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

Open-weight models did most of the work on Vercel's AI Gateway in August and collected fourteen cents of the dollar. As The New Stack reports, Vercel's September report puts open-weight models at 56% of all tokens routed through the gateway, their first month as a majority, while Anthropic's models took 64 cents of every dollar spent there.

At a glance

• Vercel's AI Gateway routes tens of trillions of tokens a month, and in August 56% of them went to open-weight models, up from 7% in December 2025 and 36% in July.
• Anthropic's models took 64 cents of every dollar spent through the gateway in August, and Vercel says its share has never dropped below 61% in any month since December 2025.
• The churn happened inside Anthropic's own catalogue: Fable 5 fell from 13.2% of gateway spend to 4.9%, while Opus 5, which Vercel says handles similar workloads at about half the price, climbed to 22.5%.

If you have not been following the gateway numbers, the split between tokens and dollars is not new. Vercel's own April breakdown put spend at 61% Anthropic and 21% Google, while token volume that same month was 38% Google and 26% Anthropic. The same open-weight climb shows up on rival plumbing: The New Stack reported on Monday that open-weight models now account for around 60% of tokens consumed by US-originating requests on OpenRouter, with Chinese models the majority of that.

Open-weight models went from 7% of tokens in December to 56% in August

Vercel's September report, published Thursday and covering activity through August, marks the first month in which open-weight models handled a majority of tokens on AI Gateway. In December 2025 their share was 7%. By April it was 13%, and it rose every month after that; July, in the previous report, came in at 36%.

A single day has already gone higher. Vercel CEO Guillermo Rauch wrote last month that August 22 was a record day for open-weight share of tokens on the gateway, at 62% of traffic, and said he read it as an early signal rather than a peak.

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

Anthropic took 64 cents of every dollar spent through the gateway

Open-weight models from DeepSeek, Moonshot AI and Z.ai are cheaper to run, so the token count and the invoice point in different directions. Those models accounted for 14 cents of every estimated dollar spent through AI Gateway in August while handling 56% of its tokens, and Vercel says the open-weight share of spending is rising.

Anthropic's side has barely moved. Its models took 64 cents of every dollar in August, and Vercel says the share has not fallen below 61% in any month since December 2025, with Claude models holding the top two places by spend throughout that period and often the third as well.

The price of the work itself keeps dropping. The average price per token across the gateway fell 23.2% in August, a third consecutive monthly decline, and among teams that processed more than 10 million tokens in both July and August, the median cost per token fell 7.6%.

Fable 5 fell from 13.2% of spend to 4.9% in a single month

Most of August's movement happened inside Anthropic's own lineup. Fable 5 dropped from 13.2% of total gateway spend in July to 4.9% in August, while the cheaper Opus 5 climbed to 22.5%. Vercel's data shows 90% of teams using Fable cut their usage of it.

More of that work went to Opus 5 than to any other model, and Opus gained almost twice as much usage as Fable lost. Vercel attributes that to the newer model handling similar workloads at roughly half the price. The dollars stayed with Anthropic while the model underneath them changed.

Lab loyalty doesn't follow brand, it follows model profile, and consistency wins.

Google's share of token volume fell from 30% to 5%

Z.ai's newest release moved volume quickly. Within five days of the launch of GLM-5.3-Flash, the model was processing three times the daily volume of GLM-5.2.

Google went the other way. More than three-quarters of the volume lost by Gemini 3 Flash moved to models from other providers, including OpenAI and Anthropic, and Google's overall share of token volume on the gateway fell from 30% to 5%. The Gemini 3 Flash decline alone accounts for 22 of those 25 percentage points.

Vercel's report authors put the rule plainly: when a new model preserves what users valued in its predecessor, the lab retains its customers, and when it doesn't, those customers fill the need through other providers.

Why does Vercel see any of this?

Vercel launched AI Gateway last year so developers could reach models from several providers through one interface, rather than managing separate API keys, accounts and rate limits. The service sits between the application and the providers, routes each request and tracks usage and cost. It is a switchboard that keeps the call logs.

Token volume measures how much inference flows through it. Vercel counts input and output tokens plus reasoning, cached-input and cache-creation tokens, so the figure tracks work handed to models, not money spent on it.

According to The New Stack, routers are hot property for economic reasons: apps used to be hard-coded to one model, while a router can choose per request, sending easier jobs to cheaper models and reserving the pricier frontier systems for work that needs them. Vercel's report frames the money the same way, saying what a workload spends per token is a function of how expensive a wrong answer is to the use case.

Worth noting what the report does not give you. Vercel calls the spend figure an estimate and publishes shares rather than absolute dollars, so both the 14 cents and the 64 cents describe a total we cannot see, and the breakdown does not say which hosts serve the open-weight tokens. In our view, the cheaper-Opus swap is how the 64% held: Anthropic undercut its own model before anyone else could.

Whether Anthropic holds 61%

Vercel has been shipping these reports monthly, the August one covering July and the September one covering August, so a read on September would follow in October, though no date has been given. Three things to watch when it lands: whether the open-weight token share clears the 62% Rauch saw on a single day, whether Anthropic's spend share stays above the 61% floor it has held since December 2025, and whether Google claws anything back from 5%.