The AI Race Nobody Asked to Run

There are mornings I open my laptop and the first thing I see is news of a new AI model…again. It has only been three weeks since the last one. I have not finished figuring out the previous one. And yet here we are, being told that some things have shifted.

Somewhere along the way in this AI journey, it somehow stopped being a tech industry story. And yet everyone in marketing or business is expected to keep up, lest you are considered “outdated”. Ready or not, like it or not.

So this is my endeavour to make sense of it all. Definitely not as a technical deep-dive (I’m far from being qualified), but as a practitioner’s honest account of how we got here, who is running, and what it actually means for the rest of us.

It started with a chat box

In November 2022, OpenAI released ChatGPT to the public. It was not the most sophisticated thing they had ever built. But it was the first time ordinary people, not engineers or researchers, could sit down, type a question, and feel like they were talking to something that genuinely understood them.

The model underneath it, GPT-3.5, was actually quite limited by today’s standards. It could only read and write text, had no access to the internet and its knowledge stopped at 2021. It could only hold about 3,000 words of conversation in memory before it started losing track of what you had said at the beginning. And it hallucinated. A lot. It would state things confidently that were simply not true, with no hint of self-awareness about it.

By today’s standards, that is like comparing a Nokia 3310 to the phone in your pocket right now. But none of that mattered in November 2022, because nothing like it had been publicly available before. A million people signed up in five days. The race had begun.

The field fills up fast

If 2022 was the starting gun, 2023 was when everyone else started showing up. The nature of the AI race became very real.

On the 14th of March 2023, OpenAI released GPT-4, a genuine leap. It could read images as well as text, reason through complex problems, and pass bar exams and medical licensing tests. It was still offline though, with no internet access yet. But it was the point where people stopped showing off ChatGPT at dinner parties and started wondering if it might actually be useful at work.

That same month, OpenAI also gave ChatGPT its first taste of internet access via a browsing plugin. It was clunky, unreliable, and still in beta, but it was a signal of where things were heading.

On the exact same day, a new challenger, Anthropic, released Claude 1. Founded by former OpenAI researchers who had left over safety concerns, Anthropic had spent two years building in the background before making their public entrance.

That is barely a coincidence. These companies watch each other obsessively, and the timing was a statement. By the end of 2023, more heavyweights had entered the ring. Google came with Gemini, Meta released its open-source Llama models that anyone could download and use for free, and Elon Musk launched Grok through his company xAI. What had been OpenAI’s race alone had become a five-horse field within twelve months, and those horses were some of the most capitalised companies on earth.

The image generation world was moving in parallel. Stability AI had released Stable Diffusion in 2022, letting anyone generate images from a text description. Midjourney was rapidly improving through multiple versions, getting to the point where creative professionals started using it for real work. The era of “you just described it and it appeared” had arrived for images as well.

The sprint begins in earnest

If 2023 was about the field filling up, 2024 was about the pace breaking into something none of us thought was possible.

In May 2024, OpenAI released GPT-4o, a model that could hear your voice, see your screen, respond in real time, and finally browse the internet properly. The earlier plugin attempt had been shut down a month before. This was the real thing. Anthropic countered almost immediately with Claude 3.5 Sonnet, which became widely regarded as the best model for coding at that moment. Meta released Llama 3.1 in the summer, with versions powerful enough to compete with frontier closed models, completely free and open-source. Grok 2 arrived from xAI.

Then in September 2024, something shifted more fundamentally than any of these launches. OpenAI released a model called o1, and with it, reasoning entered the picture.

Before o1, AI models essentially predicted the most likely next word based on what had come before. Very fast, very fluent, but not actually thinking. o1 changed that. Before answering, it worked through problems step by step, almost like a person sitting by itself, eyes closed, mouth in a slight mumble, before speaking. It scored in the top one per cent on olympiad-level maths problems, making any Asian parent beam with pride. Suddenly AI was no longer just generating text. It started working things out.

OpenAI also released Sora, a video generation model, to extraordinary hype. The demo videos were stunning. A woman walking through Tokyo rain, a woolly mammoth charging across snowy terrain, footage that looked like it had been shot by a professional camera crew. The creative industry collectively held its breath.

The East arrives, and changes the story

Up to this point, the AI race had largely been framed as an American competition. That framing did not survive January 2025.

DeepSeek R1, built by a Chinese company that had started as a side project of a data-driven hedge fund, arrived and matched the performance of the best Western models at a fraction of the compute cost. The detail that made this significant was that it had been built despite US government restrictions on exporting advanced chips to China. The message was uncomfortable as no restrictions had stopped anyone.

Some called it China’s Sputnik moment for AI. That might be slightly dramatic, but the market reaction was not. US tech stocks dropped sharply the day after DeepSeek’s release, as investors digested what it meant for the assumption of Western dominance.

What followed was an open floodgate of Chinese AI development that most Western coverage underreported. Alibaba’s Qwen, ByteDance’s Doubao, Moonshot AI’s Kimi, MiniMax, and a dozen others pushed into markets both at home and internationally. By mid-2026, ten Chinese providers accounted for essentially all meaningful Chinese AI output, a consolidation that happened in under two years. In video generation, Kuaishou’s Kling was being rated by working creative professionals as competitive with the best Western tools on motion quality and realism, at significantly lower cost.

