July 23, 2024: Meta gives away Llama 3.1, the first free-to-download AI that scores level with the paid ones, and Mark Zuckerberg publishes an essay arguing that open AI is the path forward
July 28, 2025: a Chinese lab called Zhipu releases GLM-4.5 free to download, and it ranks third in the world across twelve standard tests, first among all free models
July 2026: the newest free model, Kimi K3, is published this week, and one of the firms that runs these models for businesses reckons a free option now does about four fifths of the work of a top paid one for about a fifth of the cost
On this day, in 2025, a company most British business owners had never heard of gave away an AI that ranked third best in the world, and asked for nothing in return.
Back then that read as either a stunt or a threat, depending on who you asked. The going assumption was that good AI was expensive by nature. The labs charging you every month were charging because they had to, and anything free would be a toy. Plenty of people also read a Chinese lab publishing its best work as a political move rather than a product one.
Here is what actually happened, in plain words. Most AI companies keep the finished thing locked inside their building and rent you access through a website. A handful publish the finished thing instead, as a file anyone can download and run on their own computers. Think of a restaurant. One sells you the dish, another prints the recipe. Zhipu printed the recipe for a dish that was, on the standard tests, nearly as good as anything you could buy. Meta had promised this was coming a year earlier, in July 2024, when it gave away Llama 3.1 and Zuckerberg wrote a long essay defending the whole approach. His pitch at the time was that running the free one yourself cost about half what the paid one did. In 2024 that was an argument. In 2025 it stopped being an argument.
Look at where two years of that leaves you. The free models are not a curiosity in the corner any more, they are the ceiling on what anyone can charge. Every time one lands, the paid labs have to justify the gap. The team behind Kimi K3 published its recipe this week, on the date they promised, and the half-price claim from 2024 now looks conservative: one of the firms that runs these models for businesses puts a good free model at roughly four fifths of the work for a fifth of the price. That is the reason your AI bill this year probably reads differently to last year’s. You did not have to follow any of it. You just get the smaller number.
That is the bit worth keeping. In this business the expensive version is the preview and the cheap version is the product, usually about a year apart. So when someone tells you the tool you need costs a fortune, the useful question is not whether you can afford it today. It is how long until you do not have to.
If you want to read more:
- Mark Zuckerberg’s July 2024 essay on open AI, including the cost argument: https://about.fb.com/news/2024/07/open-source-ai-is-the-path-forward/
- A write-up of the GLM-4.5 release on July 28, 2025: https://www.marktechpost.com/2025/07/28/zhipu-ai-just-released-glm-4-5-series-redefining-open-source-agentic-ai-with-hybrid-reasoning/
Yesterday’s breakthrough is today’s normal.
