Open-weight models win the tokens, Anthropic keeps the cash

Vercel’s September report, published September 17, shows open-weight models handling 56% of the tokens routed through its AI Gateway in August while taking only 14 cents of each estimated dollar spent there, The New Stack reports. If your team reads the token share as a sign of where AI budgets are going, the dollars point elsewhere: Anthropic took 64 cents of every dollar, and its share has not fallen below 61% in any month since December 2025.

August was the first month that open-weight models carried a majority of the gateway’s monthly token volume. Their share was 7% in December 2025 and 13% by April, then rose every month after that. Vercel’s previous report put July at 36%.

Tokens and dollars

Vercel’s token count covers input and output tokens, plus reasoning, cached-input and cache-creation tokens. That makes it a measure of how much inference flows through the gateway. It says nothing about the bill. Open-weight models from DeepSeek, Moonshot AI and Z.ai are generally cheaper to run than the proprietary models from US frontier labs, so most of the work and most of the money sit on different sides of the split. Vercel says the open-weight share of spending is rising. It is still far behind their share of usage.

The report carries two price figures that measure different things. Across the whole gateway, the average price per token fell 23.2% in August, the third monthly decline in a row. The second figure covers only teams that processed more than 10 million tokens in both July and August: for them, the median cost per token fell 7.6%.

Guillermo Rauch, Vercel’s CEO, posted in August that August 22 had been a “record day for open weight share of tokens on Vercel AI Gateway,” with 62% of traffic. “This is very likely just the start, because enterprise adoption is still early, and harnesses, CLIs, IDEs, SDKs, etc need to be adapted to be model agnostic,” he wrote. Vercel’s numbers are not the only ones: on Monday, September 14, The New Stack reported that open-weight models made up 60% of OpenRouter’s US token consumption in August, with Chinese-developed models accounting for most of that volume.

Inside Anthropic’s 64%

Anthropic’s 64% survived a sharp reshuffle inside its own lineup. Fable 5 fell from 13.2% of total gateway spend in July to 4.9% in August, and the cheaper Opus 5 rose to 22.5%. Ninety percent of teams using Fable cut their usage, and more of them moved that work to Opus 5 than to any other model. Opus gained almost twice as much usage as Fable lost. Vercel puts this down to the newer model handling similar workloads at roughly half the price. Anthropic models held the top two spots by spend in every month since December 2025, and often the third as well.

“Lab loyalty doesn’t follow brand, it follows model profile, and consistency wins,” the report’s authors write. Vercel saw the pattern elsewhere too: within five days of Z.ai launching GLM-5.3-Flash, the new model was processing three times the daily volume of GLM-5.2.

What happened to Google

Customers cross lab boundaries when a replacement does not match the old model on capability and price, Vercel’s data suggests. More than three-quarters of the volume that Google’s Gemini 3 Flash lost went to models from other providers, OpenAI and Anthropic among them. Google’s overall share of the gateway’s token volume fell from 30% to 5%, and the drop in Gemini 3 Flash alone accounts for 22 of those 25 percentage points.

“When a new model preserves what users valued in its predecessor, the lab retains its customers,” the authors write. “When it doesn’t, those customers fill the need through other providers.”