For the first time, open-weight models handled more than half of the tokens passing through a commercial AI gateway in production. That is the central finding of the latest “AI Gateway Production Index,” published by Vercel, which aggregates anonymized traffic data from its AI Gateway for August 2026.
According to Vercel’s tally, open-weight models accounted for only 7% of tokens processed in December 2025. Eight months later, in August 2026, they captured 56%, overtaking the entire set of proprietary models combined. This growth has been steady since April.
In dollars, however, open-weight still accounts for only 14% of total platform spend, signaling that closed laboratories retain the bulk of value even as their volumes shrink.
This rise in prominence has mechanical effects on prices: the average cost per token fell by 23.2% in August, the third consecutive monthly decline and the steepest since April.
Among teams processing more than ten million tokens over two consecutive months, the median cost per token fell by 7.6%, more than double the decline observed in July (2.9%).
Anthropic consolidates its global position
The report documents a rapid reallocation by customers within Anthropic’s own lineup.
Fable 5, the lab’s most capable model, had seen its share of spend rise to 13.2% in July after export restrictions to the U.S. were lifted, which had previously blocked access.
One month later, that share had fallen back to 4.9%, while Opus 5, sold at roughly half the price, saw its share jump to 22.5%. Nine out of ten teams that used Fable reduced their usage, with the majority shifting to Opus 5 rather than to a rival.
Yet, this movement did not dent Anthropic’s dominant position. The Claude publisher has captured at least 61% of the total gateway spend each month since December and 64% in August. Its models have occupied the top two spend ranks since December, regardless of which exact models lead the way.
Google loses ground
The story is very different for Google.
Gemini 3 Flash has lost 95% of its volume share since May. About half of the lost volumes moved to cheaper models. At the forefront is GPT-5.6 Luna from OpenAI, whose token price is more than half the price.
The other half migrated toward more expensive models, notably Claude Opus 5 and Sonnet 5, priced at roughly nine and three times as expensive as Gemini 3 Flash, respectively.
Over the period considered, Google’s share in the gateway’s total token volume dropped from 30% to 5%, with Gemini 3 Flash alone accounting for 22 of the 25 percentage points of lost share.
Vercel thus concludes that customer retention isn’t tied to the publisher’s brand but to whether a new model offers a price-to-capacity ratio at least as favorable as the one it replaces.
An example cited to illustrate a successful transition is GLM-5.3-Flash from the Chinese firm Z.ai, which exceeded the daily volume of its predecessor GLM-5.2 the day after its launch and by August 31 was handling two-thirds of Z.ai’s tokens.
OpenAI tests Astra
The report devotes attention to a recent development: GPT-6 Astra, launched by OpenAI in early September at the same price as Anthropic’s Fable 5.1 and two and a half times more expensive than GPT-5.6 Sol.
Two days after its launch, Astra accounted for one dollar out of every three spent on OpenAI models via the gateway. That share then hovered between 28% and 39%.
Vercel notes that spending amounts are estimates based on the list prices published by LLM vendors and may differ from actual billing.
Comparisons of traffic migration between models refer to aggregated movements within the same client teams and do not constitute token-by-token tracking.
The “open-weight” classification follows the current AI Gateway model lineup, broader than the list used in previous reports; this limits strict comparability with historical series.
On the video-generation side, Google’s Veo moved into second place in spend behind Seedance, while the share of Grok Imagine (xAI) has more than halved since June, dropping from 42% to 19%.
