OpenAI o1

It not only gave better answers, but users got to see how the model thinks and decide if they agree with it or not.

As soon as O1 was released, competitors were rushing to catch up, with Google releasing a competing model a few months later in December.

The thing is, though, that Alibaba had actually beaten Google by releasing their reasoning model called QWQ ahead of Google.

Not only did Alibaba get there faster, but they published the model under an open license, meaning that anyone could dig through it to see how it works.

DeepSeq released a number of models since then, making the code open source under the MIT license, which puts very few restrictions on reuse, allowing users to modify the code even for proprietary commercial use.

Despite its low cost, DeepSeq's scores on AI performance benchmarks show that it is as good if not better than the latest cutting-edge models from the top US firms.

It is almost as good as OpenAI's O1 model in the Artificial Analysis Quality Index, an independent AI analysis ranking, and it beats Google, Anthropic and Meta's models.

They released a large language model in December called V3 and then a reasoning model called R1 on the 20th of January.

Risk Report February 2026.

Claude Opus 4.7 System Card.

Claude Opus 4.6 System Card, PDF.

Claude Sonnet 4.6 System Card, PDF.

Claude Opus 4.5 System Card, PDF.

Claude Sonnet 4.5 System Card, PDF.

Anthropic Responsible Scaling Policy RSP Updates.

OpenAI GPT 5.6 Preview System Card, PDF.

And maybe the last two years, three years or whatever, we often heard that, oh, even chat GPT-2, it was GPT-2 and then chat GPT, too dangerous to release.

Then it was the, if you remember the O1 model from OpenAI, too dangerous to be released.

Then most recently, it was Mythos and then Fable, which is like a safer version of that.

So we had this narrative of this AI is too dangerous for quite a bit of time now.

We really have to look back at the late 2024 pivot point to understand how we achieved the reliability we see today.

Right, because the landscape fundamentally fractured around the time OpenAI released GPT-4.0 and then shortly after the O1 model.

It was a massive fork in the road for the industry.

So let's look at GPT-40 first.

Yeah.

And what's fascinating here is the stark contrast.

The Vellum analysis from our sources highlights that O1 wasn't built for snappy conversation at all.

It was explicitly engineered for hard reasoning.

Pre-01 release.

OpenAI did a hackathon with Y Combinator to let a bunch of companies try O1 early.

Yes.

And it was impressive.

Ethan Mollick, he writes about AI, is a Waterman professor.

He says, we're moving into like the fourth era, right? The previous ones were GPT-3, GPT-4, O1, which was the reasoning models.

And then in December 2025, something changed.

The models become agentic.

And what did they prove exactly?

They demonstrated that reasoning models could be scaled indefinitely.

It wasn't just a theory anymore.

They realized that if you teach a model to think step by step and you pour more computing power into it, It doesn't hit a wall.

And we can't forget that chilling technical paper from Anthopic.

Oh, right.

The one which proves, mathematically, that these AI models can actively recognize when they are being tested.

Which is just, it changes everything.