Has enterprise AI entered the Multi-Model era?

For years, the AI conversation revolved around rankings: one leader, a handful of challengers, and a constant race for first place. That perspective made sense, much like in the early days of cloud computing, when providers competed to become the platform enterprises would build around. 

Production data now tells a different story. Rather than consuming more AI, organisations are becoming deliberate about where they deploy different models and how they balance cost with performance. 

According to Vercel’s AI Gateway Production Index, open-weight models accounted for 29% of all tokens processed in June 2026 while representing less than 4% of total spend. By comparison, Anthropic accounted for 61% of spend despite processing just 32% of tokens. These differences mean organisations are already assigning different models to different types of work, with the most sensitive workloads being assigned to the most capable models, while higher-volume, lower-criticality tasks are increasingly shifting toward more cost-efficient alternatives.  

A shift in AI economics 

This allocation logic is also reflected in broader usage trends across Vercel’s AI Gateway. In June alone, total token volume increased by 29% month over month, while overall spend grew by 27%, even as the average cost per token remained broadly stable. This means organisations are rebalancing how they allocate workloads, rather than responding to a sudden market shift.  

Token volume and spend are continuing to grow, but organisations are becoming more strategic about how workloads are distributed across lower-cost and premium models. Rather than a one-size-fits-all approach, they’re matching model capability to business value. 

DeepSeek illustrates how quickly these decisions can evolve. The model grew from a negligible share to 22.6% of all tokens processed by June, demonstrating how rapidly organisations can adopt a lower-cost, sufficiently capable alternative once it proves itself in production. 

From model comparison to model allocation 

In practice, technical teams are increasingly thinking which model is best suited to each workload. Which model should handle routine content generation? Which is best suited for a mission-critical coding agent? Which is best suited to regulatory analysis? Every use case becomes an architectural decision rather than simply a vendor selection. 

AI is becoming less about choosing a single provider and more about orchestrating multiple models. A model that performs well for summarisation may not be the best choice for software engineering, regulatory analysis or complex reasoning, making workload allocation an important architectural decision. 

This marks an important shift in how enterprise AI is evaluated. Early adoption often centred on selecting a single strategic provider and building around that ecosystem. Today, AI architectures increasingly resemble cloud infrastructure, where multiple services coexist and workloads are routed according to performance, resilience and cost. Instead of asking which provider is best overall, technical teams are asking which model delivers the best outcome for a specific task. The answer depends on the workload rather than the vendor. 

This specialisation is also visible across modalities. This becomes even more important as AI agents take on complex workflows involving reasoning, tool use and code execution, where different stages of a task may benefit from different models. OpenAI leads image generation with 53% of total volume, while Chinese labs account for roughly two-thirds of video-related spend. Once again, no single provider dominates every category of AI workloads. 

Governance becomes a competitive advantage 

For the most critical and agentic workloads, such as coding agents or back-office automation, Anthropic still accounts for more than 72% of total spend. This shows that trust, reliability and performance continue to outweigh cost when the business stakes are highest.  

For business leaders, this means AI governance is no longer a one-time technology decision. It is becoming an organisational capability that must continually adapt to changing pricing, shifts in model availability and the rapid emergence of new providers. 

A new stage of AI maturity 

It is too early to assume today’s competitive positions are fixed. The most advanced models continue to hold meaningful advantages for demanding use cases, and market dynamics are still evolving rapidly. 

The AI Gateway data already shows this happening. Between May and June, OpenAI’s share of total tokens declined from 12.5% to 10.3%, while its share of total spend increased from 13.3% to 16.1%. This shows a shift toward more complex and higher-value workloads rather than a change in pricing. 

The trend is still emerging, but it already points toward a new level of maturity. The organizations that pull ahead are likely to be those capable of continuously adapting their portfolio of AI models, rather than relying on a single provider to meet every need. 

Enterprise AI is entering a phase where value creation depends as much on orchestrating multiple models as on the intrinsic capabilities of any individual one. For organisations, building this capability to continuously evaluate, allocate and govern AI models could soon become as strategically important as managing procurement, finance or cybersecurity.