One in twenty tokens on Hostinger’s AI Router went to Claude Opus in September, and those tokens accounted for more than a quarter of what customers paid. AI Router lets customers choose models from Anthropic, OpenAI, DeepSeek and other providers through a single API key, with usage billed to Hostinger Credits, the balance shared across the company’s AI products.
The model mix changed quickly. Between September 28 and November 8, the open-weight DeepSeek V4.1 Flash went from 3% to 40% of AI Router usage. OpenAI’s GPT-6.1 Sol rose from 1% to 13% in its second week.
New hosting plans add a one-time allowance of 250 to 1,000 AI credits to the balance. By Hostinger’s estimate, 250 credits typically cover a first landing page or business website in AI Builder. We asked Emilis Strimaitis, head of product innovation at Hostinger, why customers choose premium models, what drove the rise of DeepSeek V4.1 Flash, when smart routing steps in and how far plan credits go.
How AI Router Customers Choose Between Opus and DeepSeek
webhosting.today: In September, Claude Opus handled 5% of the work but accounted for 27% of the spend. What do customers use premium models like Opus for, compared with open models?
Emilis Strimaitis: To be clear, the AI model usage numbers are from AI Router, where customers pick the model themselves, not from AI Builder or Hostinger Reach.
Now, regarding model usage, many people don’t think about models at all. They just want the job done. Many have had bad experiences with AI in the last few years, when the models changed fast and the results were not always reliable. So they just take the strongest option available. They pick Opus not because the task needs it, but because it feels the safest.
High-end models prove effective where a mistake is costly, such as long coding sessions or agents running multi-step jobs. Open models handle the bulk work around that, like summaries, drafts, and classification, for much less.
In reality, models have gotten very good. For most everyday work, a faster and cheaper model like GPT-6 Luna gives the same good result as a premium model. That is why the AI Router has a smart routing option. It picks the right model for each request and gets the job done for a fraction of the price. Customers can turn it on or keep choosing models themselves.
We use the same logic in our own products. In AI Builder, for example, we route small edits to faster models and use stronger models for full pages, code fixes, and complex apps.
webhosting.today: DeepSeek V4.1 Flash went from 3% to 40% of usage. What drove that jump? Did smart routing play a significant role?
Emilis Strimaitis: Customers drove it, and the reason is simple. DeepSeek V4.1 Flash is a very good model at a very low price.
The way it usually goes is someone tries it to save money, without expecting much, because cheap has usually meant lower quality. They give it one task, then another, and it handles them. After a while it becomes their main model, and then they tell others. That is what the numbers show. It went from 3% to 40% because it is cheap, fast, and handles the vast majority of what people throw at it.
We see the same pattern whenever a strong new model arrives. GPT-6.1 Sol, Claude Sonnet 5.5, and Claude Haiku 5.5 all picked up usage within their first two weeks on AI Router.
Who Picks the Model, and What Caching Saves
webhosting.today: If a customer selects a specific model in AI Router, can smart routing override that choice to reduce costs, or only when the model is unavailable? And who chooses the models used by AI Builder and Hostinger Agent?
Emilis Strimaitis: No. If a customer picks a model, that is what they get. Smart routing is a separate option, and it only kicks in when the customer chooses it. Then we pick the model for each request from a short list our team put together, based on what each model does best. So far the biggest savings are on coding, and we are working on bringing the same to everyday tasks.
For AI Builder and Hostinger Agent, our product teams choose. They try the models themselves, compare quality, cost, and speed, and use whichever fits the job. A small edit in AI Builder goes to a fast model, a full page or a complex app to a stronger one. The models themselves come from Anthropic, OpenAI, Google, and others. We don’t build our own.
webhosting.today: Hostinger says routing and caching can save up to 30% of credits. Is that based on actual customer usage or tests? Does the figure include both routing and caching, and what do customers typically save?
Emilis Strimaitis: It is based on actual, initial usage from AI Router customers, not on tests. The 30% is caching alone. When a request reuses context from an earlier one, the cached part costs less, and we pass that saving to the customer and show it in their usage view. Smart routing saves on top of that by picking a cheaper model where quality is comparable, but we don’t fold it into the same number.
What a customer actually saves depends on how they work. Someone running an agent against the same codebase or document set all day sees a lot of cache reuse. Someone sending unrelated one-off requests sees little.
Caching applies to AI Router. It is not part of how AI Builder is priced.
What Plan Credits Cover and How Long They Last
webhosting.today: What do the 250 AI credits in new Premium hosting plans cover in practice? Could you give examples of simple and complex AI Builder tasks and their credit costs?
Emilis Strimaitis: From what we see, 250 credits is typically enough for a first landing page or a business website. How far they go depends on how clearly the user describes what they want, not on coding skills. A clear, detailed prompt usually gets a usable result in fewer rounds. A vague one takes a few rounds, each taking credits.
Bigger projects, like a web app with user logins, a database, or a Customer Relationship Management integration, take more steps and use more credits. Every project is different, so we can’t put an exact number on it.
webhosting.today: The announcement says plan credits refresh monthly, while the help center describes the 250 credits in hosting plans as a one-time gift. Which plans include monthly credits, and do unused credits roll over?
Emilis Strimaitis: The 250 to 1,000 credits included with hosting plans are a one-time allowance added at purchase. Credits that refresh monthly come with AI credit plans, such as AI Builder plans and the Hostinger Agent subscription. Unused monthly credits carry over to the next month automatically, but rolled-over credits are valid for one month only. Top-ups and gifted credits expire three months after they’re added, and credits included with a hosting plan last as long as the plan.
About the Data
The model usage figures come from Hostinger in writing and cover only early AI Router usage, not AI Builder, Hostinger Agent or Reach. In the questions, work means the share of tokens processed, and spend means what customers paid for those tokens, not Hostinger’s own model costs.