Flat-rate vs per-token LLM APIs: when unlimited pays off
Per-token pricing looks cheap per request and gets expensive per month, because chat and agent workloads re-send their entire history every turn. Here is how to work out which model suits you.
Why token bills grow faster than you expect
Chat models are stateless: each request carries the whole conversation. A chat that has grown to 50,000 tokens sends those 50,000 tokens again on the next message, and the one after. Input tokens, not the replies, dominate most bills.
Three worked examples
These count tokens only; multiply by any provider's per-million price to see the bill.
At tens to hundreds of millions of tokens a month, even low per-token prices add up quickly, and you only find out at the end of the month. Prompt caching discounts help where offered, but they vary by provider and usually still bill cached tokens.
What flat rate changes
- Predictable costthe price is the same whether you send ten messages or ten thousand.
- No self-rationingyou stop trimming context or avoiding regenerations to save money, which is usually what makes output worse.
- Concurrency is the limit, not volumea plan sets how many requests run at once. Within that, use is unlimited.
When per-token is the better deal
Honestly: if you send a handful of short requests a week, pay-as-you-go will be cheaper. If you need dozens of requests in parallel for batch processing, per-token providers scale wider. Flat rate wins for sustained, interactive, long-context use: roleplay, writing, coding agents and research.
How Unquant prices it
Qwen 3.8 Flash Next64K to 262K context by tier
GLM-5.3-FlashFull 262K context on every tier
Every plan includes unlimited tokens and messages; tiers differ by model, context window and concurrent streams. See all plans.
Keep reading
- An uncensored roleplay API with unlimited messagesWhat makes a good uncensored roleplay API: refusals removed in the weights, long context that remembers the whole campaign, and flat-rate unlimited messages. Works with SillyTavern and Janitor AI.
- Uncensored AI for coding and agentsWhy developers use an uncensored model for coding and agents: no refusals on security work, reverse engineering or edgy content pipelines, plus 262K context and unlimited tokens for long agent loops.
- Unlimited and uncensored LLM API providers compared (2026)Unquant, OpenRouter, Featherless and NanoGPT compared on pricing model, usage limits, context windows, uncensored models and prompt logging. Facts checked 1 October 2026.