Jin Daily AI Trivia: Let’s Talk About LLM Trends — OpenRouter Edition
Jin Daily AI Trivia: Let’s Talk About LLM Trends — OpenRouter Edition
Here’s my kopitiam-uncle take on the latest LLM trends, especially from the perspective of technical folks (let’s be real—only devs use OpenRouter for their apps):
Claude Sonnet 4 peaked in usage this month, which explains the hype around “vibe coding.” Sonnet 4 remains the go-to model for web/app dev—less token wastage, solid performance, and great cost-to-performance (C/P) value.
DeepSeek V3 (Free + Paid) is actually the most used model right now, even surpassing Claude Sonnet. The paid version is slow (around 15–25 tokens/sec) and not exactly cheap, but people still love it. Why?Clearly, DeepSeek has met developer expectations—don’t let the media fool you. Also, V3 still beats R1 in speed and delivers solid output hence is more popular.
Google Gemini 2.0 Flash is the fastest model on the list (~160 tokens/sec) and also one of the cheapest among the top 10. But with Gemini 2.5 Flash Lite launched (faster and cheaper), it might become the top pick for devs going forward.
Google has been quietly clawing back LLM API market share since February 2025. We’re now seeing Anthropic models getting slowly replaced by DeepSeek/Kimi—especially after DeepSeek dropped their latest “minor” update. (Coder might say otherwise but I speak from datapoint)
OpenAI models aren’t popular on OpenRouter, mainly because of the strict access requirements (like needing a verified org to use the more powerful models).
LLaMA has basically fallen off the radar. Meta’s now on a massive talent hunt to regain its footing in the LLM game.
Mistral is holding onto just 3% market share, and most of that’s from people using NVIDIA’s Nemo models—likely Euro devs fine-tuning their language-specific models.
Moonshot AI (Kimi) is on the rise this month thanks to its open-source week. It’s now overtaking Qwen as the go-to alternative API provider(beside the top 4).
Qwen, on the other hand, has released a whopping 32 models, but most users seem to prefer self-hosting the smaller ones rather than using their larger parameter APIs on OpenRouter.
Grok is lagging way behind in this space. High pricing and mid performance make it a tough sell for OpenRouter’s dev-heavy user base focused on tool and app building.
Thanks to listen to Jin’s Musing. Hope you learn something today!!! See ya!!!
