A flat rate swaps the token meter for a fixed bill. The useful question is not “is it unlimited?” but “what does it cap instead?”
How it works
Flat-rate pricing means “a single fixed fee for a service, regardless of usage.” For an LLM API, that fee buys access for a period. Because GPU time is finite, every plan bounds use some other way. Common caps:
- a token or message quota per day, week or month;
- a rolling window, such as a few hours, that resets;
- a weekly limit on top of the window;
- concurrency: how many requests run at once;
- speed, which caps output per hour through tokens per second;
- a fixed model list and a fair-use clause.
To compare a plan with pay-as-you-go, price one month of your real traffic per token. If that number is above the plan price, the plan wins, as long as its caps fit how you work.
Why it matters
- Predictable budget. One number per month, with no surprise from a retry loop or a long context.
- Different habits. With no meter, you can resend large contexts and let a model reason without counting tokens.
- Different risks. If the cap is concurrency, parallel jobs queue. If it is a window, work stops until it resets.
On BroadNet Inference
BroadNet sells three plans, all for unlimited DeepSeek V4.1 Flash, priced per session:
| Plan | Price | Period |
|---|---|---|
| Full Moon | $60 | a month, renews |
| Orbit | $15 | a week |
| Meteor | $1 | an hour |
There is no token meter, no monthly cap, no 5-hour window and no weekly limit. The cap is concurrency: a session runs one request at a time, at a typical about 100 tokens per second. Run agents in parallel by buying one session per agent.
In BroadNet’s worked example, with 80% cache hits, Full Moon beats per-token pricing at about 917 million input and 46 million output tokens a month. Read flat rate vs pay per token, how sessions work and what to check in any “unlimited” plan.
Common mistakes
- Buying one session for many parallel workers. They queue behind each other.
- Comparing plan price with list token price while ignoring caching, which lowers the per-token bill.