September 28, 2026 · Industry
Open-weight models processed 56 percent of tokens on Vercel’s AI Gateway in August, up from 7 percent in December, according to Financial Times reporting summarized across AI trade desks. Mentions of open models on U.S. earnings calls rose about six times year over year.
AT&T is cited as running roughly 40 percent of its AI workloads on open-weight systems. AlphaSense tracking through August and September points to companies trying to escape rising proprietary API bills by self-hosting.
Chinese-origin checkpoints have taken a large share on aggregators such as OpenRouter in mid-September. NaiveAI’s new 309B MoE release adds another MIT-licensed option into that mix.
Closed labs answered with price cuts: Claude Opus 5.5 and GPT-6 Sol and Luna all landed cheaper than their predecessors. The market is splitting into cheap, self-hosted workhorses and expensive frontier agents—not a single winner-take-all API.
Key takeaway. Developers are voting with tokens. Open weights are no longer a research side channel; they are production infrastructure.
Photo: Photo by Carlos Muza via Unsplash.
Sources: AI Weekly · SiliconReport
