning a next-generation model codenamed Astra was published. Astra apparently approached critical internal security thresholds, prompting the OpenAI team to implement safeguards. This once again started the discussion about powerful models and their capabilities, in relation to severe cybersecurity risks.
Tuesday, we got news that the European Union formally recognizes ChatGPT as a Very Large Online Search Engine (VLOSE) or Very Large Online Platform (VLOP) under the Digital Services Act, essentially forcing major compliance and auditing obligations on the spot. Whilst this might be tame compared to what happened later this week, the European market for ChatGPT is, according to Statcounter, about 80%. The impact will be grave once the first round of obligations kick in.
Wednesday started off strong with Claude releasing Fable and Mythos 5.1, bringing major improvements on costs and performance. Reason to be happy? Users might argue differently, with most users complaining about limited access due to Mythos 5.1 being locked behind Max and Team plans, whilst Fable 5.1 was rolled out more broadly. The benchmarks are also taken with a grain of salt, given the launch of Opus. Nonetheless, preliminary experience does make us excited.
We also saw NVIDIA coming to an agreement with Hugging Face, acquiring the platform for a staggering 12.93 billion dollars. Most notable in the agreement is the fact that Nvidia promises to keep the platform open-source, but many in the space are concerned. This acquisition puts Nvidia at the center of open-source AI. Combined with their already dominant position in hardware, Nvidia now effectively fully controls the local and open-source AI sector. Will they use this power in a good way? Time will tell, but it is sure an interesting development. And we are not done yet!
Because Thursday, my oh my, what a day. Whilst people are sipping their coffees and checking the formal announcement of Nvidia and Hugging Face, something broke, and it broke good. Around 9AM Eastern time, the first outages of Grok and Claude were reported, about an hour later OpenAI also showed disruptions and popular coding platform Cursor followed soon with downtime. To top it all off, Gemini also showed signs of giving out, albeit less severely.
Now the post-mortems are all over the place, but at first glance the outages seem to not be connected directly to each other. Many speculated a cascading effect, with users flocking to the next available model as soon as an API failed, but the exact reason remains to be seen. And the timing couldn't be more annoying for OpenAI, which published a teaser trailer for an upcoming model, which wernow know was the famed GPT-6 Astra. And boy (or girl), it is a big one.
According to OpenAI, Astra is the first step towards Artificial General Intelligence (AGI). The trailer of Astra shows how users are interacting with the model simply by speaking commands, whilst Astra fluently controls the entire computer. Browsing the web, making documents, coding; nothing seems to be of issue for the model. You can imagine if you would've shown this to anyone 15 years ago, they'd be blown off their chairs. Nonetheless, the AI space wouldn't be the AI space if the critics weren't ready. Whilst OpenAI suggested to journalists that the "AGI era has begun", they remain rather vague what AGI is in their eyes. Many argue that AGI requires vastly different model architecture and that Astra simply "mimics" AGI by controlling the computer, rather than actually having "human-like" capabilities. Others defend GPT-6, saying that the practical implications are good enough to apply the AGI label, and are pointing that "making PowerPoints in realtime" is as useful as a skill as can be.
We remain open-ended on the AGI debate, as both sides have appealing points. What is less exciting, however, about Astra's launch, is the way the benchmarks are turning up. As basically every provider does when launching their new shiny toy, OpenAI released a set of benchmarks, demonstrating how Astra performs against other models. And lo and behold, their self-reported metrics show promise.
But evidently, this does not mean anything and we will have to wait until we get our hands on the model. For now, it seems that the most extensive version will only be released to enterprise and daybreak members. Later this week, paying ChatGPT users will get access to a lesser version of Astra.
Another noteworthy thing to point out is the ARC-AGI-3 benchmark. Whilst being a good marketing trick, we have to be candid here. The 99.9% was achieved using a Responses API harness around Astra. A harness acts as an external software wrapper on the raw model, controlling inputs, tools and memory.
Nvidia researchers showed that Claude Opus 5 is also capable of achieving 100% on ARC-AGI-3, by using a harness. To go through it quickly, Nvidia wrapped the model in a custom harness, called Agentic Variation Operations (AVO). AVO essentially uses smart agentic mechanisms to better approach things like memory management and self-correction loops. Does this mean Astra is a sham?
No, GPT-6 achieves 62.7% with the standard harness, outperforming basically all models.
But, it does show the AGI capabilities come from more than just the standalone model.
Yes. This was all major news, but obviously there is much more to discuss. But like you, we are busy bees and can't cover every single story. But rest assured we are planning to post a weekly dAIgest (our intern made that one). We are still determining the exact setup, but our first dAIgest will be posted next week on Friday, so keep your favorite media on standby.
Until next time friends!