# Prosus says tighter context and a model router cut agent costs 10x

> Prosus's Ioannis Zempekakis says his team cut agent costs 10x with tighter context and a basic model router.

- Status: reported (One outlet, not independently verified yet.)
- Topic: AI in Practice
- Published: 2026-11-12T09:26:58+00:00
- Updated: 2026-11-12T10:10:29.516705+00:00
- Web page: https://ai-news.ikangai.com/story/prosus-says-tighter-context-and-a-model-router-cut-agent-costs-10x/
- Seen in: Demetrios from the AAIF
- Change: other (Other)

## What changed

Prosus cut the cost of running its AI agents tenfold by tightening the context fed to models and adding a basic model router in front of them.

Ioannis Zempekakis, director of AI and data science at Prosus Group, says his team has cut the cost of running its AI agents tenfold. The team tightened the context fed to models and put a basic model router in front of them. In a video recorded on a tandem bike ride in Amsterdam, he said the earlier habit of sending every task to the most expensive model had stopped scaling. Prosus reached that point after building first on cloud APIs from GCP and OpenAI, where costs were growing exponentially day after day. In his words, agent harnesses are about as greedy as software gets for tokens, because they are given a whole computer. The sandbox itself is cheap, since it is mainly storage. But once an agent can reach many more integrations, one request can pull information from Slack, email, Google Docs, Notion and other tools, and the job can repeat many times. Each retry of a failed task, where a team throws out the codebase and prompts again, adds to the bill. His remedy is simple: give the agent as few instructions as it needs, because a crowded context costs more and, he argues, also works worse.

The router accounts for the second half of the saving. Zempekakis said the team uses a basic router, not a sophisticated one, to decide which model handles which task. That decision depends on evals, which he described as a way to make engineering work more professional. Teams benchmark models per task, cluster the evals, and then pick a model for each cluster. The post does not say how the 10x figure was measured. He also said the optimization has only just started, and that the team is building ten new features a month. The pace of new work is outrunning the savings.

For recurring work, he proposed fine-tuning a much smaller model. For an organization with logistics data, he expects a model trained on its traces to match larger models at 20 to 30 times lower cost. He said that could pay off at 5,000 users but not at five, and that the approach cannot be applied to everything. On governance, he suggested that someone with a budget own the token spend, much as a CFO would, and that teams be told how much their jobs cost. He pointed to scheduling as a source of misuse. Agents run every five minutes, and he asked whether jobs need to run every week or every Saturday at all. He also described a marketplace where teams can find apps, agents and skills others have built. He said it is mainly educational, and that he hopes it will eventually surface reusable work inside the chat itself.

## Why it matters

Routing tasks to cheaper models and assigning budget ownership can cut agent operating costs without losing model flexibility.

## What you can do

Assign one owner to the agent token budget and route simpler tasks to cheaper models.

## Original source

- [Cut One AI Cost and Two More Grow Back - Video | Agentic AI Foundation](https://home.mlops.community/public/videos/cut-one-ai-cost-and-two-more-grow-back)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Prosus says tighter context and a model router cut agent costs 10x", 2026-11-12, https://ai-news.ikangai.com/story/prosus-says-tighter-context-and-a-model-router-cut-agent-costs-10x/
