Skip to content
NewsResearch

Open-weight models handle 56 percent of Vercel AI Gateway tokens, but Anthropic still keeps 64 percent of spend

· by Pondero Newsdesk

The short version

Vercel's September AI Gateway Production Index found open-weight models processing 56 percent of gateway tokens in August, up from 7 percent in December, while capturing just 14 percent of spend as Anthropic held 64 percent.

Open-weight models handle 56 percent of Vercel AI Gateway tokens, but Anthropic still keeps 64 percent of spend

Vercel's AI Gateway crossed a milestone in August: open-weight models carried a majority of its token traffic for the first time. The dollars tell a different story. Anthropic alone still pulled in 64 cents of every dollar spent through the same gateway, according to Vercel's September Production Index.

What happened

DeepSeek, Moonshot AI, Z.ai, and other open-weight models handled 56 percent of all tokens routed through Vercel's AI Gateway in August, per Vercel's report. That share was just 7 percent in December 2025, reached 13 percent by April, and hit 36 percent in July, according to Vercel's prior monthly reports as cited by The New Stack. Cost explains most of the gap between usage and revenue: because open-weight models are priced far below frontier models, they took only 14 percent of gateway spend in August despite running the majority of tokens. Anthropic's models accounted for the other end of that split, holding 64 percent of all dollars spent, a share that Vercel says has not fallen below 61 percent in any month since December. Inside Anthropic's own lineup, the flagship Fable 5 model's share of gateway spend dropped from 13.2 percent in July to 4.9 percent in August as the cheaper Opus 5 climbed to 22.5 percent; Vercel found that 90 percent of teams using Fable cut their usage, with most of that volume moving to Opus 5 rather than to a competing lab.

Spend stays loyal to Anthropic even as tokens migrate

Vercel's own report authors put it plainly: lab loyalty tracks the model's price-to-capability profile, not the brand, and consistency wins the workload. Anthropic held onto its spend share even while customers shifted work to a cheaper model inside its own catalog, which points to a pattern budget owners should plan around: a team's token volume can swing hard toward cheap open-weight models without shrinking what it owes a frontier vendor, because the priciest, highest-value calls still tend to stay with the most capable model available. Average price per token across the whole gateway fell 23.2 percent in August, the third straight monthly decline, and that continued price drop is exactly what is pulling routine workloads toward the cheapest options while premium spend concentrates elsewhere.

What to watch next

Vercel says open-weight dollar share is rising even though frontier labs still hold most of the spend, so the pace of that convergence into Q4 is one thing to track. The other is Google: more than three-quarters of the token volume it lost on Gemini 3 Flash moved to other providers, per The New Stack, and whether Google claws any of that back will show up in Vercel's October index.

Sources