Aggregated data from extensive data flows between production applications and AI development labs offer valuable insight into current AI adoption trends within companies and startups. This information, collected from trillions of processed tokens monthly, covers language, decision, voice, image, and video models.
In the recent period, a significant shift has occurred, with decision models experiencing expansive growth and now accounting for approximately a quarter of all requests, despite contributing negligibly to overall costs. Nearly a third of all teams utilizing paid AI services in October deployed decision models, and for teams processing over 10 billion tokens monthly, this figure rose to seven out of ten. Other significant trends include the increased adoption of voice models, which were used by about 11% of paid teams in October, representing a substantial jump from July, when it was only 4%.
Production AI is increasingly bifurcating into two main functions: generation and decision-making. Previously, when software needed to make a quick decision, options included either fixed rules, which were inexpensive but inflexible, or language models, which were smart but slow and costly. Decision models now offer a third, highly efficient option. They are used for classification, scoring, or selecting between options instead of generating text, and their costs are significantly lower—approximately seventy times less per request than language models. Although they processed roughly a quarter of all requests in October, they accounted for only about 0.53% of total expenditures. These models are most popular among teams with the largest data processing volumes, who benefit most from their speed and cost-effectiveness.
The use of voice models has grown 2.6 times since July. Approximately one in nine paid teams now utilizes voice models, a result of both new and existing teams integrating them into their solutions. Interaction with an AI application via voice typically occurs in three steps: a voice model converts user speech to text, a language model fomulates a response, and another voice model reads the response back to the user. Many teams use only one or two of these components, with a significant increase in the use of both speech-to-text and text-to-speech models since July. For most teams, voice represents a new pathway for user interaction with existing applications or agents.
The market for AI solution providers is highly competitive. While one major provider dominated a year ago, in October, the two largest players were almost equal in their share of paid teams. However, newly arriving teams in October preferred one provider nearly twice as often as the other. Among the largest teams, representing the top percentile by expenditure, it is common to utilize services from both main providers simultaneously, suggesting competition lies more in the volume of work each secures. One provider saw an increase in its share of total expenditures, while the other maintained a larger overall share, albeit with a slight decline.

Need a hand with your website or app?
Get in touch and I will reply within 24 hours with a proposal for the next step.
A crucial trend is the rapid turnover of AI models. More than half of all language model requests in October were directed at models less than two months old. This marks a significant increase from the preceding period when this share was below thirty percent. The lifespan of a typical model is very short; it often peaks within a month of launch, and within four weeks of that peak, three out of four teams that used it migrate to other solutions. This "half-life" of AI models, lasting approximately six weeks, compels providers to constantly innovate and offer the latest versions to retain customers.
Furthermore, October data indicates that one recently released model became the most popular by expenditure volume. Among the most frequently used models, solutions from multiple providers are emerging. Image and video models also experienced their strongest month to date, with a growing number of teams investing in them, and the largest teams adopting them the fastest. Approximately a third of the largest teams now use image models, and one in six teams utilizes video models, marking an increase from previous periods.
