Analysis
Two frontier-scale open-weight model releases landed within roughly 24 hours of each other this week: LG AI Research's 750-billion-parameter K-EXAONE 2.0 on July 31, and DeepSeek's V4 Flash 0731 refresh the same day. Neither is an isolated event -- it's the current cadence. Major open-weight releases from Asian labs are now arriving every few weeks, not every few months, and each one resets the effective price and capability floor the whole industry is building against.
DeepSeek has held its V4 Flash pricing steady at $0.14 per million input tokens and $0.28 per million output tokens since the model's original release, with cache-hit pricing dropping to roughly $0.017 per million tokens on repeat context -- pricing that undercuts most closed frontier APIs by an order of magnitude before anyone even accounts for self-hosting. LG's K-EXAONE 2.0 goes further on distribution than price: an unrestricted Apache 2.0 license means enterprises can self-host it entirely, paying only for their own compute.
“The result is a pricing and capability floor that OpenAI, Anthropic and Google don't control -- it moves whenever the next Asian lab ships.”
What's notable is who's shipping these models. DeepSeek and Moonshot AI out of China, and now LG out of South Korea, are the labs setting the open-weight pace, not Meta, whose Llama license carries more usage restrictions, and not any US frontier lab, none of which have released a fully open, frontier-scale model in 2026. The result is a pricing and capability floor that OpenAI, Anthropic and Google don't control -- it moves whenever the next Asian lab ships.
For founders building AI products, the practical implication is that 'which model are you built on' is now a fast-decaying competitive advantage -- the price and performance gap between a chosen closed API and the best available open-weight alternative can close within weeks of a new release. For GPs, it's worth asking portfolio companies how quickly they could migrate off a closed API if pricing pressure from an open-weight competitor forced the issue.
