Article: Vercel’s June token-share report shows open-weight models handling 29 % of gateway tokens, up from 11 % in April, with DeepSeek alone accounting for 22.6 % of that volume. The jump signals a rapid shift: developers are moving away from a single “best” model and treating AI models as routable infrastructure.

Why developers are treating models as infrastructure

Cheap, open-weight models—many from Chinese providers—cost a fraction of premium offerings like Anthropic. For routine tasks such as code extraction, text cleaning, or simple data look-ups, a model that costs cents rather than dollars can be more attractive, even if it is a few percent less accurate.

The result is a new design pattern. An application first sends a request to a low-cost model for the “boring” part of the work. If the output passes a simple quality check, the pipeline proceeds; if not, a higher-priced model is invoked for a second pass or final decision. The routing logic becomes the critical piece, not the choice of a single model.

What the numbers say

Vercel’s data, drawn from its gateway that fronts multiple AI providers, shows open-weight models moving from the periphery to the mainstream. In April, they accounted for just over one-tenth of all tokens; by June, they were close to one-third. DeepSeek, a Chinese model, contributed more than one-fifth of the token volume on its own.

High-end models still dominate spend. Anthropic, for example, continues to capture the bulk of monetary outlay because its pricing per token is higher. The math is straightforward: cheap models win the high-volume, low-margin work, while expensive models retain the high-margin, high-impact tasks.

Implications for AI agents

An AI agent typically performs a sequence of steps: planning, data retrieval, tool invocation, and result refinement. When each step can be assigned to the most cost-effective model, the overall operating cost drops dramatically.

  • Data retrieval and extraction often need only basic language understanding, a task cheap models excel at.
  • Final judgment—the part that decides whether the answer is acceptable—remains the domain of premium models that have higher reliability.

This layered approach lets developers automate larger workloads without blowing up budgets. It also forces a shift from “pick the best model” to “measure success per step and route accordingly.”

Risks and limits

The economics are compelling, but the transition is not frictionless.

  • Compliance: Some jurisdictions impose strict data-handling rules that may limit the use of foreign-hosted models.
  • Hosting complexity: Large premium models are difficult to run in-house, so organizations must rely on external APIs, which can raise latency and vendor-lock concerns.

Because of these factors, cheap models are unlikely to replace premium ones entirely. Instead, they become the default for routine work, while premium models stay in the “decision layer” where their higher cost is justified.

What to watch next

  • Tooling for routing: As the routing layer becomes more central, we can expect a wave of SDKs and platforms that automate model selection based on latency, cost, and confidence thresholds.

Developers who start measuring task-level success, tracking retries, and clearly separating “volume” from “judgment” work will be best positioned to profit from the emerging infrastructure mindset.

Takeaway: The AI stack is morphing from a single model choice into a routing problem, where cheap open-weight models handle the bulk of work and premium models are reserved for high-impact decisions. Mastering that routing will be the next competitive edge for AI builders.