For the first time, open weight AI models account for the majority of AI usage via Vercel’s AI Gateway. In August, open weight models processed 56% of all tokens, while the average price per token fell by 23.2%. The latest AI Gateway Production Index shows that the price–performance ratio is becoming increasingly important when it comes to choosing an AI model.
A token is a small piece of text – a character, part of a word, a whole word or a punctuation mark – that a language model (such as GPT) uses as the basic unit to read, understand and generate language.
In August, open weight models processed 56% of all tokens via AI Gateway, compared with 13% in April. This means that, for the first time, open-weight models processed more tokens than all closed weight models combined. Despite this strong increase, open weight models in August accounted for only 14% of total spending via the AI Gateway.
The growth of relatively inexpensive open weight models coincides with a further drop in average costs. The price per token fell by 23.2% in August; the third consecutive monthly decline and the steepest since April. Lower costs were also visible among larger users. Among teams that used more than 10 million tokens in both July and August, the median cost per token fell by 7.6%. This is double the 2.9% decline recorded in July.
Within Anthropic’s offering, usage is also shifting towards cheaper models. Claude Fable 5’s share of total AI Gateway spending fell from 13.2% in July to 4.9% in August. At the same time, Opus 5’s share grew to 22.5%.
Per token, Opus 5 costs roughly half as much as Fable 5. Nine out of 10 teams that were using Fable 5 scaled back their usage. More of those teams switched to Opus 5 than to other models. As a result, a large share of usage remained with Anthropic, which in August accounted for 64% of all spending via the AI Gateway. Since December, at least 61% of spending via AI Gateway has gone to Anthropic every month.
This shift shows that users are more likely to choose a cheaper model when it meets their requirements. For AI providers, this means they can retain users if they offer an appropriate alternative within their own line-up.
At OpenAI, a similar split between premium and cheaper models can be seen. GPT-6 Astra was added to AI Gateway on 3 September and costs twice as much per token as GPT-5.6 Sol. Yet two days later, the model already accounted for a third of all spending on OpenAI models. Since then, its share has fluctuated between 28% and 39%.
Astra is also performing well compared with other premium models. Anthropic launched Fable 5.1 two days earlier at the same price. In the first 12 days, Astra accounted for 7.7% of all Gateway spending, more than twice as much as Fable 5.1’s 3.7%. Over that period, Astra was also used by twice as many teams.
The difference between premium and cheaper models is also visible in usage. From 4-16 September, GPT-6 Astra and GPT-5.6 Sol together accounted for 27% of all OpenAI tokens via the AI Gateway, but for 71% of spending. Luna and Nano processed more than twice as many tokens, while spending on those models was nine times lower.
Specialised AI models are also quickly finding their way to users. Within 24 hours of the introduction of Jev by TypeSafe AI, almost 13% of paying teams were using the model. This meant Jev reached more than twice as many teams as earlier models in their first 24 hours and became the fastest-adopted model in the history of the AI Gateway. Within 18 hours, 10% of paying teams were already using Jev, whereas other recent models remained below 7% after 24 hours.
Jev is a specialised model for structured decision-making within software. Unlike general language models, it does not generate free text, but produces structured answers that can be used directly by software. Its rapid adoption shows how quickly specialised models can find a place in production environments.
The data from the AI Gateway Production Index shows that price is playing an increasingly important role in the choice of an AI model. Open-weight and cheaper models are handling a growing share of the token volume, while premium models are retaining a substantial share of spending. This suggests that teams are weighing up, for each use case, which model best fits their needs and budget.
Emerce
Ltd
