Open-weight models ran 56% of Vercel's gateway tokens for 14 cents of every dollar
Vercel's September Production Index shows the gateway's average price per token down 23.2% in August, a third straight decline, and the median heavy-usage team paying 7.6% less, so most of the saving came from switching models.
The Engineer · Build desk
What happened
- Open-weight models ran 56% of all tokens on Vercel's AI Gateway in August, the first month they took a majority, up from 36% in July and 7% in December 2025.
- The gateway's average price per token fell 23.2% in August, a third consecutive monthly decline, and the median team running more than ten million tokens in both months paid 7.6% less.
- Anthropic took 64 cents of every dollar spent through the gateway in August, has not been below 61 cents in any month since December 2025, and has held the top two spots by spend throughout.
- GPT-6 Astra, launched September 3, took 7.7% of all gateway spend in its first twelve days against 3.7% for Fable 5.1, which arrived two days apart at the same price.
Why it matters
- What doesn't add up contradiction The New Stack reads cheaper inference as money moving to the harness, while the same gateway data shows workloads stepping down a tier inside one lab and Rauch says harnesses, CLIs, IDEs and SDKs still need adapting to be model agnostic.
- What it limits constraint Capturing the gateway-wide decline requires re-qualifying models, because Vercel attributes much of it to adoption of open weights; a team with a frozen approved-model list stays near its July price.
- Who has to decide decision Tier selection inside a single vendor becomes the recurring buying decision: the cheaper model that absorbed Fable's workloads came from Anthropic, so a team has to retest the model without changing its integration work.
- Who's exposed exposure The cheap tokens arrive with a provenance question, since most open-weight volume on OpenRouter's US traffic is Chinese-developed and Deloitte's survey puts 77% of companies weighing country of origin in vendor selection.
Next story · Build · Oct 8
OpenAI sells the same GPT-6.1 Sol model faster for six times the token priceConfirmed2 publishers
Take the two shares Vercel published for August and divide them. Open-weight models ran 56% of gateway tokens and took 14 cents of every estimated dollar [1][7]. Closed models ran the other 44% for 86 cents [26]. The average closed token therefore cost about 7.8 times the average open one: 86 over 44 is 1.95, 14 over 56 is 0.25, and 1.95 divided by 0.25 is 7.8 [28].
That ratio is why the gateway-wide price number moves faster than most bills. The average fell about three times as fast as the median team's own price per token, 23.2% against 7.6% [27]. Vercel credits growth in open-weight adoption for part of the gateway-wide drop, which it calls its steepest since April [5]. Vercel also says the average token now costs less than half what it did five months ago [4]. Most of that came from which models the traffic ran on, so a team holding its approved model list steady tracked the median.
Inside the spend table, the substitution stayed in one vendor's price list. Fable 5, Anthropic's most capable model, went from 13.2% of gateway spend in July to 4.9% in August, and Opus 5 climbed to 22.5% [9]. Nine in ten teams running Fable cut their usage, and more of them moved to Opus 5 than to any other model [10]. Opus gained almost twice the usage Fable lost, which Vercel attributes to the newer model handling the same workloads at roughly half the price [11]. "Lab loyalty doesn't follow brand, it follows model profile, and consistency wins," Vercel's report says [12].
The New Stack argues the falling price is buyers spending on the harness, the software that supplies context, connects tools, routes work and checks results [19]. Vercel's chief executive describes that layer as unfinished. "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," Guillermo Rauch wrote [17]. Where the data does show cross-lab movement, price explains most of it: Gemini 3 Flash has lost 95% of its token share since May, more than three-quarters of that going to other labs, with about half the departing volume landing on cheaper models [14][15].
Before carrying a 23.2% decline into a budget, check the denominator. Vercel counts input and output tokens plus reasoning, cached-input and cache-creation tokens, across applications running on its own infrastructure [18][13]. An average computed that way moves when prompt-cache hit rates move, with an identical model list. For the decline to show up in someone else's bill, their mix of reasoning and cached tokens has to resemble the gateway's.
Zed's launch puts the harness argument into a product. Zed launched Delta in public beta, organising code collaboration around shared threads instead of pull requests, and says 33 of its team members landed 570 changes to Delta's main branch without opening one [20][21]. The public Zed editor repository still accepts conventional pull requests [22]. The new workflow gets the new repo first. "It seems like everyone is in a race to replace GitHub right now," Zed chief executive Nathan Sobo said [23].
