There’s so much to discuss about LLMs and “AI companies”. Do they take away jobs? Do they kill our planet? Do they infringe copyright laws? Do they have to draw assholes for their logos? I may have hid one question with a different answer than others.
Anyway, I still use LLMs. It started with ChatGPT, then there appeared Claude, the Batman in evil Gotham City. These two companies held monopoly for some time until Chinese companies published strong competitors. Many heard of Deepseek, a whale in the sea of assholes… Alright, alright, I’ll cut it! So DeepSeek got into every respected media. Now we got Xiaomi, Z.ai, MiniMax, Moonshot AI, and many others. These are all Chinese companies which published their models, models’ weights1, and these models are actually competitive with giants!
I had tough time in April. Opening IT chats and Hacker News, I had strong FOMO—everyone there have been talking about AI! Job hunt became a cat-and-mouse game of recruiters with their LLMs filtering out people with wrong keywords against hunters with their LLMs bruteforcing correct keywords. Websites have been becoming worse. And then I found out about open models, tried agents for the first time, and became a person I’d despised. We met with an acquaintance who I haven’t seen for years. In the middle of the silence, I asked what LLM agent he uses.
Many months past since then but I still want to share about open models. They are being developed very fast. Currently the smartest models are:
- Claude Opus 5,
- Claude Fable 5 (unavailable for general public),
- GPT-5.6 Sol,
- Grok 4.6,
- Kimi K3 (open-weight Chinese model),
- GLM-5.3 (open-weight Chinese model),
- Qwen3.8 Max (closed Chinese model),
- Qwen3.8 2.4T A95B (open-weight Chinese model),
- Claude Opus 4.8,
- Muse Spark 1.2.
It’s mind-blowing that open-weight models compete with frontier models (that’s how they call the best of the best) from Anthropic, OpenAI, Elon Musk2, Google, and Meta. If you know this, you start to notice how Claude offers you free credits here, OpenAI discounts rates of their model usage there, and currently I see “Gemini 3.7 Flash is 75% off for a limited time” on openrouter.ai. They are scared.
Today I had to use a smart model to learn some programming stuff. Claude went down so I switched to Kimi K3. It felt less verbose than Opus 5 and more useful. I learned so much and at the end of the day I paid $1.43 for over 80 requests. That’s more expensive than a subsided Claude subscription but less than an alternative usage through Anthropic API.
Open-weight models mean LLMs become a public utility. A Chinese model can run on my laptop and it’s better than something you had to pay $200 per month a few years ago. One day everyone can run their own model smarter than current leaders on a box under their table or in a local neighborhood data center. I don’t know if LLMs make our world brighter but if we must live in the world of thinking computers then better offer these technologies to every thinking human.
Even though, I talk about LLMs, this post was fully written by me. As every other post in 2026. I’m not an asshole.