AgentsModels 🇷🇺 05.08.2026 13:03

Agent for a Dollar: The Economics of the Hermes Agent + DeepSeek V4 Flash Combination

DeepSeekDeepSeek
A viral screenshot claimed hundreds of millions of tokens processed for just over a dollar using the open-source Hermes Agent on DeepSeek V4 Flash. This news explains the mechanics behind such low costs, focusing on automatic prefix caching, and outlines what breaks the cache and the vendor's claims.
In early August, a screenshot circulated on X showing a dashboard with hundreds of millions of processed tokens and a bill of just over a dollar, reportedly from using the open Hermes Agent on the DeepSeek V4 Flash model. The Hermes Agent, developed by Nous Research under the MIT license, works with any OpenAI-compatible endpoint and features a closed learning loop, saving learned skills in markdown files. DeepSeek V4 Flash, in public beta since July 31, uses the same architecture as the April preview but with improved agentic capabilities after fine-tuning. The main cost saving comes not from the low base rate but from automatic prefix caching: repeated prompt prefixes are billed at a rate dozens of times lower. The agentic loop fits this mechanic almost perfectly, as each call resends the same system prompt and history with only a few new lines appended. However, the savings break if elements in the prompt header change, such as timestamps, session identifiers, or dynamically reordered tool lists. The vendor claims the smaller model outperforms the larger preview model on agentic benchmarks, but recommends using the V4-Pro for heavy reasoning. Peak-hour surcharges are announced but not yet active, and hybrid routing with escalation to Pro resets the cache. The article provides pricing rates, lists common cache-breakers, and advises on planning and operational considerations such as rate limits and security perimeter.
Source: Habr — хаб ИИ — original
Our earlier posts on this topic ↓
Fresh news