What happened

Vercel publishes a monthly AI Gateway Production Index, built from real traffic through its AI Gateway, a service that routes requests between apps and AI providers like OpenAI and Anthropic. The July report covers June 2026 data: gateway token volume, the chunks of text models process, grew 29% month over month, spend grew 27%, and the price per token stayed flat after rising nearly 20% the month before.

The more interesting number is the mix. Open-weight models, meaning models whose trained files are available to download rather than accessible only through a vendor’s service, jumped to 29% of gateway tokens, up from 11% in April, while using under 4% of total spend. DeepSeek alone reached 22.6% of token volume, just behind Google in third place.

The report also has a cautionary example. Anthropic’s Claude Fable 5 launched June 9 and reached 22% of Opus 4.8’s request volume within four days, a fast ramp for a new model. Then a US export-control directive suspended access for the rest of the month.

Why it matters

Anthropic still took 61% of gateway spend on 32% of tokens, and captured more than 72% of spend in every high-stakes use case Vercel tracks. That’s a real signal that builders are paying a premium for quality where mistakes are expensive, even as cheaper open-weight options grow fast everywhere else.

The Claude Fable 5 disruption is the part worth sitting with longer than the growth numbers. A model you’re actively using can become unavailable for reasons that have nothing to do with your code, your account, or the model’s quality. Four days of strong adoption, then an export-control order took it offline for everyone. That’s not a bug you can fix. It’s a dependency you don’t control.

Who should care

If you build with a single hardcoded model or provider, this is worth ten minutes of thought even if you never touch the AI Gateway itself. The open-weight cost gap, 29% of tokens on under 4% of spend, means there’s real headroom to test a cheaper model for lower-stakes parts of your app, like drafting or summarization, while keeping a stronger model for anything user-facing or high-stakes.

If you’re choosing a provider for something new, the high-stakes-spend concentration around Anthropic is a data point worth weighing alongside your own testing, not a substitute for it.

What builders should do next

Check whether your app calls one model directly or uses a shared connection layer that lets you swap providers without a code change. If it’s the former, that’s the actual risk the Claude Fable 5 example points to, not the model’s quality.

If you use a Gateway or router, check its fallback rules for slow, unavailable, or retired models. Configure and test a fallback for your primary model before an outage.

End of article