The AI race was no longer American. It had become global, and geopolitical.

The government enters the room

Which brings us to June 2026, and the moment that made the AI conversation a whole lot different.

Anthropic launched Claude Fable 5 on the 9th of June. By any measure it was the most capable publicly available AI model ever released. Stronger reasoning, longer memory, better at complex autonomous tasks than anything before it. Three days later, the US Department of Commerce issued an export control directive and pulled it offline. This extended to every user on the planet, including Anthropic employees who were foreign nationals. The first time in history that a government had switched off a frontier AI model by regulatory order.

The reason was a jailbreak, a technique that Amazon’s cybersecurity researchers had discovered for getting the model to reveal information about software vulnerabilities. Anthropic could not verify users’ nationalities in real time, so they had to pull the model for everyone.

It came back on the 1st of July, with new safety constraints that would filter out certain kinds of requests. Slightly more cautious but basically the same model under new rules. I wrote about the comeback in more detail in a separate post if you want the full story.

Two weeks later, Moonshot AI in Beijing released Kimi K3. It was massive in scale, open for anyone to use, and performed neck and neck with Fable 5 on many of the same tests. What really turned heads was the price. Chinese AI had always been known for being cheap, but K3 didn’t go that route. Moonshot had decided it was good enough to charge what the Western labs charge. This could possibly be the start of a shift in how Chinese AI companies think about pricing. After all, you only need one mover to move the needle.

The week before K3, OpenAI had released GPT-5.6 Sol. Google had Gemini 3.5 in the market. The interval between major releases was now measured in weeks. Geopolitics had entered the product roadmap. All of this, by the way, in under four years, from November 2022 to June 2026.

Oh, and we haven’t even talked about images and video yet

All of this, every model and every milestone above, is only the language model story. There is an entirely separate race running in parallel on image and video generation, and in some ways it has moved even faster in terms of what ordinary people can feel.

In early 2024, AI video meant four-second clips with blurry motion and melting faces. By 2026, you can describe a scene in plain English and get back footage that would pass for a professional camera crew on a budget shoot.

Sora, the OpenAI video model that launched to such extraordinary hype, was shut down in March 2026. It was a combination of brutal economics and an IP war it could not win. The numbers were painful, losing $15 million a day in computing costs against just $2.1 million in total lifetime revenue. Hollywood came out swinging at the same time, with CAA calling it a serious and harmful risk to their clients’ intellectual property, and the Motion Picture Association demanding OpenAI take immediate and decisive action over copyright infringement. The market moved on without it.

Google’s Veo 3 now generates native audio, ambient sound, dialogue and music, in the same pass as the video rather than stitching it in afterwards. Runway’s Gen-4 became the tool that ad agencies and production houses actually adopted into real workflows. Kling from China’s Kuaishou is rated as competitive with all of them on human motion.

A five-second video clip that cost $2.50 to generate in early 2025 costs somewhere between 18 and 30 cents today. That is a ten-fold drop in under eighteen months. But that is a whole separate story, and one I will be covering in a dedicated post.

What this means if you are not in the industry

As I take this to a landing, this is what I feel. The pace of change is real, and it is faster than any prior technological shift, not just incrementally faster but structurally different.

Previous technological eras did not have a recursive quality. The printing press could not help design a better printing press. The steam engine could not help engineer a more efficient one. Humans did that work, slowly and over generations. With AI, the models themselves are being used to help build the next generation of models. The tool is accelerating its own development, and this is what changes the pace of everything.

The context window, essentially how much an AI model can hold in its memory at once, went from around 3,000 words in 2022 to 750,000 words today. We are talking about a 250-fold expansion in under four years, in a single metric.

So here is the thing I kept pondering from a practitioner’s lens. The question that occupied everyone for the first few years was, “which model is the best?” That question is becoming less useful now. Think of it this way. The gap in quality between the best model and the third best model today is so narrow that for most everyday business tasks, you would barely notice the difference. Fable 5 edges K3 on some tests, K3 edges Fable 5 on others. But for a marketing team putting together a campaign brief or a strategist making sense of research, the practical difference is almost nothing.

The more important question, and the one most businesses should be looking at, is how you build around these models in a way that does not leave you stranded every time a new one drops. The Fable 5 suspension was a timely case study. Teams that had built their entire workflow around Fable 5 found out the hard way. In just three days, the thing that took their AI tool offline was not a software update or a pricing change. It was a government.

The race is real and the pace extraordinary. But running it on someone else’s schedule, treating every new release as a reason to stop and re-evaluate everything, is its own kind of losing.

The better question is not what just came out. It is what you are actually building, and whether it would survive the next one.

The race at a glance

If you want to see all the major launches laid out in one place, from ChatGPT in 2022 to Kimi K3 in 2026, I put together a table below. You can filter by US and Chinese labs.

Enjoyed this? I would love to hear how you are making sense of the AI race in your own work — feel free to share your perspective in the comments below.