Korva is a scroll of the internet’s best ideas.
Across the 12 weeks, token use moved from an 80/20 split of closed versus open to an 80/20 split of open versus closed.
Earlier charts that went viral put open source at 70 to 80 percent of token usage. Those tokens are untracked except on routers.
In a chart of the last 100 days of models, the Claude and Astra models sit highest in price, with Muse, GLM, and Kimi below them and Mimo and GLM from Z.ai lower still. Those are hosted prices. Running some of these open models yourself costs less than 10 cents per million tokens.
How do routers measure token usage?
A router such as OpenRouter sits between developers and many model providers, forwarding each request and billing for it. Because every call passes through the service, it can count tokens by model. Calls made straight to a provider's own API never appear in those counts, and neither does a model running on someone's own hardware. Router charts therefore describe the traffic on that platform rather than the whole market.
How can running a model cost so little?
Open-weight models can be downloaded and run on rented or owned GPUs, so the cost is hardware and electricity rather than a vendor's margin. Servers batch many requests together on one chip, spreading its cost across them. Mixture-of-experts designs, used in several open models, activate only a fraction of their parameters for each token. That cuts the computing needed per token.
Why do labs release their model weights?
Releasing weights spreads a lab's model across developers, who then build tools and research around it. Meta released its Llama 2 weights in 2023 under a license allowing most commercial use. DeepSeek published its R1 reasoning model in January 2025 under the permissive MIT license. Both releases led to many derived models built by outside developers.
More from All-In with Chamath, Jason, Sacks & Friedberg