What to watch
- Whether Vercel's October index shows open-weight dollar share rising as fast as its token share already has.
- Whether Astra's share of gateway spend holds past its first two weeks, and which lab's share it comes out of.
- Whether Zed moves the public editor repository off pull requests, or keeps threads confined to Delta's own repo.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence58
- Adoption80
- Hype gap+14
- Incentives70
- Confidence64
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
In August, open-weight models ran 56% of all tokens on Vercel's AI Gateway, the first month they took the majority of monthly token volume.
- [2]
Open-weight share of Vercel AI Gateway tokens was 7% in December 2025, 13% by April, and 36% in July, rising every month.
- [3]
Across Vercel's AI Gateway, the average price per token fell 23.2% in August, a third consecutive monthly decline.
- [4]
Vercel says the average token costs less than half what it did five months ago.
- [5]
Vercel says growth in open-weight model adoption helped push the average price per token down 23.2% in August, its third consecutive drop and the steepest since April.
- [6]
Among teams that processed more than 10 million tokens in both July and August, the median cost per token fell 7.6%, more than double July's 2.9% decline.
- [7]
Open-weight models accounted for 14 cents of every estimated dollar spent through AI Gateway in August, despite processing 56% of its tokens.
- [8]
Anthropic's models accounted for 64 cents of every dollar spent through the gateway in August, its share has never fallen below 61% in any month since December 2025, and its models held the top two positions by spend throughout that period.
- [9]
Fable 5, Anthropic's most capable model, fell from 13.2% of total gateway spend in July to 4.9% in August, while the cheaper Opus 5 climbed to 22.5%.
- [10]
Vercel's data indicates 90% of teams using Fable reduced their usage, with more moving those workloads to Opus 5 than to any other model.
- [11]
Opus 5 gained almost twice as much usage as Fable lost, which Vercel attributes to the newer model handling similar workloads at roughly half the price.
- [12]
"Lab loyalty doesn't follow brand, it follows model profile, and consistency wins."
- [13]
AI Gateway routes tens of trillions of tokens each month between production applications and AI labs, across the applications running on Vercel's infrastructure.
- [14]
Gemini 3 Flash has lost 95% of its share of gateway tokens since May, and more than three-quarters of the volume it lost went to models from other labs.
- [15]
About half of the volume that left Gemini 3 Flash went to cheaper models, led by GPT-5.6 Luna.
- [16]
Open-weight models accounted for roughly 60% of OpenRouter's US-originating token consumption in August, with Chinese-developed models making up most of the volume.
- [17]
"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", Vercel CEO Guillermo Rauch wrote.
- [18]
Vercel counts input and output tokens, along with reasoning, cached-input and cache-creation tokens, in its token volume figures.
- [19]
The New Stack wrote that inference is getting cheaper and companies are buying the harness around inference, defining the harness as the software around the model that supplies context, connects tools, routes work and checks results.
- [20]
Zed launched Delta in public beta, building code collaboration around shared threads instead of pull requests.
- [21]
Zed says 33 of its team members landed 570 changes to Delta's main branch without opening a single pull request.
- [22]
The public Zed editor repository still accepts conventional pull requests.
- [23]
"It seems like everyone is in a race to replace GitHub right now," Zed CEO Nathan Sobo said.
- [24]
A Deloitte global survey cited by The New Stack found that 77% of companies factor an AI solution's country of origin into vendor selection.
- [25]
GPT-6 Astra, launched September 3, took 7.7% of all gateway spend in its first twelve days while Fable 5.1 took 3.7%; the two were introduced two days apart at the same price.
- [26]
Closed-weight models ran 44% of gateway tokens in August and took 86 cents of every estimated dollar.
- [27]
The gateway-wide average price per token fell about three times as fast as the median heavy-usage team's cost per token.
- [28]
The average closed-weight token cost about 7.8 times the average open-weight token through the gateway in August.
Sources
2 independent publishers whose own reporting we read for this story.
- blog.vercel.comOpen-weight models take 56% of token volume, Astra doubles Fable 5.1 spend
1 article · September 17, 2026
- thenewstack.ioOpen-weight models now handle a majority of tokens on Vercel’s AI Gateway. But Anthropic still takes 64% of the spend.
2 articles · September 19, 2026