Vercel’s AI Gateway Data Proves the Inference Barbell: Anthropic Takes 61% of Spend on 32% of Tokens, DeepSeek Takes 22% of Volume on Under 4%

Vercel’s July 2026 production routing data — not list prices, not benchmarks — shows volume and spend compounding while price per token goes flat, and a Chinese open-weight model ramping 50x in two weeks. The efficiency-era barbell is no longer a thesis. It’s in the numbers.

What Happened

The Vercel AI Gateway Production Index for July 2026 is the most useful data release in AI right now, precisely because it measures what developers actually route in production — not what labs publish on their pricing pages. The headline finding is the three-line chart: through June 2026, usage and spend both climbed steeply while the blended price per token went essentially flat. Token volume rose roughly 29% month-over-month, total spend rose roughly 27%, but the market, as Vercel puts it, “spent more overall, but not more per token.”

The mechanism is a tug-of-war hiding inside the aggregate. Frontier model prices rose approximately 12% per token in June. But cheap open-weight volume kept growing faster, pulling the blended rate back down. The two forces roughly cancelled, leaving the price-per-token line flat after a 20% spike in May. More intelligence was consumed; the per-unit cost of that intelligence did not increase.

The lab-share breakdown is where the structural story becomes undeniable. In June, Anthropic claimed roughly 32% of all tokens routed through the gateway but approximately 61% of all spend — it dominates the premium, high-value, high-stakes workloads. DeepSeek was the mirror image: roughly 22.6% of tokens and under 4% of spend, having moved from essentially zero in April to nearly a third of total volume in about two months. Google sat around 24% of tokens (off an April surge). OpenAI slipped to roughly 10.3% of tokens — down from around 12.5% — even as its spend share rose to approximately 16%, driven by a roughly 50% jump in cost per token. Developers are using OpenAI less often but paying significantly more when they do.

Lab Share in Production — June 2026

Token share vs. spend share (Vercel AI Gateway — developer-weighted sample)

The Structural Read

This is the Map of AI playing out in production. Vercel’s gateway sits at the routing layer — the exact point where inference demand disaggregates by price, capability, and task. What it sees is not what a lab’s pricing page projects. It is what engineers actually select when they have every model available and a cost constraint in the prompt.

Three structural forces are visible in this data simultaneously, and they are not independent of each other.

Force one: value and volume have decoupled. Anthropic at 32% of tokens and 61% of spend versus DeepSeek at 22.6% of tokens and under 4% of spend is the efficiency-era barbell in a single line of data. The majority of volume flows to cheap open-weight models. The majority of value still accrues to the smartest frontier models. This is exactly the split investors like Gavin Baker have argued — now visible in real gateway data rather than argued in a thread. It also maps directly onto the list-price version of the same gap — but list prices tell you what labs want to charge. Gateway data tells you what the market actually routes.

Force two: flat price-per-token is Jevons in the wild. Unit price holds flat while total spend compounds at roughly 27% per month. Cheaper intelligence expands usage faster than it deflates margin. The money is not leaving the inference market; it is redistributing toward whoever has the lowest per-token cost at acceptable quality. The aggregate spend line keeps climbing because the availability of cheaper tokens unlocks demand that did not previously exist — use cases that were too expensive at frontier prices, now viable at open-weight prices. This is the state of the inference economy: deflation at the token, inflation in aggregate.

Force three: Chinese open weights are no longer theoretical. DeepSeek from zero to approximately a third of gateway volume in two months is a production routing fact. GLM 5.2’s 50x daily token ramp at one-fifth the price of Opus — with customer adoption growing even faster than usage — means developers are not just benchmarking Chinese open-weight models. They are routing production workloads through them today. Z.ai is not stopping at the model layer either; its 91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.