DeepSeek’s new models are catching up to the big dogs

DeepSeek just dropped a preview of two new models, and they’re making some bold claims. According to the team, these things are more efficient and perform better than DeepSeek V3.2. The headline grabber? They’ve almost “closed the gap” with the current leading models—both open and closed—on reasoning benchmarks.

Now, I’ve seen a lot of “we beat GPT-4” claims that turned out to be cherry-picked metrics. But DeepSeek has been quietly putting out solid work. Their V3 series was already competitive, and these new models seem to address some of the architectural bottlenecks that held V3.2 back.

What’s interesting here is the phrasing: “closed the gap.” That’s not saying they’re ahead. It’s saying they’re close enough that the difference might not matter for most use cases. I’ve been burned by models that look great on benchmarks but stumble on real-world tasks, so I’ll reserve judgment until I can run my own tests.

The efficiency angle is worth noting too. If these models can deliver near-frontier performance with lower compute costs, that’s a bigger deal than a few points on a leaderboard. We’ve seen this pattern before—DeepSeek tends to optimize for practical deployment rather than just chasing top scores.

I’m not ready to call this a game-changer yet, but it’s another sign that the gap between open and closed models is shrinking faster than many expected. If you’re building on top of open models, this is good news. If you’re selling access to frontier models, you might want to start worrying.