Vercel’s AI Gateway reported that open‑weight models processed 56% of all tokens in August, crossing the majority threshold for the first time. Although they dominate token volume, they represent only 14 ¢ of each dollar spent, while Anthropic’s models still consume 64 ¢ of every dollar, keeping spend concentration on proprietary providers.
Shift in token distribution
The gateway’s token accounting shows a rapid climb from 7% in December 2025 to 56% in August 2026, with a record day in August 22 where open‑weight traffic peaked at 62%. The rise is driven largely by Chinese‑developed models, and the overall average price per token dropped 23.2% in August, marking a third straight month of decline. For teams processing more than 10 million tokens in both July and August, the median cost per token fell 7.6%.
Spend concentration on Anthropic
Despite the token shift, Anthropic’s models still account for nearly two‑thirds of dollar spend on the gateway. Their share has never slipped below 61% since December 2025, with Opus 5 gaining ground (22.5% of spend in August) while the older Fable 5 fell to 4.9%. The internal reallocation shows customers moving to cheaper models within the same lab, preserving overall spend on Anthropic while reducing token usage on higher‑priced variants.
Operational implications
For engineers, the divergence between token volume and spend highlights the need for granular cost monitoring. Since the gateway aggregates input, output, reasoning, and cache tokens, observability pipelines must capture these dimensions to avoid surprise billing. Model‑agnostic tooling—CLIs, SDKs, and CI/CD integrations—should be updated to handle dynamic provider selection, as Vercel’s data suggests enterprises are still early in adopting open‑weight options. Budget forecasts must factor in the lower per‑token price of open‑weight models while accounting for the higher spend share of proprietary labs.
Security and compliance considerations
Routing traffic through a multi‑provider gateway introduces a broader attack surface. Teams should verify that token accounting and usage logs are protected, as they now contain detailed breakdowns of model providers and cost metrics. When crossing lab boundaries—e.g., moving workloads from Google’s Gemini 3 Flash to OpenAI or Anthropic—data residency and compliance requirements may change, requiring a review of policy enforcement at the gateway level.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Monitor token‑level metrics alongside dollar spend to detect cost anomalies early. Adopt or extend model‑agnostic deployment pipelines so you can shift workloads between open‑weight and proprietary models without code changes. Review gateway logging and access controls to ensure that detailed usage data remains secure and compliant. Finally, incorporate price‑per‑token trends into capacity planning, recognizing that lower token costs do not automatically translate to lower spend when proprietary models dominate the budget.