Vercel reveals top production AI models: Anthropic leads in spending

Vercel analyzed seven months of traffic through its AI Gateway, which handles trillions of tokens across hundreds of models in real applications and agents. The results show what the production AI market actually looks like, as opposed to synthetic benchmarks that change every week.

Who spends more, who processes more

By spending in April 2026, Anthropic leads: 61% of all spending goes to Claude. That holds despite a higher price per token. Developers pay more because the results are worth more to them. Google takes 21%, OpenAI 12%, and the rest is split between xAI and open models.

By volume of tokens processed, the picture is reversed. Google is first here: 38% of all traffic goes through Gemini (mostly Flash, the fast, cheap version). Anthropic handles 26%, OpenAI 13%, and xAI and others 23%.

The gap looks odd, but the logic is simple. Different models compete at different layers:

• Claude Opus goes to complex, expensive tasks, where a mistake costs money
• Gemini Flash soaks up the volume, on tasks where speed matters more than accuracy
• GPT-5.5 is spread evenly across both layers

It is like two different markets inside one market. When a developer picks a model, they don't think about reputation. They think about the balance of price and risk.

The cost of an error determines model choice

Behind this pattern is a simple principle: a model is expensive if an error is expensive.

Personal assistants account for 20% of spending and 40% of tokens. They can run on cheap models because if an assistant makes a mistake, the user notices and corrects it quickly. The error stays local.

Coding agents account for 22% of spending and 20% of tokens. A bug in code costs developer time and debugging. That is pricier than a mistake in a chat, but not critical.

Back-office systems account for 6% of spending and 15% of tokens. Teams economize here because volumes are huge, but they still don't pick the cheapest option. An error can affect finances or operations.

App generation accounts for 7% of spending and 11% of tokens. Generated code goes through code review before use, so there is a safety net.

There is also a larger pattern: B2B applications spend roughly twice as much on tokens as B2C ones. In B2B, an error can lead to financial losses, litigation or downtime. In B2C, an error costs less.

Who wins in which tasks

Sliced by type of work, the market looks fragmented.

Anthropic is clearly ahead in software development: developers choose Claude for complex coding and code analysis. This reflects the model's reputation in ML and system design.

Google dominates consumer applications. Gemini Flash has captured the mass-market layer thanks to low prices and acceptable quality. The strategy is cheap, good enough, high volume.

OpenAI is the most evenly distributed across all categories, meaning GPT-5.5 is used everywhere, from mobile apps to enterprise systems.

xAI and open models pick up cases in specialized niches, for example companies that want to run without the cloud or need full customization.

Over six months, this picture has shifted quickly. The release of a new GPT version in April noticeably increased OpenAI's share of spending. Gemini Flash was much more modest in March, but quickly gained volume. This shows the market responds sharply to quality and price, not to inertia.

What this means

The AI market in 2026 is not a search for one best choice. Developers choose models by task, not by prestige. Expensive models go to high-stakes scenarios (where an error is costly), cheap ones to low-stakes scenarios (where speed and volume matter). New versions gain share quickly if they solve real problems better and cheaper than competitors. And every model wins in its own segment at the same time.