AI is entering a more mature phase, where companies are now choosing between several models depending on uses, costs and the level of sensitivity of the tasks.
For a long time, the debate on AI was structured around a logic of classification: a champion, challengers, a permanent race for first place. This reading has long made sense, like the early years of the cloud, when each player sought to impose its platform as a unique reference.
Production data today tells a different story. On Vercel’s AI Gateway Production Index, so-called open-weight models represented 29% of tokens processed in June 2026, for less than 4% of total spending. Conversely, Anthropic concentrated 61% of the expenditure for 32% of the tokens. These differences suggest a segmentation already at work with the most sensitive tasks entrusted to the most robust models, while higher volume and lower criticality uses migrate towards more economical options.
A silent rocker
This allocation logic can also be seen in the overall evolution of flows on Vercel’s AI Gateway Production Index: the total volume of tokens increased by 29% in one month in June, and the total expenditure by 27%, while the average price per token remained generally stable. The signal suggests a continuous adjustment of trade-offs rather than a sudden shift.
The example of DeepSeek illustrates this speed as the model went from negligible volume to 22.6% of tokens in June. This shows how quickly choices can evolve as soon as a cheaper and sufficiently efficient alternative becomes necessary in production.
From the comparison reflex to allocation logic
Concretely, technical departments are thinking more and more in terms of allocation: which model for a routine writing task, which other for a critical development agent, which third for an analysis with high regulatory implications. Each use case becomes an architectural decision rather than a simple choice of supplier.
This logic also appears by type of use with OpenAI dominating image generation with 53% of the volume, and Chinese labs capturing around two thirds of video spending. Here again, no supplier imposes itself on all uses.
Governance becomes a competitive advantage
On the most critical and agentic cases, such as development agents or back-office automation, Anthropic still concentrates more than 72% of the expenditure. This suggests that trust and reliability remain critical where the stakes are highest.
For managers, this implies rethinking the governance of AI beyond a one-off technological choice, as a continuous organizational skill capable of absorbing price changes, disruptions in access to certain models or the rapid emergence of new entrants.
A new maturity of artificial intelligence
It would be premature to conclude that the positions are fixed. The most advanced models stay ahead of many demanding uses and the balances continue to evolve rapidly.
Some maintain a real lead in the most demanding uses, and positions can still evolve quickly. Data from the AI Gateway illustrates this: between May and June, OpenAI’s share of tokens increased from 12.5% to 10.3%, while its share of expenditure increased from 13.3% to 16.1%, a sign of a change in the type of tasks entrusted to it rather than a change in pricing.
This movement is still emerging, but it is already showing a new form of maturity. The best-positioned companies are likely to be those able to continuously adapt their portfolio of models, rather than seeking the single vendor that fits all needs.
Artificial intelligence is thus entering a phase where its value creation depends as much on its orchestration as on the intrinsic quality of the models. For companies, the challenge now consists of developing this management capacity, which could quickly become a competitive advantage as structuring as mastery of purchasing, finance or cybersecurity.
