Open-weight models close the gap with frontier at a fraction of cost

Kimi K3 landed this week with benchmark scores competitive with Fable, currently considered state-of-the-art, at roughly one-third the cost and as an open-weight model. Separately, Laguna S 2.1 showed up with performance allegedly close to DeepSeek V4 at a size comparable to Nemotron 3 Super. The HN reaction to Laguna was pure shock: 'This is INSANE. How did they do this?' Both threads share the same energy: the efficiency curve for open models is bending faster than most people expected.

This is a recurring theme that keeps getting stronger. A few months ago, the argument was that closed frontier models had a meaningful quality lead that justified their cost. That argument is eroding. The commenters on the Kimi thread were already asking about routing layers, specifically which providers support model routers so you can blend Kimi K3 with other models dynamically.

The counterpoint worth noting: Gemini 3.6 Flash dropped simultaneously and got a rough reception. Commenters called it less intelligent and more expensive than GLM-5.2 while being closed-weight. Google releasing a paid closed model into a market where open-weight models are catching up on quality is a hard sell.