DeepSeek Flash Reduces Browser Agent Costs by 100x

Retriever AI developers have introduced the Code-as-Plan architecture, which utilizes DeepSeek Flash to generate JavaScript code instead of relying on expensive LLM calls for every single action.

Compiled by Sergey KostenchukPublished 2026-06-25Updated 2026-06-26

2026-06-25 Business

💼 Browser agents will become 100x cheaper

Retriever AI developers have introduced the Code-as-Plan architecture. Instead of using an LLM as an expensive "processor" for every click, the system uses DeepSeek Flash as a "compiler" to generate JavaScript code that is executed locally. This allows logic to be moved from the expensive LLM context into efficient code execution.

🌍 A shift is occurring from the "multimodal tax" toward efficient text-based agents. Using cheap models like DeepSeek Flash as planners breaks the economic monopoly that large laboratories hold over agent creation.

👤 The development of autonomous AI agents is becoming orders of magnitude cheaper and faster. This means that complex browser assistants will soon emerge that work instantly and do not cost a fortune per session.

Source 1: https://www.rtrvr.ai/blog/code-as-plan-deepseek-flash-text-only-browser-agent

Sources