Vercel AI Updates: September 18, 2026

1. Open-Weight Models Passed Half of AI Gateway Token Volume While Taking 14% of Spend

Vercel published its September AI Gateway Production Index, reporting that open-weight models carried 56% of token volume in August, up from 7% in December 2025, while accounting for only 14% of spending. Token prices fell 23.2% over the month, a third consecutive monthly decline, and the median price for teams above 10 million tokens per month dropped 7.6%, more than double July’s fall. Source

2. The Same Index Shows Astra Taking 7.7% of Gateway Spend in Twelve Days

Anthropic held 64% of gateway spending in August, but the internal mix moved sharply: when Opus 5 launched at half the price of Fable 5, Fable 5 fell from 13.2% to 4.9% of spend while Opus 5 climbed to 22.5%. GPT-6 Astra took 7.7% of total gateway spending in its first twelve days, roughly double Fable 5.1’s 3.7% over the same window and across twice as many teams. Gemini 3 Flash lost 95% of its token share since May, with more than three-quarters of that volume leaving Google’s lineup entirely. Source

3. Vercel Sandbox Became a Backend for Terminal-Bench and Other Harbor Evals

Vercel added support for running Harbor evals on Vercel Sandbox. Harbor is the open-source harness behind Terminal-Bench, and its registry also covers SWE-bench, tau3-bench, and OSWorld. Passing --env vercel to harbor run executes each trial in its own isolated Firecracker microVM, which removes the local-machine bottleneck from agent benchmark runs. Source

4. The skills CLI Can Install Agent Skills Straight From a Notion Database

Vercel released skills@1.7.0, which adds Notion skills databases as an install source for agent skills. Teams author and review skills as Notion pages in the workspace they already use, then install them into any agent the skills CLI supports, with no Git repository in the loop. Source

5. Static Artifact Deploys Now Complete in Under a Second

Vercel cut static artifact deployment through the Vercel CLI to under one second, aimed at the case where a coding agent produces a prototype, an HTML report, or a single page that needs a shareable URL. Vercel detects eligible deployments automatically and skips the build step for valid artifacts. Source