Anyone who works with language models in web and software projects knows the discussion: which model do we use, and what does running it cost?A current analysis provides a surprisingly clear figure on this.According to data from Vercel's AI Gateway – an intermediary layer through which applications route requests to various model providers – 65.1 per cent of total spending in July went to Anthropic.But only around 30 per cent of the tokens were processed there.

A price premium that is accepted

Tokens are the billing unit of language models: small text or code building blocks into which inputs and outputs are broken down.If you weigh spending against volume, an Anthropic token costs on average 4.4 times as much as a token from the other providers in the same gateway.

What is remarkable is less the premium itself than the reaction to it.Development teams apparently pay it deliberately, even though cheaper alternatives would be technically available immediately – a gateway typically allows switching models with little effort.The obvious conclusion: in tasks such as code generation and agentic workflows, where a model independently plans and executes several steps, teams weight output quality and reliability higher than the token price.However, the published figures do not allow any robust statements about the reasons; they show spending and volume, not motives.

Why the token price alone says little

For budget planning, the price per token is a deceptive metric.What matters is how many tokens a task consumes until the result is usable.A cheaper model that needs several attempts, drags along longer contexts or creates rework for developers can be more expensive in total than a high-priced model that delivers on the first pass.

The reverse is equally true: not every subtask requires the most expensive model.Classification, summarisation, format conversion or simple extraction can often be covered with smaller, significantly cheaper models.The spending distribution in the gateway suggests that in practice exactly this split emerges – a large part of the volume runs through cheap models, a large part of the money through a few demanding use cases.

The infrastructure around the model

At the same time, a second development shows that model choice is only part of the bill.Artificial Analysis has published a benchmark with the Search Index that compares search APIs for AI agents – that is, services through which an agent researches and retrieves web content.Quality, cost and speed are assessed.

Seven providers were tested, with GPT-5.6 Luna used as the model.In the evaluation, Parallel, Exa and Firecrawl are in the lead.For project managers, the structure of the benchmark is particularly interesting: it treats search as a standalone component with its own price and latency profile, not as an afterthought to the model.

That reflects the reality of agentic systems.An agent that researches incurs costs in several places: on every search call, when retrieving and preparing the content, and again at the model, which has to process the returned texts as context.A search API that delivers precise and concise results thus indirectly also lowers token costs.One that spews out a lot of raw text drives them up – even if its own call price is low.

What this means for budgeting

Some practical consequences can be derived from both reports without overstretching the figures:

  • Measure cost per task, not per token. A sensible reference quantity is the completed process: a processed ticket, a generated pull request, an answered query.
  • Separate models by task class. Demanding code and agent steps justify an expensive model, routine steps rarely do.
  • Plan for switchability. A gateway or your own abstraction layer keeps a provider change within the scope of a configuration change.Given the current price differences, that is a real lever.
  • Treat search and retrieval as their own cost item. The quality, price and latency of the search API have a direct effect on model costs and response times.
  • Use benchmarks for shortlisting, not for deciding. The Search Index was compiled with a specific model; your own content, languages and domains may produce a different picture.

Limits of the data

Restraint is appropriate when interpreting this.The spending distribution comes from a single gateway and a single month.It reflects the usage behaviour of the development teams active there, not the overall market.Those who work through this gateway tend to have a particular project profile – web and application development with a strong code component.The fact that Anthropic is particularly present there need not apply in other segments.

The factor of 4.4 is also an average across models, context lengths and billing types.It is useful as an indication of the order of magnitude, not as the basis for a concrete calculation.For your own planning, the way forward remains measured values from your own operations.

Assessment for projects

For web and software projects, this above all means: AI costs belong in monitoring from the outset, broken down by use case and component – model, search, retrieval.Anyone with this transparency can view price differences of several hundred per cent calmly and invest where quality actually matters.

And anyone who keeps model choice and search infrastructure interchangeable remains able to act in a market in which prices and rankings will continue to shift.

Sources