Your agent is a whale. Per-token pricing is built for sardines.
Here's the story every heavy agent user lives — and the math that decides who wins.
Your agent re-reads everything, every turn.
A coding agent doesn't send a short question. It sends the whole conversation, the whole file, the whole plan — again, on every single turn. One prompt into an existing conversation at 1M context is already more than 100,000 input tokens. Ten turns an hour is a million tokens an hour. A serious setup — several agents, long horizons, big repos — runs 10 million tokens an hour, all day, every day.
Per-token APIs bill that whale at sardine prices… per sardine.
At 10M input tokens an hour, GLM-5.2's meter reads $14/hr. Kimi K3 reads $30/hr. And that's before output. The more your agent works, the more you pay — linearly, forever. The per-token model was built for people who ask one question and walk away. It is the exact opposite of what you are.
On a dedicated endpoint, the tokens are free.
You pay for the hour, not the tokens. 10M tokens an hour? 100M? 1 billion tokens an hour, per user, and that's fine. The meter doesn't care. Your server runs only your traffic, at 47–160 tok/s, with up to 1M of context — and the price is the same whether your agent is idle or overflowing it. Heavy use is not a penalty here. It's the point.
And nothing you send is kept.
Your endpoint is dedicated to you. Your prompts, your code, your context — zero data retention. No training on your work, no logs of your conversations, no shared anything. The server is yours for the hour; when it stops, it's gone.