Open-weight models have never beaten closed models on Chatbot Arena Elo, but the gap is narrowing fast and the price advantage is already decisive. The median open-weight frontier model runs 15% cheaper than GPT-5.2 at a 90/10 input-to-output blend. DeepSeek V4 Flash undercuts it by roughly 90%. Three massive open releases landed in a single week: Moonshot's Kimi K3 at 2.8 trillion parameters on July 16, Alibaba's Qwen 3.8 at 2.4 trillion on July 19, and Thinking Machines' Inkling, a 975B Apache-2.0 multimodal model, on July 15.
The closed-model side is not standing still. GPT-5.2 and Fable 5, the first Blackwell-trained models, produced a measurable step change in Elo scores in 2026. OpenAI cut inference costs 50%. Kimi shipped a new attention architecture called KDA. Anthropic is about to post its first profitable quarter. The concern that open-source commoditization slows innovation is not supported by what is actually happening.
The full piece is worth reading for the Chatbot Arena Elo chart tracing closed versus open performance from 2023 through mid-2026, and for the blended API pricing breakdown that makes the cost argument concrete. The structural question Tunguz poses is where margins go from here: open-source competitive pressure keeps pricing honest, but closed labs are still setting the pace on capability. That tension is the real story.
[READ ORIGINAL →]